System Identification Toolbox™

User's Guide

R2014b

Lennart Ljung
How to Contact MathWorks

Latest news: www.mathworks.com

Sales and services: www.mathworks.com/sales_and_services

User community: www.mathworks.com/matlabcentral

Technical support: www.mathworks.com/support/contact_us

Phone: 508-647-7000

The MathWorks, Inc.

3 Apple Hill Drive
Natick, MA 01760-2098
System Identification Toolbox™ User's Guide
© COPYRIGHT 1988–2014 by The MathWorks, Inc.
The software described in this document is furnished under a license agreement. The software may be used
or copied only under the terms of the license agreement. No part of this manual may be photocopied or
reproduced in any form without prior written consent from The MathWorks, Inc.
FEDERAL ACQUISITION: This provision applies to all acquisitions of the Program and Documentation
by, for, or through the federal government of the United States. By accepting delivery of the Program
or Documentation, the government hereby agrees that this software or documentation qualifies as
commercial computer software or commercial computer software documentation as such terms are used
or defined in FAR 12.212, DFARS Part 227.72, and DFARS 252.227-7014. Accordingly, the terms and
conditions of this Agreement and only those rights specified in this Agreement, shall pertain to and
govern the use, modification, reproduction, release, performance, display, and disclosure of the Program
and Documentation by the federal government (or other entity acquiring for or through the federal
government) and shall supersede any conflicting contractual terms or conditions. If this License fails
to meet the government's needs or is inconsistent in any respect with federal procurement law, the
government agrees to return the Program and Documentation, unused, to The MathWorks, Inc.
Trademarks
MATLAB and Simulink are registered trademarks of The MathWorks, Inc. See
www.mathworks.com/trademarks for a list of additional trademarks. Other product or brand
names may be trademarks or registered trademarks of their respective holders.
Patents
MathWorks products are protected by one or more U.S. patents. Please see
www.mathworks.com/patents for more information.
Revision History
April 1988 First printing
July 1991 Second printing
May 1995 Third printing
November 2000 Fourth printing Revised for Version 5.0 (Release 12)
April 2001 Fifth printing
July 2002 Online only Revised for Version 5.0.2 (Release 13)
June 2004 Sixth printing Revised for Version 6.0.1 (Release 14)
March 2005 Online only Revised for Version 6.1.1 (Release 14SP2)
September 2005 Seventh printing Revised for Version 6.1.2 (Release 14SP3)
March 2006 Online only Revised for Version 6.1.3 (Release 2006a)
September 2006 Online only Revised for Version 6.2 (Release 2006b)
March 2007 Online only Revised for Version 7.0 (Release 2007a)
September 2007 Online only Revised for Version 7.1 (Release 2007b)
March 2008 Online only Revised for Version 7.2 (Release 2008a)
October 2008 Online only Revised for Version 7.2.1 (Release 2008b)
March 2009 Online only Revised for Version 7.3 (Release 2009a)
September 2009 Online only Revised for Version 7.3.1 (Release 2009b)
March 2010 Online only Revised for Version 7.4 (Release 2010a)
September 2010 Online only Revised for Version 7.4.1 (Release 2010b)
April 2011 Online only Revised for Version 7.4.2 (Release 2011a)
September 2011 Online only Revised for Version 7.4.3 (Release 2011b)
March 2012 Online only Revised for Version 8.0 (Release 2012a)
September 2012 Online only Revised for Version 8.1 (Release 2012b)
March 2013 Online only Revised for Version 8.2 (Release 2013a)
September 2013 Online only Revised for Version 8.3 (Release 2013b)
March 2014 Online only Revised for Version 9.0 (Release 2014a)
October 2014 Online only Revised for Version 9.1 (Release 2014b)
 Contents

Choosing Your System Identification Approach

1
Acknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-2

What Are Model Objects? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-3

Model Objects Represent Linear Systems . . . . . . . . . . . . . . . 1-3
About Model Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-3

Types of Model Objects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-5

Dynamic System Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-7

Numeric Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-9

Numeric Linear Time Invariant (LTI) Models . . . . . . . . . . . . 1-9
Identified LTI Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-9
Identified Nonlinear Models . . . . . . . . . . . . . . . . . . . . . . . . 1-10

About Identified Linear Models . . . . . . . . . . . . . . . . . . . . . . . 1-11

What are IDLTI Models? . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-11
Measured and Noise Component Parameterizations . . . . . . 1-12
Linear Model Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-15

Linear Model Structures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-19

About System Identification Toolbox Model Objects . . . . . . . 1-19
Available Linear Models . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-20
When to Construct a Model Structure Independently of
Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-22
Commands for Constructing Model Structures . . . . . . . . . . 1-22
Model Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-24
See Also . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-26

Imposing Constraints on Model Parameter Values . . . . . . . 1-27

Available Nonlinear Models . . . . . . . . . . . . . . . . . . . . . . . . . . 1-29

v
 Recommended Model Estimation Sequence . . . . . . . . . . . . . 1-30

Supported Models for Time- and Frequency-Domain Data . 1-32

Supported Models for Time-Domain Data . . . . . . . . . . . . . . 1-32
Supported Models for Frequency-Domain Data . . . . . . . . . . 1-33
See Also . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-34

Supported Continuous- and Discrete-Time Models . . . . . . . 1-35

Model Estimation Commands . . . . . . . . . . . . . . . . . . . . . . . . . 1-37

Modeling Multiple-Output Systems . . . . . . . . . . . . . . . . . . . . 1-38

About Modeling Multiple-Output Systems . . . . . . . . . . . . . . 1-38
Modeling Multiple Outputs Directly . . . . . . . . . . . . . . . . . . 1-38
Modeling Multiple Outputs as a Combination of Single-Output
Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-39
Improving Multiple-Output Estimation Results by Weighing
Outputs During Estimation . . . . . . . . . . . . . . . . . . . . . . . 1-39

Regularized Estimates of Model Parameters . . . . . . . . . . . . 1-41

What Is Regularization? . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-41
When to Use Regularization . . . . . . . . . . . . . . . . . . . . . . . . 1-44
Choosing Regularization Constants . . . . . . . . . . . . . . . . . . . 1-46

Estimate Regularized ARX Model Using System Identification

App . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1-49

Data Import and Processing

2
Supported Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-3

Ways to Obtain Identification Data . . . . . . . . . . . . . . . . . . . . . 2-5

Ways to Prepare Data for System Identification . . . . . . . . . . 2-6

Requirements on Data Sampling . . . . . . . . . . . . . . . . . . . . . . . 2-8

Representing Data in MATLAB Workspace . . . . . . . . . . . . . . 2-9

Time-Domain Data Representation . . . . . . . . . . . . . . . . . . . . 2-9

vi Contents
 Time-Series Data Representation . . . . . . . . . . . . . . . . . . . . 2-10
Frequency-Domain Data Representation . . . . . . . . . . . . . . . 2-11

Import Time-Domain Data into the App . . . . . . . . . . . . . . . . 2-16

Import Frequency-Domain Data into the App . . . . . . . . . . . 2-19

Importing Frequency-Domain Input/Output Signals into the
App . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-19
Importing Frequency-Response Data into the App . . . . . . . . 2-21

Import Data Objects into the App . . . . . . . . . . . . . . . . . . . . . 2-25

Specifying the Data Sampling Interval . . . . . . . . . . . . . . . . . 2-28

Specify Estimation and Validation Data in the App . . . . . . 2-30

Preprocess Data Using Quick Start . . . . . . . . . . . . . . . . . . . . 2-32

Create Data Sets from a Subset of Signal Channels . . . . . . 2-33

Create Multiexperiment Data Sets in the App . . . . . . . . . . . 2-35

Why Create Multiexperiment Data? . . . . . . . . . . . . . . . . . . 2-35
Limitations on Data Sets . . . . . . . . . . . . . . . . . . . . . . . . . . 2-35
Merging Data Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-35
Extracting Specific Experiments from a Multiexperiment Data
Set into a New Data Set . . . . . . . . . . . . . . . . . . . . . . . . . 2-39

Managing Data in the App . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-42

Viewing Data Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-42
Renaming Data and Changing Display Color . . . . . . . . . . . 2-43
Distinguishing Data Types . . . . . . . . . . . . . . . . . . . . . . . . . 2-46
Organizing Data Icons . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-46
Deleting Data Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-47
Exporting Data to the MATLAB Workspace . . . . . . . . . . . . 2-48

Representing Time- and Frequency-Domain Data Using

iddata Objects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-50
iddata Constructor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-50
iddata Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-52
Select Data Channels, I/O Data and Experiments in iddata
Objects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-55
Increasing Number of Channels or Data Points of iddata
Objects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-58

vii
 Create Multiexperiment Data at the Command Line . . . . . 2-60
Why Create Multiexperiment Data Sets? . . . . . . . . . . . . . . 2-60
Limitations on Data Sets . . . . . . . . . . . . . . . . . . . . . . . . . . 2-60
Entering Multiexperiment Data Directly . . . . . . . . . . . . . . . 2-60
Merging Data Sets . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-61
Adding Experiments to an Existing iddata Object . . . . . . . . 2-61

Dealing with Multi-Experiment Data and Merging Models 2-63

Managing iddata Objects . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-78

Modifying Time and Frequency Vectors . . . . . . . . . . . . . . . 2-78
Naming, Adding, and Removing Data Channels . . . . . . . . . 2-80
Subreferencing iddata Objects . . . . . . . . . . . . . . . . . . . . . . . 2-82
Concatenating iddata Objects . . . . . . . . . . . . . . . . . . . . . . . 2-82

Representing Frequency-Response Data Using idfrd

Objects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-83
idfrd Constructor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-83
idfrd Properties . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-84
Select I/O Channels and Data in idfrd Objects . . . . . . . . . . 2-85
Adding Input or Output Channels in idfrd Objects . . . . . . . 2-86
Managing idfrd Objects . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-88
Operations That Create idfrd Objects . . . . . . . . . . . . . . . . . 2-88

Analyzing Data Quality . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-90

Is Your Data Ready for Modeling? . . . . . . . . . . . . . . . . . . . 2-90
Plotting Data in the App Versus at the Command Line . . . . 2-91
How to Plot Data in the App . . . . . . . . . . . . . . . . . . . . . . . . 2-91
How to Plot Data at the Command Line . . . . . . . . . . . . . . . 2-96
How to Analyze Data Using the advice Command . . . . . . . . 2-98

Selecting Subsets of Data . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-100

Why Select Subsets of Data? . . . . . . . . . . . . . . . . . . . . . . . 2-100
Extract Subsets of Data Using the App . . . . . . . . . . . . . . . 2-100
Extract Subsets of Data at the Command Line . . . . . . . . . 2-102

Handling Missing Data and Outliers . . . . . . . . . . . . . . . . . . 2-104

Handling Missing Data . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-104
Handling Outliers . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-105
Extract and Model Specific Data Segments . . . . . . . . . . . . 2-106
See Also . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-107

viii Contents
Handling Offsets and Trends in Data . . . . . . . . . . . . . . . . . 2-108
When to Detrend Data . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-108
Alternatives for Detrending Data in App or at the Command-
Line . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-109
Next Steps After Detrending . . . . . . . . . . . . . . . . . . . . . . . 2-110

How to Detrend Data Using the App . . . . . . . . . . . . . . . . . . 2-111

How to Detrend Data at the Command Line . . . . . . . . . . . 2-112

Detrending Steady-State Data . . . . . . . . . . . . . . . . . . . . . . 2-112
Detrending Transient Data . . . . . . . . . . . . . . . . . . . . . . . . 2-112
See Also . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-113

Resampling Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-114

What Is Resampling? . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-114
Resampling Data Without Aliasing Effects . . . . . . . . . . . . 2-115
See Also . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-119

Resampling Data Using the App . . . . . . . . . . . . . . . . . . . . . 2-120

Resampling Data at the Command Line . . . . . . . . . . . . . . . 2-121

Filtering Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-123

Supported Filters . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-123
Choosing to Prefilter Your Data . . . . . . . . . . . . . . . . . . . . 2-123
See Also . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-124

How to Filter Data Using the App . . . . . . . . . . . . . . . . . . . . 2-125

Filtering Time-Domain Data in the App . . . . . . . . . . . . . . 2-125
Filtering Frequency-Domain or Frequency-Response Data in
the App . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-126

How to Filter Data at the Command Line . . . . . . . . . . . . . 2-128

Simple Passband Filter . . . . . . . . . . . . . . . . . . . . . . . . . . . 2-128
Defining a Custom Filter . . . . . . . . . . . . . . . . . . . . . . . . . . 2-129
Causal and Noncausal Filters . . . . . . . . . . . . . . . . . . . . . . 2-130

Generate Data Using Simulation . . . . . . . . . . . . . . . . . . . . . 2-131

Commands for Generating Data Using Simulation . . . . . . 2-131
Create Periodic Input Data . . . . . . . . . . . . . . . . . . . . . . . . 2-132
Generate Output Data Using Simulation . . . . . . . . . . . . . 2-134
Simulating Data Using Other MathWorks Products . . . . . 2-136

ix
 Transforming Between Time- and Frequency-Domain Data 2-137
Transforming Data Domain in the App . . . . . . . . . . . . . . . 2-137
Transforming Data Domain at the Command Line . . . . . . 2-141

Manipulating Complex-Valued Data . . . . . . . . . . . . . . . . . . 2-145

Supported Operations for Complex Data . . . . . . . . . . . . . . 2-145
Processing Complex iddata Signals at the Command Line . 2-145

Linear Model Identification

3
Black-Box Modeling . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-3
Selecting Black-Box Model Structure and Order . . . . . . . . . . 3-3
When to Use Nonlinear Model Structures? . . . . . . . . . . . . . . 3-4
Black-Box Estimation Example . . . . . . . . . . . . . . . . . . . . . . . 3-5

Identifying Frequency-Response Models . . . . . . . . . . . . . . . . 3-8

What Is a Frequency-Response Model? . . . . . . . . . . . . . . . . . 3-8
Data Supported by Frequency-Response Models . . . . . . . . . . 3-8
Estimate Frequency-Response Models in the App . . . . . . . . . 3-9
Estimate Frequency-Response Models at the Command Line 3-10
Selecting the Method for Computing Spectral Models . . . . . 3-11
Controlling Frequency Resolution of Spectral Models . . . . . 3-12
Spectrum Normalization . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-13

Identifying Impulse-Response Models . . . . . . . . . . . . . . . . . 3-16

What Is Time-Domain Correlation Analysis? . . . . . . . . . . . . 3-16
Data Supported by Correlation Analysis . . . . . . . . . . . . . . . 3-16
Estimate Impulse-Response Models Using System
Identification App . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-17
Estimate Impulse-Response Models at the Command Line . 3-18
Compute Response Values . . . . . . . . . . . . . . . . . . . . . . . . . . 3-19
Identify Delay Using Transient-Response Plots . . . . . . . . . . 3-19
Correlation Analysis Algorithm . . . . . . . . . . . . . . . . . . . . . . 3-21

Identifying Process Models . . . . . . . . . . . . . . . . . . . . . . . . . . 3-24

What Is a Process Model? . . . . . . . . . . . . . . . . . . . . . . . . . . 3-24
Data Supported by Process Models . . . . . . . . . . . . . . . . . . . 3-25
Estimate Process Models Using the App . . . . . . . . . . . . . . . 3-25
Estimate Process Models at the Command Line . . . . . . . . . 3-29

x Contents
 Process Model Structure Specification . . . . . . . . . . . . . . . . . 3-35
Estimating Multiple-Input, Multi-Output Process Models . . 3-36
Disturbance Model Structure for Process Models . . . . . . . . . 3-37
Assigning Estimation Weightings . . . . . . . . . . . . . . . . . . . . 3-37
Specifying Initial Conditions for Iterative Estimation
Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-38

Identifying Input-Output Polynomial Models . . . . . . . . . . . 3-40

What Are Polynomial Models? . . . . . . . . . . . . . . . . . . . . . . . 3-40
Data Supported by Polynomial Models . . . . . . . . . . . . . . . . 3-45
Preliminary Step – Estimating Model Orders and Input
Delays . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-46
Estimate Polynomial Models in the App . . . . . . . . . . . . . . . 3-54
Estimate Polynomial Models at the Command Line . . . . . . . 3-58
Polynomial Sizes and Orders of Multi-Output Polynomial
Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-61
Assigning Estimation Weightings . . . . . . . . . . . . . . . . . . . . 3-65
Specifying Initial States for Iterative Estimation Algorithms 3-65
Polynomial Model Estimation Algorithms . . . . . . . . . . . . . . 3-66
Estimate Models Using armax . . . . . . . . . . . . . . . . . . . . . . 3-67

Identifying Transfer Function Models . . . . . . . . . . . . . . . . . 3-70

What are Transfer Function Models? . . . . . . . . . . . . . . . . . 3-70
Data Supported by Transfer Function Models . . . . . . . . . . . 3-71
How to Estimate Transfer Function Models in the System
Identification App . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-72
How to Estimate Transfer Function Models at the Command
Line . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-78
Transfer Function Structure Specification . . . . . . . . . . . . . . 3-79
Estimate Transfer Function Models by Specifying Number of
Poles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-80
Estimate Transfer Function Models with Transport Delay to Fit
Given Frequency Response Data . . . . . . . . . . . . . . . . . . . 3-80
Estimate Transfer Function Models With Prior Knowledge of
Model Structure and Constraints . . . . . . . . . . . . . . . . . . 3-81
Estimate Transfer Function Models with Unknown Transport
Delays . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-82
Estimate Transfer Functions with Delays . . . . . . . . . . . . . . 3-83
Specifying Initial Conditions for Iterative Estimation
Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-84

Refining Linear Parametric Models . . . . . . . . . . . . . . . . . . . 3-86

When to Refine Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-86
What You Specify to Refine a Model . . . . . . . . . . . . . . . . . . 3-86

xi
 Refine Linear Parametric Models Using System Identification
App . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-86
Refine Linear Parametric Models at the Command Line . . . 3-88

Refine ARMAX Model with Initial Parameter Guesses at

Command Line . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-89

Refine Initial ARMAX Model at Command Line . . . . . . . . . 3-91

Extracting Numerical Model Data . . . . . . . . . . . . . . . . . . . . . 3-93

Transforming Between Discrete-Time and Continuous-Time

Representations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-96
Why Transform Between Continuous and Discrete Time? . . 3-96
Using the c2d, d2c, and d2d Commands . . . . . . . . . . . . . . . 3-96
Specifying Intersample Behavior . . . . . . . . . . . . . . . . . . . . . 3-98
Effects on the Noise Model . . . . . . . . . . . . . . . . . . . . . . . . . 3-98

Continuous-Discrete Conversion Methods . . . . . . . . . . . . . 3-100

Choosing a Conversion Method . . . . . . . . . . . . . . . . . . . . . 3-100
Zero-Order Hold . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-101
First-Order Hold . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-102
Impulse-Invariant Mapping . . . . . . . . . . . . . . . . . . . . . . . . 3-103
Tustin Approximation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-104
Zero-Pole Matching Equivalents . . . . . . . . . . . . . . . . . . . . 3-108

Effect of Input Intersample Behavior on Continuous-Time

Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-110

Transforming Between Linear Model Representations . . 3-114

Subreferencing Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-116

What Is Subreferencing? . . . . . . . . . . . . . . . . . . . . . . . . . . 3-116
Limitation on Supported Models . . . . . . . . . . . . . . . . . . . . 3-116
Subreferencing Specific Measured Channels . . . . . . . . . . . 3-116
Separation of Measured and Noise Components of Models . 3-117
Treating Noise Channels as Measured Inputs . . . . . . . . . . 3-118

Concatenating Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-120

About Concatenating Models . . . . . . . . . . . . . . . . . . . . . . . 3-120
Limitation on Supported Models . . . . . . . . . . . . . . . . . . . . 3-120
Horizontal Concatenation of Model Objects . . . . . . . . . . . . 3-121
Vertical Concatenation of Model Objects . . . . . . . . . . . . . . 3-121

xii Contents
 Concatenating Noise Spectrum Data of idfrd Objects . . . . . 3-122
See Also . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-123

Merging Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-124

Building and Estimating Process Models Using System

Identification Toolbox™ . . . . . . . . . . . . . . . . . . . . . . . . . . 3-125

Determining Model Order and Delay . . . . . . . . . . . . . . . . . 3-151

Model Structure Selection: Determining Model Order and

Input Delay . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 3-152

Frequency Domain Identification: Estimating Models Using

Frequency Domain Data . . . . . . . . . . . . . . . . . . . . . . . . . . 3-167

Building Structured and User-Defined Models Using System

Identification Toolbox™ . . . . . . . . . . . . . . . . . . . . . . . . . . 3-193

Identifying State-Space Models

4
What Are State-Space Models? . . . . . . . . . . . . . . . . . . . . . . . . . 4-2
Definition of State-Space Models . . . . . . . . . . . . . . . . . . . . . . 4-2
Continuous-Time Representation . . . . . . . . . . . . . . . . . . . . . 4-2
Discrete-Time Representation . . . . . . . . . . . . . . . . . . . . . . . . 4-3
Relationship Between Continuous-Time and Discrete-Time
State Matrices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-3
State-Space Representation of Transfer Functions . . . . . . . . 4-4

Data Supported by State-Space Models . . . . . . . . . . . . . . . . . 4-5

Supported State-Space Parameterizations . . . . . . . . . . . . . . . 4-6

Estimate State-Space Model With Order Selection . . . . . . . . 4-7

Estimate Model With Selected Order in the App . . . . . . . . . . 4-7
Estimate Model With Selected Order at the Command Line 4-10
Using the Model Order Selection Window . . . . . . . . . . . . . . 4-10

Estimate State-Space Models in System Identification App 4-12

xiii
 Estimate State-Space Models at the Command Line . . . . . . 4-21
Black Box vs. Structured State-Space Model Estimation . . . 4-21
Estimating State-Space Models Using ssest, ssregest and
n4sid . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-22
Choosing the Structure of A, B, C Matrices . . . . . . . . . . . . . 4-23
Choosing Between Continuous-Time and Discrete-Time
Representations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-23
Choosing to Estimate D, K, and X0 Matrices . . . . . . . . . . . . 4-24

Estimate State-Space Models with Free-Parameterization . 4-27

Estimate State-Space Models with Canonical

Parameterization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-28
What Is Canonical Parameterization? . . . . . . . . . . . . . . . . . 4-28
Estimating Canonical State-Space Models . . . . . . . . . . . . . . 4-28

Estimate State-Space Models with Structured

Parameterization . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-30
What Is Structured Parameterization? . . . . . . . . . . . . . . . . 4-30
Specify the State-Space Model Structure . . . . . . . . . . . . . . . 4-30
Are Grey-Box Models Similar to State-Space Models with
Structured Parameterization? . . . . . . . . . . . . . . . . . . . . . 4-32
Estimate Structured Discrete-Time State-Space Models . . . 4-33
Estimate Structured Continuous-Time State-Space Models . 4-34

Estimate the State-Space Equivalent of ARMAX and OE

Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-37

Assigning Estimation Weightings . . . . . . . . . . . . . . . . . . . . . 4-39

Specifying Initial States for Iterative Estimation

Algorithms . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-40

State-Space Model Estimation Methods . . . . . . . . . . . . . . . . 4-41

References . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 4-41

xiv Contents
 Nonlinear Black-Box Model Identification
5
About Nonlinear Model Identification . . . . . . . . . . . . . . . . . . 5-2
What Are Nonlinear Models? . . . . . . . . . . . . . . . . . . . . . . . . 5-2
When to Fit Nonlinear Models . . . . . . . . . . . . . . . . . . . . . . . 5-2
Available Nonlinear Models . . . . . . . . . . . . . . . . . . . . . . . . . 5-4

Preparing Data for Nonlinear Identification . . . . . . . . . . . . . 5-6

Identifying Nonlinear ARX Models . . . . . . . . . . . . . . . . . . . . . 5-7

Nonlinear ARX Model Extends the Linear ARX Structure . . . 5-7
Structure of Nonlinear ARX Models . . . . . . . . . . . . . . . . . . . 5-8
Nonlinearity Estimators for Nonlinear ARX Models . . . . . . . 5-9
Ways to Configure Nonlinear ARX Estimation . . . . . . . . . . 5-11
How to Estimate Nonlinear ARX Models in the System
Identification App . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-14
How to Estimate Nonlinear ARX Models at the Command
Line . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-17
Using Linear Model for Nonlinear ARX Estimation . . . . . . . 5-28
Validating Nonlinear ARX Models . . . . . . . . . . . . . . . . . . . 5-33
Using Nonlinear ARX Models . . . . . . . . . . . . . . . . . . . . . . . 5-39
How the Software Computes Nonlinear ARX Model Output . 5-40

Identifying Hammerstein-Wiener Models . . . . . . . . . . . . . . . 5-48

Applications of Hammerstein-Wiener Models . . . . . . . . . . . 5-48
Structure of Hammerstein-Wiener Models . . . . . . . . . . . . . 5-49
Nonlinearity Estimators for Hammerstein-Wiener Models . 5-50
Ways to Configure Hammerstein-Wiener Estimation . . . . . . 5-52
Estimation Algorithm for Hammerstein-Wiener Models . . . 5-53
How to Estimate Hammerstein-Wiener Models in the System
Identification App . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-54
How to Estimate Hammerstein-Wiener Models at the
Command Line . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-56
Using Linear Model for Hammerstein-Wiener Estimation . . 5-64
Validating Hammerstein-Wiener Models . . . . . . . . . . . . . . . 5-69
Using Hammerstein-Wiener Models . . . . . . . . . . . . . . . . . . 5-77
How the Software Computes Hammerstein-Wiener Model
Output . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-78

Linear Approximation of Nonlinear Black-Box Models . . . 5-81

Why Compute a Linear Approximation of a Nonlinear
Model? . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-81

xv
 Choosing Your Linear Approximation Approach . . . . . . . . . 5-81
Linear Approximation of Nonlinear Black-Box Models for a
Given Input . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-81
Tangent Linearization of Nonlinear Black-Box Models . . . . 5-82
Computing Operating Points for Nonlinear Black-Box
Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5-83

ODE Parameter Estimation (Grey-Box Modeling)

6
Supported Grey-Box Models . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-2

Data Supported by Grey-Box Models . . . . . . . . . . . . . . . . . . . 6-3

Choosing idgrey or idnlgrey Model Object . . . . . . . . . . . . . . . 6-4

Estimating Linear Grey-Box Models . . . . . . . . . . . . . . . . . . . . 6-6

Specifying the Linear Grey-Box Model Structure . . . . . . . . . . 6-6
Create Function to Represent a Grey-Box Model . . . . . . . . . . 6-7
Estimate Continuous-Time Grey-Box Model for Heat
Diffusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-9
Estimate Discrete-Time Grey-Box Model with Parameterized
Disturbance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-12

Estimating Nonlinear Grey-Box Models . . . . . . . . . . . . . . . . 6-15

Specifying the Nonlinear Grey-Box Model Structure . . . . . . 6-15
Constructing the idnlgrey Object . . . . . . . . . . . . . . . . . . . . . 6-17
Using pem to Estimate Nonlinear Grey-Box Models . . . . . . 6-17
Nonlinear Grey-Box Model Estimation Algorithm Options . . 6-18
Represent Nonlinear Dynamics Using MATLAB File for Grey-
Box Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6-20

After Estimating Grey-Box Models . . . . . . . . . . . . . . . . . . . . 6-38

Estimating Coefficients of ODEs to Fit Given Solution . . . . 6-39

Estimate Model Using Zero/Pole/Gain Parameters . . . . . . . 6-46

Creating IDNLGREY Model Files . . . . . . . . . . . . . . . . . . . . . 6-52

xvi Contents
 Time Series Identification
7
What Are Time-Series Models? . . . . . . . . . . . . . . . . . . . . . . . . 7-2

Preparing Time-Series Data . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-3

Estimate Time-Series Power Spectra . . . . . . . . . . . . . . . . . . . 7-4

How to Estimate Time-Series Power Spectra Using the App . 7-4
How to Estimate Time-Series Power Spectra at the Command
Line . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-5

Estimate AR and ARMA Models . . . . . . . . . . . . . . . . . . . . . . . 7-7

Definition of AR and ARMA Models . . . . . . . . . . . . . . . . . . . 7-7
Estimating Polynomial Time-Series Models in the App . . . . . 7-7
Estimating AR and ARMA Models at the Command Line . . 7-10

Estimate State-Space Time-Series Models . . . . . . . . . . . . . . 7-11

Definition of State-Space Time-Series Model . . . . . . . . . . . . 7-11
Estimating State-Space Models at the Command Line . . . . 7-11

Identify Time-Series Models at Command Line . . . . . . . . . . 7-12

Estimate Nonlinear Models for Time-Series Data . . . . . . . . 7-14

Estimate ARIMA Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-15

Spectrum Estimation Using Complex Data - Marple's Test

Case . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7-18

Analyzing Time-Series Models . . . . . . . . . . . . . . . . . . . . . . . . 7-28

Recursive Model Identification

8
What Is Recursive Estimation? . . . . . . . . . . . . . . . . . . . . . . . . 8-2

Data Supported for Recursive Estimation . . . . . . . . . . . . . . . 8-3

xvii
 Algorithms for Recursive Estimation . . . . . . . . . . . . . . . . . . . 8-4
Types of Recursive Estimation Algorithms . . . . . . . . . . . . . . 8-4
General Form of Recursive Estimation Algorithm . . . . . . . . . 8-4
Kalman Filter Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-5
Forgetting Factor Algorithm . . . . . . . . . . . . . . . . . . . . . . . . . 8-7
Unnormalized and Normalized Gradient Algorithms . . . . . . . 8-9

Data Segmentation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 8-11

Recursive Estimation and Data Segmentation Techniques in

System Identification Toolbox™ . . . . . . . . . . . . . . . . . . . . 8-12

Online Estimation
9
What Is Online Estimation? . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-2
Requirements . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-3

How Online Estimation Differs from Offline Estimation . . . 9-4

Preprocess Online Estimation Data . . . . . . . . . . . . . . . . . . . . 9-6

Validate Online Estimation Results . . . . . . . . . . . . . . . . . . . . 9-7

Troubleshooting Online Estimation . . . . . . . . . . . . . . . . . . . . 9-9

Check Model Structure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-9
Check Model Order . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 9-10
Preprocess Estimation Data . . . . . . . . . . . . . . . . . . . . . . . . 9-10
Check Initial Guess for Parameter Values . . . . . . . . . . . . . . 9-10
Check Estimation Settings . . . . . . . . . . . . . . . . . . . . . . . . . 9-10

Generate Online Estimation Code . . . . . . . . . . . . . . . . . . . . . 9-12

Recursive Algorithms for Online Estimation . . . . . . . . . . . . 9-14

General Form of Recursive Estimation . . . . . . . . . . . . . . . . 9-14
Types of Recursive Estimation Algorithms . . . . . . . . . . . . . 9-15

Online Recursive Least Squares Estimation . . . . . . . . . . . . 9-19

Online ARMAX Polynomial Model Estimation . . . . . . . . . . . 9-32

xviii Contents
 State Estimation Using Time-Varying Kalman Filter . . . . . 9-48

Model Analysis
10
Validating Models After Estimation . . . . . . . . . . . . . . . . . . . 10-2
Ways to Validate Models . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-2
Data for Model Validation . . . . . . . . . . . . . . . . . . . . . . . . . . 10-3
Supported Model Plots . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-3
Definition of Confidence Interval for Specific Model Plots . . 10-4

Plotting Models in the App . . . . . . . . . . . . . . . . . . . . . . . . . . 10-6

Simulating and Predicting Model Output . . . . . . . . . . . . . . 10-8

Why Simulate or Predict Model Output . . . . . . . . . . . . . . . 10-8
Definition: Simulation and Prediction . . . . . . . . . . . . . . . . . 10-9
Simulation and Prediction in the App . . . . . . . . . . . . . . . . 10-11
Simulation and Prediction at the Command Line . . . . . . . 10-15
Compare Simulated Output with Measured Data . . . . . . . 10-17
Simulate Model Output with Noise . . . . . . . . . . . . . . . . . . 10-18
Simulate a Continuous-Time State-Space Model . . . . . . . . 10-19
Predict Using Time-Series Model . . . . . . . . . . . . . . . . . . . 10-20

Residual Analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-23

What Is Residual Analysis? . . . . . . . . . . . . . . . . . . . . . . . . 10-23
Supported Model Types . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-24
What Residual Plots Show for Different Data Domains . . . 10-24
Displaying the Confidence Interval . . . . . . . . . . . . . . . . . . 10-25
How to Plot Residuals Using the App . . . . . . . . . . . . . . . . 10-26
How to Plot Residuals at the Command Line . . . . . . . . . . 10-27
Examine Model Residuals . . . . . . . . . . . . . . . . . . . . . . . . . 10-27

Impulse and Step Response Plots . . . . . . . . . . . . . . . . . . . . 10-31

Supported Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-31
How Transient Response Helps to Validate Models . . . . . . 10-31
What Does a Transient Response Plot Show? . . . . . . . . . . 10-32
Displaying the Confidence Interval . . . . . . . . . . . . . . . . . . 10-33

How to Plot Impulse and Step Response Using the App . . 10-34

xix
 How to Plot Impulse and Step Response at the Command
Line . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-36

Frequency Response Plots . . . . . . . . . . . . . . . . . . . . . . . . . . 10-38

What Is Frequency Response? . . . . . . . . . . . . . . . . . . . . . . 10-38
How Frequency Response Helps to Validate Models . . . . . 10-39
What Does a Frequency-Response Plot Show? . . . . . . . . . . 10-39
Displaying the Confidence Interval . . . . . . . . . . . . . . . . . . 10-40

How to Plot Bode Plots Using the App . . . . . . . . . . . . . . . . 10-42

How to Plot Bode and Nyquist Plots at the Command

Line . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-44

Noise Spectrum Plots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-46

Supported Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-46
What Does a Noise Spectrum Plot Show? . . . . . . . . . . . . . 10-46
Displaying the Confidence Interval . . . . . . . . . . . . . . . . . . 10-47

How to Plot the Noise Spectrum Using the App . . . . . . . . 10-49

How to Plot the Noise Spectrum at the Command Line . . 10-52

Pole and Zero Plots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-54

Supported Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-54
What Does a Pole-Zero Plot Show? . . . . . . . . . . . . . . . . . . 10-54
Reducing Model Order Using Pole-Zero Plots . . . . . . . . . . 10-56
Displaying the Confidence Interval . . . . . . . . . . . . . . . . . . 10-56

How to Plot Model Poles and Zeros Using the App . . . . . . 10-58

How to Plot Poles and Zeros at the Command Line . . . . . 10-60

Analyzing MIMO Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-61

Overview of Analyzing MIMO Models . . . . . . . . . . . . . . . . 10-61
Array Selector . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-62
I/O Grouping for MIMO Models . . . . . . . . . . . . . . . . . . . . 10-64
Selecting I/O Pairs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-65

Customizing Response Plots Using the Response Plots

Property Editor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-67
Opening the Property Editor . . . . . . . . . . . . . . . . . . . . . . . 10-67
Overview of Response Plots Property Editor . . . . . . . . . . . 10-68

xx Contents
 Labels Pane . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-70
Limits Pane . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-70
Units Pane . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-71
Style Pane . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-77
Options Pane . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-77
Editing Subplots Using the Property Editor . . . . . . . . . . . 10-78

Akaike's Criteria for Model Validation . . . . . . . . . . . . . . . . 10-80

Definition of FPE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-80
Computing FPE . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-81
Definition of AIC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-81
Computing AIC . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-82

Computing Model Uncertainty . . . . . . . . . . . . . . . . . . . . . . . 10-83

Why Analyze Model Uncertainty? . . . . . . . . . . . . . . . . . . . 10-83
What Is Model Covariance? . . . . . . . . . . . . . . . . . . . . . . . . 10-83
Types of Model Uncertainty Information . . . . . . . . . . . . . . 10-84

Troubleshooting Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-86

About Troubleshooting Models . . . . . . . . . . . . . . . . . . . . . 10-86
Model Order Is Too High or Too Low . . . . . . . . . . . . . . . . 10-86
Nonlinearity Estimator Produces a Poor Fit . . . . . . . . . . . 10-87
Substantial Noise in the System . . . . . . . . . . . . . . . . . . . . 10-87
Unstable Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 10-88
Missing Input Variables . . . . . . . . . . . . . . . . . . . . . . . . . . 10-89
Complicated Nonlinearities . . . . . . . . . . . . . . . . . . . . . . . . 10-89

Next Steps After Getting an Accurate Model . . . . . . . . . . . 10-91

Setting Toolbox Preferences

11
Toolbox Preferences Editor . . . . . . . . . . . . . . . . . . . . . . . . . . 11-2
Overview of the Toolbox Preferences Editor . . . . . . . . . . . . 11-2
Opening the Toolbox Preferences Editor . . . . . . . . . . . . . . . 11-2

Units Pane . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-4

Style Pane . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-7

xxi
 Options Pane . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-8

SISO Tool Pane . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11-9

Control Design Applications

12
Using Identified Models for Control Design Applications . . 12-2
How Control System Toolbox Software Works with Identified
Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-2
Using balred to Reduce Model Order . . . . . . . . . . . . . . . . . . 12-2
Compensator Design Using Control System Toolbox
Software . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-3
Converting Models to LTI Objects . . . . . . . . . . . . . . . . . . . . 12-3
Viewing Model Response Using the LTI Viewer . . . . . . . . . 12-4
Combining Model Objects . . . . . . . . . . . . . . . . . . . . . . . . . . 12-5

Create and Plot Identified Models Using Control System

Toolbox Software . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 12-6

System Identification Toolbox Blocks

13
Using System Identification Toolbox Blocks in Simulink
Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13-2

Preparing Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13-3

Identifying Linear Models . . . . . . . . . . . . . . . . . . . . . . . . . . . 13-4

Simulating Identified Model Output in Simulink . . . . . . . . 13-5

When to Use Simulation Blocks . . . . . . . . . . . . . . . . . . . . . 13-5
Summary of Simulation Blocks . . . . . . . . . . . . . . . . . . . . . . 13-5
Specifying Initial Conditions for Simulation . . . . . . . . . . . . 13-6

Simulate Identified Model Using Simulink Software . . . . . 13-8

xxii Contents
 System Identification App
14
Steps for Using the System Identification App . . . . . . . . . . 14-2

Working with System Identification App . . . . . . . . . . . . . . . 14-3

Starting and Managing Sessions . . . . . . . . . . . . . . . . . . . . . 14-3
Managing Models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14-7
Working with Plots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 14-11
Customizing the System Identification App . . . . . . . . . . . . 14-15

xxiii
xxiv
 1

Choosing Your System Identification

Approach

• “Acknowledgments” on page 1-2

• “What Are Model Objects?” on page 1-3
• “Types of Model Objects” on page 1-5
• “Dynamic System Models” on page 1-7
• “Numeric Models” on page 1-9
• “About Identified Linear Models” on page 1-11
• “Linear Model Structures” on page 1-19
• “Imposing Constraints on Model Parameter Values” on page 1-27
• “Available Nonlinear Models” on page 1-29
• “Recommended Model Estimation Sequence” on page 1-30
• “Supported Models for Time- and Frequency-Domain Data” on page 1-32
• “Supported Continuous- and Discrete-Time Models” on page 1-35
• “Model Estimation Commands” on page 1-37
• “Modeling Multiple-Output Systems” on page 1-38
• “Regularized Estimates of Model Parameters” on page 1-41
• “Estimate Regularized ARX Model Using System Identification App” on page 1-49
1 Choosing Your System Identification Approach

Acknowledgments
System Identification Toolbox™ software is developed in association with the following
leading researchers in the system identification field:

Lennart Ljung. Professor Lennart Ljung is with the Department of Electrical

Engineering at Linköping University in Sweden. He is a recognized leader in system
identification and has published numerous papers and books in this area.

Qinghua Zhang. Dr. Qinghua Zhang is a researcher at Institut National de

Recherche en Informatique et en Automatique (INRIA) and at Institut de Recherche
en Informatique et Systèmes Aléatoires (IRISA), both in Rennes, France. He conducts
research in the areas of nonlinear system identification, fault diagnosis, and signal
processing with applications in the fields of energy, automotive, and biomedical systems.

Peter Lindskog. Dr. Peter Lindskog is employed by NIRA Dynamics AB, Sweden. He
conducts research in the areas of system identification, signal processing, and automatic
control with a focus on vehicle industry applications.

Anatoli Juditsky. Professor Anatoli Juditsky is with the Laboratoire Jean Kuntzmann
at the Université Joseph Fourier, Grenoble, France. He conducts research in the areas of
nonparametric statistics, system identification, and stochastic optimization.

1-2
 What Are Model Objects?

What Are Model Objects?

In this section...
“Model Objects Represent Linear Systems” on page 1-3
“About Model Data” on page 1-3

Model Objects Represent Linear Systems

In Control System Toolbox™, System Identification Toolbox, and Robust Control
Toolbox™ software, you represent linear systems as model objects. Model objects are
specialized data containers that encapsulate model data and other attributes in a
structured way. Model objects allow you to manipulate linear systems as single entities
rather than keeping track of multiple data vectors, matrices, or cell arrays.

Model objects can represent single-input, single-output (SISO) systems or multiple-input,

multiple-output (MIMO) systems. You can represent both continuous- and discrete-time
linear systems.

The main families of model objects are:

• Numeric Models — Basic representation of linear systems with fixed numerical

coefficients. This family also includes identified models that have coefficients
estimated with System Identification Toolbox software.
• Generalized Models — Representations that combine numeric coefficients
with tunable or uncertain coefficients. Generalized models support tasks such as
parameter studies or compensator tuning using Robust Control Toolbox tuning
commands.

About Model Data

The data encapsulated in your model object depends on the model type you use. For
example:

• Transfer functions store the numerator and denominator coefficients

• State-space models store the A, B, C, and D matrices that describe the dynamics of
the system
• PID controller models store the proportional, integral, and derivative gains

1-3
1 Choosing Your System Identification Approach

Other model attributes stored as model data include time units, names for the model
inputs or outputs, and time delays.

Note: All model objects are MATLAB® objects, but working with them does not require
a background in object-oriented programming. To learn more about objects and object
syntax, see “Classes in the MATLAB Language” in the MATLAB documentation.

More About
• “Types of Model Objects” on page 1-5

1-4
 Types of Model Objects

Types of Model Objects

The following diagram illustrates the relationships between the types of model objects
in Control System Toolbox, Robust Control Toolbox, and System Identification Toolbox
software. Model types that begin with id require System Identification Toolbox software.
Model types that begin with u require Robust Control Toolbox software. All other model
types are available with Control System Toolbox software.

Analysis Point Block

AnalysisPoint

The diagram illustrates the following two overlapping broad classifications of model
object types:

1-5
1 Choosing Your System Identification Approach

• Dynamic System Models vs. Static Models — In general, Dynamic System Models
represent systems that have internal dynamics, while Static Models represent static
input/output relationships.
• Numeric Models vs. Generalized Models — Numeric Models are the basic
numeric representation of linear systems with fixed coefficients. Generalized Models
represent systems with tunable or uncertain components.

More About
• “What Are Model Objects?” on page 1-3
• “Dynamic System Models” on page 1-7
• “Numeric Models” on page 1-9

1-6
 Dynamic System Models

Dynamic System Models

Dynamic System Models generally represent systems that have internal dynamics or
memory of past states (such as integrators).

Most commands for analyzing linear systems, such as bode, margin, and ltiview, work
on most Dynamic System Model objects. For Generalized Models, analysis commands use
the current value of tunable parameters and the nominal value of uncertain parameters.

The following table lists the Dynamic System Models.

Model Family Model Types

Numeric LTI models — Basic numeric tf
representation of linear systems zpk
ss
frd
pid
pidstd
Identified LTI models — Representations idtf
of linear systems with tunable coefficients, idss
whose values can be identified using
measured input/output data. idfrd
idgrey
idpoly
idproc
Identified nonlinear models — idnlarx
Representations of nonlinear systems with idnlhw
tunable coefficients, whose values can be
identified using input/output data. Limited idnlgrey
support for commands that analyze linear
systems.
Generalized LTI models — Representations genss
of systems that include tunable or genfrd
uncertain coefficients
uss

1-7
1 Choosing Your System Identification Approach

Model Family Model Types

ufrd
Dynamic Control Design Blocks — ltiblock.gain
Tunable, uncertain, or switch components ltiblock.tf
for constructing models of control systems
ltiblock.ss
ltiblock.pid
ltiblock.pid2
ultidyn
udyn
AnalysisPoint

More About
• “Numeric Linear Time Invariant (LTI) Models” on page 1-9
• “Identified LTI Models” on page 1-9
• “Identified Nonlinear Models” on page 1-10

1-8
 Numeric Models

Numeric Models

Numeric Linear Time Invariant (LTI) Models

Numeric LTI models are the basic numeric representation of linear systems or
components of linear systems. Use numeric LTI models for modeling dynamic
components, such as transfer functions or state-space models, whose coefficients are
fixed, numeric values. You can use numeric LTI models for linear analysis or control
design tasks.

The following table summarizes the available types of numeric LTI models.

Model Type Description

tf Transfer function model in polynomial form
zpk Transfer function model in zero-pole-gain (factorized) form
ss State-space model
frd Frequency response data model
pid Parallel-form PID controller
pidstd Standard-form PID controller

Identified LTI Models

Identified LTI Models represent linear systems with coefficients that are identified
using measured input/output data. You can specify initial values and constraints for the
estimation of the coefficients.

The following table summarizes the available types of identified LTI models.

Model Type Description

idtf Transfer function model in polynomial form, with
identifiable parameters
idss State-space model, with identifiable parameters
idpoly Polynomial input-output model, with identifiable
parameters

1-9
1 Choosing Your System Identification Approach

Model Type Description

idproc Continuous-time process model, with identifiable
parameters
idfrd Frequency-response model, with identifiable parameters
idgrey Linear ODE (grey-box) model, with identifiable parameters

Identified Nonlinear Models

Identified Nonlinear Models represent nonlinear systems with coefficients that are
identified using measured input/output data. You can specify initial values and
constraints for the estimation of the coefficients.

The following table summarizes the available types of identified nonlinear models.

Model Type Description

idnlarx Nonlinear ARX model, with identifiable
parameters
idnlgrey Nonlinear ODE (grey-box) model, with
identifiable parameters
idnlhw Hammerstein-Wiener model, with
identifiable parameters

1-10
 About Identified Linear Models

About Identified Linear Models

In this section...
“What are IDLTI Models?” on page 1-11
“Measured and Noise Component Parameterizations” on page 1-12
“Linear Model Estimation” on page 1-15

What are IDLTI Models?

System Identification Toolbox software uses objects to represent a variety of linear
and nonlinear model structures. These linear model objects are collectively known as
Identified Linear Time-Invariant (IDLTI) models.

IDLTI models contain two distinct dynamic components:

• Measured component — Describes the relationship between the measured inputs

and the measured output (G)
• Noise component — Describes the relationship between the disturbances at the
output and the measured output (H)

Models that only have the noise component H are called time-series or signal models.
Typically, you create such models using time-series data that consist of one or more
outputs y(t) with no corresponding input.

The total output is the sum of the contributions from the measured inputs and the
disturbances: y = G u + H e, where u represents the measured inputs and e the
disturbance. e(t) is modeled as zero-mean Gaussian white noise with variance Λ. The
following figure illustrates an IDLTI model.

IDLTI
H

u y
G

1-11
1 Choosing Your System Identification Approach

When you simulate an IDLTI model, you study the effect of input u(t) (and possibly
initial conditions) on the output y(t). The noise e(t) is not considered. However, with
finite-horizon prediction of the output, both the measured and the noise components of
the model contribute towards computation of the (predicted) response.

u
H-1G
y_predicted

y_measured
1-H -1

One-step ahead prediction model corresponding to a linear identified model (y = Gu+He)

Measured and Noise Component Parameterizations

The various linear model structures provide different ways of parameterizing the
transfer functions G and H. When you construct an IDLTI model or estimate a model
directly using input-output data, you can configure the structure of both G and H, as
described in the following table:

Model Type Transfer Functions G and H Configuration Method

State space Represents an identified state-space Construction: Use idss to create a model,
model model structure, governed by the specifying values of state-space matrices
(idss) equations: A, B, C, D and K as input arguments
(using NaNs to denote unknown entries).
x& = Ax + Bu + Ke
y = Cx + Du + e Estimation: Use ssest or n4sid,
specifying name-value pairs for various
configurations, such as, canonical
where the transfer function between parameterization of the measured
the measured input u and output dynamics ('Form'/'canonical'),
y is G ( s) = C( sI - A) -1 B + D and denoting absence of feedthrough by fixing
the noise transfer function is D to zero ('Feedthrough'/false), and
absence of noise dynamics by fixing K to
H ( s) = C( sI - A) -1 K + I . zero ('DisturbanceModel'/'none').

1-12
 About Identified Linear Models

Model Type Transfer Functions G and H Configuration Method

Polynomial Represents a polynomial model such Construction: Use idpoly to create a
model as ARX, ARMAX and BJ. An ARMAX model using values of active polynomials
(idpoly) model, for example, uses the input-output as input arguments. For example, to
equation Ay(t) = Bu(t)+Ce(t), so that create an Output-Error model which uses
the measured transfer function G is G = B/F as the measured component and
has a trivial noise component (H = 1).
G ( s) = A -1 B , while the noise transfer
enter:
function is H ( s) = A -1C .
y = idpoly([],B,[],[],F)
The ARMAX model is a special Estimation: Use the armax, arx, or bj,
configuration of the general polynomial
specifying the orders of the polynomials
model whose governing equation is:
as input arguments. For example, bj
requires you to specify the orders of the
Ay(t) =
B C
u(t) + e(t) B, C, D, and F polynomials to construct a
F D model with governing equation

The autoregressive component, A, is B C

common between the measured and noise y( t) =
F
u( t) + e(t)
D
components. The polynomials B and F
constitute the measured component while
the polynomials C and D constitute the
noise component.
Transfer Represents an identified transfer function Construction: Use idtf to create
function model, which has no dynamic elements a model, specifying values of the
model to model noise behavior. This object uses numerator and denominator coefficients
(idtf) the trivial noise model H(s) = I. The as input arguments. The numerator
governing equation is and denominator vectors constitute the
measured component G = num(s)/
num den(s). The noise component is fixed to H
y( t) = u(t) + e(t)
den = 1.

Estimation: Use tfest, specifying the

number of poles and zeros of the measured
component G.

1-13
1 Choosing Your System Identification Approach

Model Type Transfer Functions G and H Configuration Method

Process Represents a process model, which For process (and grey-box) models,
model provides options to represent the noise the noise component is often treated
(idproc) dynamics as either first- or second- as an on-demand extension to an
order ARMA process (that is, H(s)= otherwise measured component-centric
C(s)/A(s), where C(s) and A(s) are representation. For these models, you
monic polynomials of equal degree). The can add a noise component by using the
measured component, G(s), is represented DisturbanceModel estimation option.
by a transfer function expressed in pole- For example:
zero form.
model = procest(data,'P1D')

estimates a model whose equation is:

1
y( s) = K p e-sTd u( s) + e(s).
( Tp1 s + 1)

To add a second order noise component to

the model, use:
Options = procestOptions(‘DisturbanceModel’, ‘AR
model = procest(data, ‘P1D’, Options);

This model has the equation:

1 1 + c1 s
y( s) = K p e-sTd u( s) + e( s)
( Tp1 s + 1) 1 + d1 s

where the coefficients c1 and d1

parameterize the noise component of the
model. If you are constructing a process
model using the idproc command, specify
the structure of the measured component
using the Type input argument and the
noise component by using the NoiseTF
name-value pair. For example,
model=idproc('P1','Kp',1,'Tp1',1,'NoiseTF',struc

1-14
 About Identified Linear Models

Model Type Transfer Functions G and H Configuration Method

creates the process model y(s) = 1/(s+1)
u(s) + (s + 0.1)/(s + 0.5) e(s)

Sometimes, fixing coefficients or specifying bounds on the parameters are not sufficient.
For example, you may have unrelated parameter dependencies in the model or
parameters may be a function of a different set of parameters that you want to identify
exclusively. For example, in a mass-spring-damper system, the A and B parameters
both depend on the mass of the system. To achieve such parameterization of linear
models, you can use grey-box modeling where you establish the link between the actual
parameters and model coefficients by writing an ODE file. To learn more, see “Grey-Box
Model Estimation”.

Linear Model Estimation

You typically use estimation to create models in System Identification Toolbox. You
execute one of the estimation commands, specifying as input arguments the measured
data, along with other inputs necessary to define the structure of a model. To illustrate,
the following example uses the state-space estimation command, ssest, to create a state
space model. The first input argument data specifies the measured input-output data.
The second input argument specifies the order of the model.
sys = ssest(data, 4)

The estimation function treats the noise variable e(t) as prediction error – the
residual portion of the output that cannot be attributed to the measured inputs. All
estimation algorithms work to minimize a weighted norm of e(t) over the span of
available measurements. The weighting function is defined by the nature of the noise
transfer function H and the focus of estimation, such as simulation or prediction error
minimization.

• “Black Box (“Cold Start”) Estimation” on page 1-15

• “Structured Estimations” on page 1-16
• “Estimation Options” on page 1-17
• “Estimation Report” on page 1-17

Black Box (“Cold Start”) Estimation

In a black-box estimation, you only have to specify the order to configure the structure of
the model.

1-15
1 Choosing Your System Identification Approach

sys = estimator(data, orders)

where estimator is the name of an estimation command to use for the desired model
type.

For example, you use tfest to estimate transfer function models, arx for ARX-structure
polynomial models, and procest for process models.

The first argument, data, is time- or frequency domain data represented as an iddata
or idfrd object. The second argument, orders, represents one or more numbers whose
definitions depends upon the model type:

• For transfer functions, orders refers to the number of poles and zeros.
• For state-space models, orders is a scalar that refers to the number of states.
• For process models, orders is a string denoting the structural elements of a process
model, such as, the number of poles and presence of delay and integrator.

When working with the app, you specify the orders in the appropriate edit fields of
corresponding model estimation dialogs.

Structured Estimations

In some situations, you want to configure the structure of the desired model more closely
than what is achieved by simply specifying the orders. In such cases, you construct a
template model and configure its properties. You then pass that template model as an
input argument to the estimation commands in place of orders.

To illustrate, the following example assigns initial guess values to the numerator and the
denominator polynomials of a transfer function model, imposes minimum and maximum
bounds on their estimated values, and then passes the object to the estimator function.

% initial guess for numerator

num = [1 2] den = [1 2 1 1]
% initial guess for the denominator
sys = idtf(num, den);
% set min bound on den coefficients to 0.1
sys.Structure.den.Minimum = [1 0.1 0.1 0.1];
sysEstimated = tfest(data, sys);

The estimation algorithm uses the provided initial guesses to kick-start the estimation
and delivers a model that respects the specified bounds.

1-16
 About Identified Linear Models

You can use such a model template to also configure auxiliary model properties such
as input/output names and units. If the values of some of the model’s parameters are
initially unknown, you can use NaNs for them in the template.

Estimation Options

There are many options associated with a model’s estimation algorithm that
configure the estimation objective function, initial conditions and numerical search
algorithm, among other things. For every estimation command, estimator, there
is a corresponding option command named estimatorOptions. To specify options
for a particular estimator command, such as tfest, use the options command that
corresponds to the estimation command, in this case, tfestOptions. The options
command returns an options set that you then pass as an input argument to the
corresponding estimation command.

For example, to estimate an Output-Error structure polynomial model, you use oe.
To specify simulation as the focus and lsqnonlin as the search method, you use
oeOptions:
load iddata1 z1
Options = oeOptions('Focus','simulation','SearchMethod','lsqnonlin');
sys= oe(z1, [2 2 1], Options);

For information about how to view the list of options used to create an estimated model,
see “Estimation Report” on page 1-17.

Estimation Report

The model returned by the estimation command contains information about the
estimation operation in the Report property. Information includes:

• Information on initial conditions used for estimation

• Termination conditions for iterative estimation algorithms under a category called
Termination
• Final prediction error (FPE), fit % and mean-square error (MSE) under a category
called Fit
• The nature of estimation data under a category called DataUsed
• All estimated quantities - parameter values and their covariance, initial state values
for state-space models and values of input level for process models under a category
called Parameters

1-17
1 Choosing Your System Identification Approach

• The option set used for estimation under a category called OptionsUsed

To illustrate, this example displays the contents of the Report.OptionsUsed property

of the model estimated in “Estimation Options” on page 1-17.

sys.Report

Status: 'Estimated using POLYEST with Focus = "prediction"'

Method: 'OE'
InitialCondition: 'zero'
Fit: [1x1 struct]
Parameters: [1x1 struct]
OptionsUsed: [1x1 idoptions.polyest]
RandState: [1x1 struct]
DataUsed: [1x1 struct]
Termination: [1x1 struct]

Exploring one of the categories provides additional data:

sys.Report.Fit shows:

FitPercent: 70.5666
FPE: 1.7558
LossFcn: 1.7094
MSE: 1.6865

1-18
 Linear Model Structures

Linear Model Structures

In this section...
“About System Identification Toolbox Model Objects” on page 1-19
“Available Linear Models” on page 1-20
“When to Construct a Model Structure Independently of Estimation” on page 1-22
“Commands for Constructing Model Structures” on page 1-22
“Model Properties” on page 1-24
“See Also” on page 1-26

About System Identification Toolbox Model Objects

Objects are instances of model classes. Each class is a blueprint that defines the following
information about your model:

• How the object stores data

• Which operations you can perform on the object

This toolbox includes nine classes for representing models. For example, idpoly
represents linear input-output polynomial models, and idss represents linear state-
space models. For a complete list of available model objects, see “Available Linear
Models” on page 1-20 and “Available Nonlinear Models” on page 1-29.

Model properties define how a model object stores information. Model objects store
information about a model, including the mathematical form of a model, names of input
and output channels, units, names and values of estimated parameters, parameter
uncertainties, algorithm specifications, and estimation information. For example, an
idpoly model has an InputName property for storing one or more input channel names.
Different model objects have different properties.

The allowed operations on an object are called methods. In System Identification Toolbox
software, some methods have the same name but apply to multiple model objects. For
example, bode creates a bode plot for all linear model objects. However, other methods
are unique to a specific model object. For example, the estimation method canon is
unique to the state-space idss model.

Every class has a special method for creating objects of that class, called the constructor.
Using a constructor creates an instance of the corresponding class or instantiates the
object. The constructor name is the same as the class name. For example, idpoly is both

1-19
1 Choosing Your System Identification Approach

the name of the class representing linear black-box polynomial models and the name of
the constructor for instantiating the model object.

Available Linear Models

A linear model is often sufficient to accurately describe the system dynamics and, in
most cases, you should first try to fit linear models. Available linear structures include
transfer functions and state-space models, summarized in the following table.

Model Type Usage Learn More

Transfer function (idtf) Use this structure to “Identifying Transfer
represent transfer Function Models”
functions:

num
y= u+ e
den

where num and den are

polynomials of arbitrary
lengths. You can specify
initial guesses for, and
estimate, num, den, and
transport delays.
Process model (idproc) Use this structure to “Identifying Process Models”
represent process models on page 3-24
that are low order transfer
functions expressed in pole-
zero form. They include
integrator, delay, zero, and
up to 3 poles.
State-space model (idss) Use this structure to “Identifying State-Space
represent known state- Models”
space structures and black-
box structures. You can
fix certain parameters to
known values and estimate
the remaining parameters.
You can also prescribe

1-20
 Linear Model Structures

Model Type Usage Learn More

minimum/maximum
bounds on the values
of the free parameters.
If you need to specify
parameter dependencies
or parameterize the state-
space matrices using your
own parameters, use a grey-
box model.
Polynomial models Use to represent linear “Identifying Input-Output
(idpoly) transfer functions based Polynomial Models” on page
on the general form input- 3-40
output polynomial form:

B C
Ay = u+ e
F D

where A, B, C, D and F
are polynomials with
coefficients that the toolbox
estimates from data.

Typically, you begin

modeling using simpler
forms of this generalized
structure (such as
ARX: Ay = Bu + e and
B
OE: y = u + e ) and, if
F
necessary, increase the
model complexity.

1-21
1 Choosing Your System Identification Approach

Model Type Usage Learn More

Grey-box model (idgrey) Use to represent arbitrary “Estimating Linear Grey-Box
parameterizations of Models” on page 6-6
state-space models. For
example, you can use this
structure to represent your
ordinary differential or
difference equation (ODE)
and to define parameter
dependencies.

When to Construct a Model Structure Independently of Estimation

You use model constructors to create a model object at the command line by specifying all
required model properties explicitly.

You must construct the model object independently of estimation when you want to:

• Simulate or analyze the effect of model's parameters on its response, independent of

estimation.
• Specify an initial guess for specific model parameter values before estimation. Specify
bounds on parameter values, or set up the auxiliary model information in advance,
or both. Auxiliary model information includes specifying input/output names, units,
Notes, user data, etc.

In most cases, you can use the estimation commands to both construct and estimate
the model—without having to construct the model object independently. For example,
the estimation command tfest creates a transfer function model using data and some
information on the order of the model - the number of poles and zeros. For information
about how to both construct and estimate models with a single command, see “Model
Estimation Commands” on page 1-37.

In case of grey-box models, you must always construct the model object first and then
estimate the parameters of the ordinary differential or difference equation.

Commands for Constructing Model Structures

The following table summarizes the model constructors available in the System
Identification Toolbox product for representing various types of models.

1-22
 Linear Model Structures

After model estimation, you can recognize the corresponding model objects in the
MATLAB Workspace browser by their class names. The name of the constructor matches
the name of the object it creates.

For information about how to both construct and estimate models with a single
command, see “Model Estimation Commands” on page 1-37.

Summary of Model Constructors

Model Constructor Resulting Model Class

idfrd Nonparametric frequency-response model.
idproc Continuous-time, low-order transfer functions (process
models).
idpoly Linear input-output polynomial models:

• ARX
• ARMAX
• Output-Error
• Box-Jenkins
idss Linear state-space models.
idtf Linear transfer function models.
idgrey Linear ordinary differential or difference equations
(grey-box models). You write a function that translates
user parameters to state-space matrices. Can also
be viewed as state-space models with user-specified
parameterization.
idnlgrey Nonlinear ordinary differential or difference equation
(grey-box models). You write a function or MEX-file to
represent the governing equations.
idnlarx Nonlinear ARX models, which define the predicted
output as a nonlinear function of past inputs and
outputs.
idnlhw Nonlinear Hammerstein-Wiener models, which
include a linear dynamic system with nonlinear static
transformations of inputs and outputs.

1-23
1 Choosing Your System Identification Approach

For more information about when to use these commands, see “When to Construct a
Model Structure Independently of Estimation” on page 1-22.

Model Properties
• “Categories of Model Properties” on page 1-24
• “Viewing Model Properties and Estimated Parameters” on page 1-25

Categories of Model Properties

The way a model object stores information is defined by the properties of the
corresponding model class.

Each model object has properties for storing information that are relevant only to that
specific model type. The idtf, idgrey, idpoly, idproc, and idss model objects are
based on the idlti superclass and inherit all idlti properties.

Similarly, the nonlinear models idnlarx, idnlhw, and idnlgrey are based on the
idnlmodel superclass and inherit all idnlmodel properties.

In general, all model objects have properties that belong to the following categories:

• Names of input and output channels, such as InputName and OutputName

• Sampling interval of the model, such as Ts
• Units for time or frequency
• Model order and mathematical structure (for example, ODE or nonlinearities)
• Properties that store estimation results (Report)
• User comments, such as Notes and Userdata
• Estimation algorithm information (nonlinear models only)

• Algorithm

Structure includes fields that specify the estimation method. Algorithm includes
another structure, called Advanced, which provides additional flexibility for
setting the search algorithm. Different fields apply to different estimation
techniques.

For linear parametric models, Algorithm specifies the frequency weighing of the
estimation using the Focus property.

For information about getting help on object properties, see the model reference pages.

1-24
 Linear Model Structures

Viewing Model Properties and Estimated Parameters

The following table summarizes the commands for viewing and changing model property
values. Property names are not case sensitive. You do not need to type the entire
property name if the first few letters uniquely identify the property.
Task Command Example
View all model get Load sample data, compute an ARX model, and
properties and list the model properties:
their values
load iddata8
m_arx=arx(z8,[4 3 2 3 0 0 0]);
get(m_arx)
Access a specific Use dot notation View the A matrix containing the estimated
model property parameters in the previous model:
m_arx.a
For properties, such as View the method used in ARX model estimation:
Report, that are configured
like structures, use dot m_arx.Report.Method
notation of the form
model.PropertyName.FieldName.
FieldName is the name of
any field of the property.
Change model dot notation Change the input delays for all three input
property values channels to [1 1 1] for an ARX model:

m_arx.InputDelay = [1 1 1]
Access model Use getpar, getpvec and • View a table of all parameter attributes:
parameter values getcov
and uncertainty See Also: polydata, getpar(m_arx)
information idssdata, tfdata, • View the A polynomial and 1 standard
zpkdata uncertainty of an ARX model:
[a,~,~,~,~,da] = polydata(m_arx)
Set model Use setpar, setpvec and • Set default parameter labels:
property values setcov
and uncertainty m_arx = setpar(m_arx,'label','default')
information • Set parameter covariance data:
m_arx = setcov(m_arx,cov)

1-25
1 Choosing Your System Identification Approach

See Also
Validate each model directly after estimation to help fine-tune your modeling strategy.
When you do not achieve a satisfactory model, you can try a different model structure
and order, or try another identification algorithm. For more information about validating
and troubleshooting models, see “Validating Models After Estimation”.

1-26
 Imposing Constraints on Model Parameter Values

Imposing Constraints on Model Parameter Values

All identified linear (IDLTI) models, except idfrd, contain a Structure property. The
Structure property contains the adjustable entities (parameters) of the model. Each
parameter has attributes such as value, minimum/maximum bounds, and free/fixed
status that allow you to constrain them to desired values or a range of values during
estimation. You use the Structure property to impose constraints on the values of
various model parameters.

The Structure property contains the essential parameters that define the structure of a
given model:

• For identified transfer functions, includes the numerator, denominator and delay
parameters
• For polynomial models, includes the list of active polynomials
• For state-space models, includes the list of state-space matrices

For information about other model types, see the model reference pages.

For example, the following example constructs an idtf model, specifying values for the
num and den parameters:

num = [1 2];
den = [1 2 2];
sys = idtf(num,den)

You can update the value of the num and den properties after you create the object as
follows:

new_den = [1 1 10];
sys.den = new_den;

To fix the denominator to the value you specified (treat its coefficients as fixed
parameters), use the Structure property of the object as follows:

sys.Structure.den.Value = new_den;
sys.Structure.den.Free = false(1,3);

For a transfer function model, the num, den, and ioDelay model properties are simply
pointers to the Value attribute of the corresponding parameter in the Structure
property.

1-27
1 Choosing Your System Identification Approach

IDTF Model Properties Parameters

num double vector num:

Value: double vector
den double vector Minimum: double vector
Maximum: double vector
Free: logical vector
ioDelay scalar
Scale: double vector
Info: struct
Structure
den:
InputDelay scalar Value:
Minimum:
Maximum:
Ts scalar Free:
Scale:
. . Info:

ioDelay:
. .
Value:
Minimum:
. . Maximum:
Free:
Scale:
Info:

Similar relationships exist for other model structures. For example, the a property of a
state-space model contains the double value of the state matrix. It is an alias to the A
parameter value stored in Structure.a.Value.

1-28
 Available Nonlinear Models

Available Nonlinear Models

System Identification Toolbox provides several nonlinear black-box model structures,
which have traditionally been useful for representing dynamic systems.

Model Type Usage Learn More

Nonlinear ARX model Use to represent nonlinear “Identifying Nonlinear ARX
(idnlarx object) extensions of linear models. Models” on page 5-7
This structure allows you
to model complex nonlinear
behavior using flexible
nonlinear functions, such
as wavelet and sigmoid
networks.
Linear models with input/ Use to represent linear “Identifying Hammerstein-
output nonlinearities, or models with static Wiener Models” on page
Hammerstein-Wiener model nonlinearities. 5-48
(idnlhw object)
Nonlinear grey-box model Use to represent nonlinear “Estimating Nonlinear Grey-
(idnlgrey object) ODEs with unknown Box Models” on page 6-15
parameters.

1-29
1 Choosing Your System Identification Approach

Recommended Model Estimation Sequence

System identification is an iterative process, where you identify models with different
structures from data and compare model performance. You start by estimating the
parameters of simple model structures. If the model performance is poor, you gradually
increase the complexity of the model structure. Ultimately, you choose the simplest
model that best describes the dynamics of your system.

Another reason to start with simple model structures is that higher-order models are
not always more accurate. Increasing model complexity increases the uncertainties in
parameter estimates and typically requires more data (which is common in the case of
nonlinear models).

Note: Model structure is not the only factor that determines model accuracy. If your
model is poor, you might need to preprocess your data by removing outliers or filtering
noise. For more information, see “Ways to Prepare Data for System Identification”.

Estimate impulse-response and frequency-response models first to gain insight into

the system dynamics and assess whether a linear model is sufficient. Then, estimate
parametric models in the following order:

1 Transfer function, ARX polynomial and state-space models provide the simplest
structures. Estimation of ARX and state-space models let you determine the model
orders.

In the System Identification app. Choose to estimate the Transfer function

models, ARX polynomial models and the state-space model using the n4sid method.

At the command line. Use the tfest, arx, and the n4sid commands,
respectively.

For more information, see “Input-Output Polynomial Models” and “State-Space

Models”.
2 ARMAX and BJ polynomial models provide more complex structures and require
iterative estimation. Try several model orders and keep the model orders as low as
possible.

In the System Identification app. Select to estimate the BJ and ARMAX

polynomial models.

1-30
 Recommended Model Estimation Sequence

At the command line. Use the bj or armax commands.

For more information, see “Identifying Input-Output Polynomial Models”.

3 Nonlinear ARX or Hammerstein-Wiener models provide nonlinear structures. For
more information, see “Nonlinear Model Identification”.

For general information about choosing you model strategy, see “System Identification
Overview”. For information about validating models, see “Validating Models After
Estimation”.

1-31
1 Choosing Your System Identification Approach

Supported Models for Time- and Frequency-Domain Data

In this section...
“Supported Models for Time-Domain Data” on page 1-32
“Supported Models for Frequency-Domain Data” on page 1-33
“See Also” on page 1-34

Supported Models for Time-Domain Data

Continuous-Time Models

You can directly estimate the following types of continuous-time models:

• Transfer function models. See “Identifying Transfer Function Models”.

• Process models. See “Identifying Process Models” on page 3-24.
• State-space models. See “Identifying State-Space Models”.

You can also use d2c to convert an estimated discrete-time model into a continuous-time
model.

Discrete-Time Models

You can estimate all linear and nonlinear models supported by the System Identification
Toolbox product as discrete-time models, except process models, which are defined only in
continuous-time..

ODEs (Grey-Box Models)

You can estimate both continuous-time and discrete-time models from time-domain data
for linear and nonlinear differential and difference equations.

Nonlinear Models

You can estimate discrete-time Hammerstein-Wiener and nonlinear ARX models from
time-domain data.

You can also estimate nonlinear grey-box models from time-domain data. See
“Estimating Nonlinear Grey-Box Models” on page 6-15.

1-32
 Supported Models for Time- and Frequency-Domain Data

Supported Models for Frequency-Domain Data

There are two types of frequency-domain data:

• Frequency response data

• Frequency domain input/output signals which are Fourier Transforms of the
corresponding time domain signals.

The data is considered continuous-time if its sample time (Ts) is 0, and is considered
discrete-time if the sample time is nonzero.

Continuous-Time Models

You can estimate the following types of continuous-time models directly:

• Transfer function models using continuous- or discrete-time data. See “Identifying

Transfer Function Models”.
• Process models using continuous- or discrete-time data. See “Identifying Process
Models” on page 3-24.
• Input-output polynomial models of output-error structure using continuous time data.
See “Identifying Input-Output Polynomial Models” on page 3-40.
• State-space models using continuous- or discrete-time data.

From continuous-time frequency-domain data, you can only estimate continuous-time

models.

You can also use d2c to convert an estimated discrete-time model into a continuous-time
model.

Discrete-Time Models

You can estimate all linear model types supported by the System Identification Toolbox
product as discrete-time models, except process models, which are defined in continuous-
time only. For estimation of discrete-time models, you must use discrete-time data.

The noise component of a model cannot be estimated using frequency domain data, with
the exception of ARX models. Thus, the K matrix of an identified state-space model, the
noise component, is zero. An identified polynomial model has output-error (OE) or ARX
structure; BJ/ARMAX or other polynomial structure with nontrivial values of C or D
polynomials cannot be estimated.

1-33
1 Choosing Your System Identification Approach

ODEs (Grey-Box Models)

For linear grey-box models, you can estimate both continuous-time and discrete-time
models from frequency-domain data. The noise component of the model, the K matrix,
cannot be estimated using frequency domain data; it remains fixed to 0.

Nonlinear grey-box models are supported only for time-domain data.

Nonlinear Black-Box Models

Nonlinear black box (nonlinear ARX and Hammerstein-Wiener models) cannot be

estimated using frequency domain data.

See Also
“Supported Continuous- and Discrete-Time Models” on page 1-35

1-34
 Supported Continuous- and Discrete-Time Models

Supported Continuous- and Discrete-Time Models

For linear and nonlinear ODEs (grey-box models), you can specify any ordinary
differential or difference equation to represent your continuous-time or discrete-time
model in state-space form, respectively. In the linear case, both time-domain and
frequency-domain data are supported. In the nonlinear case, only time-domain data is
supported.

For black-box models, the following tables summarize supported continuous-time and
discrete-time models.

Supported Continuous-Time Models

Model Type Description

“Identifying Transfer Function Estimate continuous-time transfer function models directly
Models” using tfest from either time- and frequency-domain data.
If you estimated a discrete-time transfer function model
from time-domain data, then use d2c to transform it into a
continuous-time model.
“Low-order transfer functions Estimate low-order process models for up to three free poles
(process models)” from either time- or frequency-domain data.
“Linear input-output polynomial To get a linear, continuous-time model of arbitrary structure
models” from time-domain data, you can estimate a discrete-time
model, and then use d2c to transform it into a continuous-time
model.
You can estimate only polynomial models of Output Error
structure using continuous-time frequency domain data..
Other structures that include noise models, such as Box-
Jenkins (BJ) and ARMAX, are not supported for frequency-
domain data.
“State-space models” Estimate continuous-time state-space models directly using
the estimation commands from either time- and frequency-
domain data.
If you estimated a discrete-time state-space model from time-
domain data, then use d2c to transform it into a continuous-
time model.

1-35
1 Choosing Your System Identification Approach

Model Type Description

“Linear ODEs (grey-box) models” If the MATLAB file returns continuous-time model matrices,
then estimate the ordinary differential equation (ODE)
coefficients using either time- or frequency-domain data.
“Nonlinear ODEs (grey-box) If the MATLAB file returns continuous-time output and state
models” derivative values, estimate arbitrary differential equations
(ODEs) from time-domain data.

Supported Discrete-Time Models

Model Type Description

“Linear, input-output polynomial Estimate arbitrary-order, linear parametric models from time-
models” or frequency-domain data.
To get a discrete-time model, your data sampling interval
must be set to the (nonzero) value you used to sample in your
experiment.
“Nonlinear Model Identification” Estimate from time-domain data only.
“Linear ODEs (grey-box) models” If the MATLAB file returns discrete-time model matrices, then
estimate ordinary difference equation coefficients from time-
domain or discrete-time frequency-domain data.
“Nonlinear ODEs (grey-box) If the MATLAB file returns discrete-time output and state
models” update values, estimate ordinary difference equations from
time-domain data.

1-36
 Model Estimation Commands

Model Estimation Commands

In most cases, a model can be created by using a model estimation command on a
dataset. For example, SSEST(DATA, Nx) creates a continuous-time state-space model of
order Nx using the input/output of frequency response data DATA.

Note: For ODEs (grey-box models), you must first construct the model structure and then
apply an estimation command (either greyest or pem) to the resulting model object.

The following table summarizes System Identification Toolbox estimation commands. For
detailed information about using each command, see the corresponding reference page.

Commands for Constructing and Estimating Models

Model Type Estimation Commands

Transfer function models tfest
Process models procest
Linear input-output polynomial armax (ARMAX only)
models arx (ARX only)
bj (BJ only)
iv4 (ARX only)
oe (OE only)
polyest (for all models)
State-space models n4sid
ssest
Time-series models ar
arx (for multiple outputs)
ivar
nlarx(for nonlinear time-series models)
Nonlinear ARX models nlarx
Hammerstein-Wiener models nlhw

1-37
1 Choosing Your System Identification Approach

Modeling Multiple-Output Systems

In this section...
“About Modeling Multiple-Output Systems” on page 1-38
“Modeling Multiple Outputs Directly” on page 1-38
“Modeling Multiple Outputs as a Combination of Single-Output Models” on page
1-39
“Improving Multiple-Output Estimation Results by Weighing Outputs During
Estimation” on page 1-39

About Modeling Multiple-Output Systems

You can estimate multiple-output model directly using all the measured inputs and
outputs, or you can try building models for subsets of the most important input and
output channels. To learn more about each approach, see:

• “Modeling Multiple Outputs Directly” on page 1-38

• “Modeling Multiple Outputs as a Combination of Single-Output Models” on page
1-39

Modeling multiple-output systems is more challenging because input/output couplings

require additional parameters to obtain a good fit and involve more complex models.
In general, a model is better when more data inputs are included during modeling.
Including more outputs typically leads to worse simulation results because it is harder to
reproduce the behavior of several outputs simultaneously.

If you know that some of the outputs have poor accuracy and should be less important
during estimation, you can control how much each output is weighed in the estimation.
For more information, see “Improving Multiple-Output Estimation Results by Weighing
Outputs During Estimation” on page 1-39.

Modeling Multiple Outputs Directly

You can perform estimation with linear and nonlinear models for multiple-output data.

Tip Estimating multiple-output state-space models directly generally produces better

results than estimating other types of multiple-output models directly.

1-38
 Modeling Multiple-Output Systems

Modeling Multiple Outputs as a Combination of Single-Output Models

You may find that it is harder for a single model to explain the behavior of several
outputs. If you get a poor fit estimating a multiple-output model directly, you can try
building models for subsets of the most important input and output channels.

Use this approach when no feedback is present in the dynamic system and there are no
couplings between the outputs. If you are unsure about the presence of feedback, see
“How to Analyze Data Using the advice Command”.

To construct partial models, use subreferencing to create partial data sets, such that each
data set contains all inputs and one output. For more information about creating partial
data sets, see the following sections in the System Identification Toolbox User's Guide:

• For working in the System Identification app, see “Create Data Sets from a Subset of
Signal Channels” on page 2-33.
• For working at the command line, see the “Select Data Channels, I/O Data and
Experiments in iddata Objects”.

After validating the single-output models, use vertical concatenation to combine

these partial models into a single multiple-output model. For more information about
concatenation, see “Increasing Number of Channels or Data Points of iddata Objects” or
“Adding Input or Output Channels in idfrd Objects”.

You can try refining the concatenated multiple-output model using the original (multiple-
output) data set.

Improving Multiple-Output Estimation Results by Weighing Outputs

During Estimation
When estimating linear and nonlinear black-box models for multiple-output systems, you
can control the relative importance of output channels during the estimation process.
The ability to control how much each output is weighed during estimation is useful when
some of the measured outputs have poor accuracy or should be treated as less important
during estimation. For example, if you have already modeled one output well, you might
want to focus the estimation on modeling the remaining outputs. Similarly, you might
want to refine a model for a subset of outputs.

For linear models, use the OutputWeight estimation to indicate the desired output
weighting. If you set this option to 'noise', an automatic weighting, equal to the

1-39
1 Choosing Your System Identification Approach

inverse of the estimated noise variance, is used for model estimation. You can also specify
a custom weighting matrix, which must be a positive semi-definite matrix.

Note:

• The OutputWeight option is not available for polynomial models, except ARX
models, since their estimation algorithm estimates the parameters one output at a
time.

• Transfer function (idtf) and process models (idproc) ignore OutputWeight

when they contain nonzero or free transport delays. In the presence of delays, the
estimation is carried out one output at a time.

For nonlinear models, you can specify output weights directly in the estimation command
using the Criterion and Weighting estimation options. You must set the Criterion
field to Trace, and set the Weighting field to the matrix that contains the output
weights. The Trace criterion minimizes the weighted sum of the prediction errors using
the weights specified by Weighting.

For more information about the OutputWeight estimation option, see arxOptions,
ssestOptions, tfestOptions, procestOptions, etc. For more information about
the Algorithm fields for nonlinear estimation, see the idnlarx and idnlhw reference
pages.

Note: For multiple-output idnlarx models containing neuralnet or treepartition

nonlinearity estimators, output weighting is ignored because each output is estimated
independently.

1-40
 Regularized Estimates of Model Parameters

Regularized Estimates of Model Parameters

In this section...
“What Is Regularization?” on page 1-41
“When to Use Regularization” on page 1-44
“Choosing Regularization Constants” on page 1-46

What Is Regularization?
Regularization is the technique for specifying constraints on a model’s flexibility, thereby
reducing uncertainty in the estimated parameter values.

Model parameters are obtained by fitting measured data to the predicted model response,
such as a transfer function with three poles or a second-order state-space model.
The model order is a measure of its flexibility — higher the order, the greater the
flexibility. For example, a model with three poles is more flexible than one with two
poles. Increasing the order causes the model to fit the observed data with increasing
accuracy. However, the increased flexibility comes with the price of higher uncertainty
in the estimates, measured by a higher value of random or variance error. On the other
hand, choosing a model with too low an order leads to larger systematic errors. Such
errors cannot be attributed to measurement noise and are also known as bias error.

Ideally, the parameters of a good model should minimize the mean square error (MSE),
given by a sum of systematic error (bias) and random error (variance):

MSE = |Bias|2 + Variance

The minimization is thus a tradeoff in constraining the model. A flexible (high order)
model gives small bias and large variance, while a simpler (low order) model results
in larger bias and smaller variance errors. Typically, you can investigate this tradeoff
between bias and variance errors by cross-validation tests on a set of models of
increasing flexibility. However, such tests do not always give full control in managing the
parameter estimation behavior. For example:

• You cannot use the known (a priori) information about the model to influence the
quality of the fits.
• In grey-box and other structured models, the order is fixed by the underlying ODEs
and cannot be changed. If the data is not rich enough to capture the full range of
dynamic behavior, this typically leads to high uncertainty in the estimated values.

1-41
1 Choosing Your System Identification Approach

• Varying the model order does not let you explicitly shape the variance of the
underlying parameters.

Regularization gives you a better control over the bias versus variance tradeoff by
introducing an additional term in the minimization criterion that penalizes the model
flexibility. Without regularization, the parameter estimates are obtained by minimizing a
weighted quadratic norm of the prediction errors ε(t,θ):

N
VN (q ) = Âe 2 (t,q )
t =1

where t is the time variable, N is the number of data samples, and ε(t,θ) is the predicted
error computed as the difference between the observed output and the predicted output
of the model.

Regularization modifies the cost function by adding a term proportional to the square of
the norm of the parameter vector θ, so that the parameters θ are obtained by minimizing:

N
2
Vˆ N (q ) = Âe 2 (t,q ) +   l q
t =1

where λ is a positive constant that has the effect of trading variance error in VN(θ)
for bias error — the larger the value of λ, the higher the bias and lower the variance
of θ. The added term penalizes the parameter values with the effect of keeping their
values small during estimation. In statistics, this type of regularization is called ridge
regression. For more information, see “Ridge Regression” in the Statistics Toolbox™
documentation.

Note: Another choice for the norm of θ vector is the L1-norm, known as lasso
regularization. However, System Identification Toolbox supports only the 2-norm based
penalty, known as L2 regularization, as shown in the previous equation.

The penalty term is made more effective by using a positive definite matrix R, which
allows weighting and/or rotation of the parameter vector:

N
Vˆ N (q ) = Âe 2 (t,q ) +   lq T Rq
t =1

1-42
 Regularized Estimates of Model Parameters

The square matrix R gives additional freedom for:

• Shaping the penalty term to meet the required constraints, such as keeping the model
stable
• Adding known information about the model parameters, such as reliability of the
individual parameters in the θ vector

For structured models such as grey-box models, you may want to keep the estimated
parameters close to their guess values to maintain the physical validity of the estimated
T
model. This can be achieved by generalizing the penalty term to l q - q * ( ) R (q - q * ) ,
such that the cost function becomes:

N T
Vˆ N (q ) = Âe 2 (t,q ) + l (q - q * ) (
R q -q* )
t =1

Minimizing this cost function has the effect of estimating θ such that their values remain
close to initial guesses θ*.

In regularization:

• θ* represents prior knowledge about the unknown parameters.

• λ*R represents the confidence in the prior knowledge of the unknown parameters.
This implies that the larger the value, the higher the confidence.

A formal interpretation in a Bayesian setting is that θ has a prior distribution that is

Gaussian with mean θ* and covariance matrix s 2 / l R-1 , where σ2 is the variance of
ε(t). The use of regularization can therefore be linked to some prior information about the
system. This could be quite soft, such as the system is stable.

You can use the regularization variables λ and R as tools to find a good model that
balances complexity and provides the best tradeoff between bias and variance. You can
obtain regularized estimates of parameters for transfer function, state-space, polynomial,
grey-box, process and nonlinear black-box models. The three terms defining the penalty
term, λ, R and θ*, are represented by regularization options Lambda, R, and Nominal,
respectively in the toolbox. You can specify their values in the estimation options set
for linear models and the Algorithm property for nonlinear models. In the System
Identification app, click Regularization in the linear model estimation dialog box or
Estimation Options in the Nonlinear Models dialog box.

1-43
1 Choosing Your System Identification Approach

When to Use Regularization

Use regularization for:

• “Identifying Overparameterized Models” on page 1-44

• “Imposing A Priori Knowledge of Model Parameters in Structured Models” on page
1-45
• “Incorporating Knowledge of System Behavior in ARX and FIR Models” on page
1-45

Identifying Overparameterized Models

Over-parameterized models are rich in parameters. Their estimation typically yields

parameter values with a high level of uncertainty. Over-parameterization is quite
common for nonlinear ARX (idnlarx) models and could also be for linear state-space
models using free parameterization.

In such cases, regularization improves the numerical conditioning of the estimation. You
can explore bias-vs.-variance tradeoff using various values of the regularization constant
Lambda. Typically, the Nominal option is its default value of 0, and R is an identity
matrix such that the following cost function is minimized:

N
2
Vˆ N (q ) = Âe 2 (t,q ) +   l q
t =1

In the following example, a nonlinear ARX model estimation using a large number of
neurons leads to an ill-conditioned estimation problem.
% Load estimation data
load regularizationExampleData.mat nldata
% Estimate model without regularization
Orders = [1 2 1];
NL = sigmoidnet('NumberOfUnits',30);
sys = nlarx(nldata, Orders, NL);
compare(nldata, sys)

Applying even a small regularizing penalty produces a good fit for the model to the data.
% Estimate model using regularization constant # = 1e-8;
al = sys.Algorithm;
al.Regularization.Lambda = 1e-8;
sysr = nlarx(nldata, Orders, NL, 'Algorithm',al);

1-44
 Regularized Estimates of Model Parameters

compare(nldata, sysr)

Imposing A Priori Knowledge of Model Parameters in Structured Models

In models derived from differential equations, the parameters have physical significance.
You may have a good guess for typical values of those parameters even if the reliability
of the guess may be different for each parameter. Because the model structure is fixed in
such cases, you cannot simplify the structure to reduce variance errors.

Using the regularization constant Nominal, you can keep the estimated values close to
their initial guesses. You can also design R to reflect the confidence in the initial guesses
of the parameters. For example, if θ is a 2-element vector and you can guess the value of
the first element with more confidence than the second one, set R to be a diagonal matrix
of size 2-by-2 such that R(1,1) >> R(2,2).

In the following example, a model of a DC motor is parameterized by static gain G and

time constant τ. From prior knowledge, suppose you know that G is about 4 and τ is about
1. Also, assume that you have more confidence in the value of τ than G and would like to
guide the estimation to remain close to the initial guess.
% Load estimation data
load regularizationExampleData.mat motorData
% Create idgrey model for DC motor dynamics
mi = idgrey(@DCMotorODE,{'G', 4; 'Tau', 1},'cd',{}, 0);
mi = setpar(mi, 'label', 'default');
% Configure Regularization options
opt = greyestOptions;
opt.Regularization.Lambda = 100;
% Specify that the second parameter better known than the first
opt.Regularization.R = [1, 1000];
% Specify initial guess as Nominal
opt.Regularization.Nominal = 'model';
% Estimate model
sys = greyest(motorData, mi, opt)
getpar(sys)

Incorporating Knowledge of System Behavior in ARX and FIR Models

In many situations, you may know the shape of the system impulse response from impact
tests. For example, it is quite common for stable systems to have an impulse response
that is smooth and exponentially decaying. You can use such prior knowledge of system
behavior to derive good values of regularization constants for linear-in-parameter models
such as ARX and FIR structure models using the arxRegul command.

1-45
1 Choosing Your System Identification Approach

For black-box models of arbitrary structure, it is often difficult to determine the optimal
values of Lambda and R that yield the best bias-vs.-variance tradeoff. Therefore, it
is recommended that you start by obtaining the regularized estimate of an ARX or
FIR structure model. Then, convert the model to a state-space, transfer function or
polynomial model using the idtf, idss, or idpoly commands, followed by order
reduction if required.

In the following example, direct estimation of a 15th order continuous-time transfer

function model fails due to numerical ill-conditioning.
% Load estimation data
load dryer2
Dryer = iddata(y2,u2,0.08);
Dryerd = detrend(Dryer,0);
Dryerde = Dryerd(1:500);
xe = Dryerd(1:500);
ze = Dryerd(1:500);
zv = Dryerd(501:end);
% Estimate model without regularization
sys1 = tfest(ze, 15);

Therefore, use regularized ARX estimation and then convert the model to transfer
function structure.
% Specify regularization constants
[L, R] = arxRegul(ze, [15 15 1]);
optARX = arxOptions;
optARX.Regularization.Lambda = L;
optARX.Regularization.R = R;
% Estimate ARX model
sysARX = arx(ze, [15 15 1], optARX);
% Convert model to continuous time
sysc = d2c(sysARX);
% Convert model to transfer function
sys2 = idtf(sysc);
% Validate the models sys1 and sys2
compare(zv, sys1, sys2)

Choosing Regularization Constants

A guideline for selecting the regularization constants λ and R is in the Bayesian
interpretation. The added penalty term is an assumption that the parameter vector θ is a
Gaussian random vector with mean θ* and covariance matrix s 2 / l R-1 .

1-46
 Regularized Estimates of Model Parameters

You can relate naturally to such an assumption for a grey-box model, where the
parameters are of known physical interpretation. In other cases, this may be more
difficult. Then, you have to use ridge regression (R = 1; θ* = 0) and tune λ by trial and
error.

Use the following techniques for determining λ and R values:

• “Incorporate Prior Information Using Tunable Kernels” on page 1-47

• “Perform Cross-Validation Tests” on page 1-47

Incorporate Prior Information Using Tunable Kernels

Tuning the regularization constants for ARX models in arxRegul is based on simple
assumptions about the properties of the true impulse responses.

In the case of an FIR model, the parameter vector contains the impulse response
coefficients bk for the system. From prior knowledge of the system, it is often known that
the impulse response is smooth and exponentially decaying:

2
E [ bk ] = Cm k , corr {bkbk -1 } = r

where corr means correlation. The equation is a parameterization of the regularization

constants in terms of coefficients C, μ, and ρ and the chosen shape (decaying polynomial)
is called a kernel. The kernel thus contains information about parameterization of the
prior covariance of the impulse response coefficients.

You can estimate the parameters of the kernel by adjusting them to the measured data
using the RegulKernel input of the arxRegul command. For example, the DC kernel
estimates all three parameters while the TC kernel links r = m . This technique of
tuning kernels applies to all linear-in-parameter models such as ARX and FIR models.

Perform Cross-Validation Tests

A general way to test and evaluate any regularization parameters is to estimate a model
based on certain parameters on an estimation data set, and evaluate the model fit for
another validation data set. This is known as cross-validation.

Cross-validation is entirely analogous to the method for selecting model order:

1 Generate a list of candidate λ and R values to be tested.

1-47
1 Choosing Your System Identification Approach

2 Estimate a model for each candidate regularization constant set.

3 Compare the model fit to the validation data.
4 Use the constants that give the best fit to the validation data.

For example:
% Create estimation and validation data sets
ze = z(1:N/2);
zv = z(N/2:end);
% Specify regularization options and estimate models
opt = ssestOptions;
for tests = 1:M
opt.Regularization.Lambda = Lvalue(test);
opt.Regularization.R = Rvalue(test);
m{test} = ssest(ze,order,opt);
end
% Compare models with validation data for model fit
[~,fit] = compare(zv,m{:))

References
[1] L. Ljung. “Some Classical and Some New Ideas for Identification of Linear Systems.”
Journal of Control, Automation and Electrical Systems. April 2013, Volume 24,
Issue 1-2, pp 3-10.

[2] L. Ljung, and T. Chen. “What can regularization offer for estimation of dynamical
systems?” In Proceedings of IFAC International Workshop on Adaptation and
Learning in Control and Signal Processing, ALCOSP13, Caen, France, July 2013.

[3] L. Ljung, and T. Chen. “Convexity issues in system identification.” In Proceedings of

the 10th IEEE International Conference on Control & Automation, ICCA 2013,
Hangzhou, China, June 2013.

Related Examples
• “Estimate Regularized ARX Model Using System Identification App” on page
1-49
• “Regularized Identification of Dynamic Systems”

1-48
 Estimate Regularized ARX Model Using System Identification App

Estimate Regularized ARX Model Using System Identification App

This example shows how to estimate regularized ARX models using automatically
generated regularization constants in the System Identification app.

Open a saved System Identification App session.

filename = fullfile(matlabroot,'help','toolbox',...
'ident','examples','ex_arxregul.sid');
systemIdentification(filename)

The session imports the following data and model into the System Identification app:

• Estimation data eData

The data is collected by simulating a system with the following known transfer
function:

0 .02008 + 0 .04017 z-1 + 0.02008 z -2

G ( z) =
1 - 1.56 z-1 + 0.6414 z-2
• Transfer function model trueSys

trueSys is the transfer function model used to generate the estimation data eData
described previously. You also use the impulse response of this model later to compare
the impulse responses of estimated ARX models.

Estimate a 50th-order ARX model.

1 In the System Identification app, select Estimate > Polynomial Models to open
the Polynomial Models dialog box.

1-49
1 Choosing Your System Identification Approach

2 Verify that ARX is selected in the Structure list.

3 In the Orders field, specify [0 50 0] as the ARX model order and delay.

4 Click Estimate to estimate the model.

A model arx0500 is added to the System Identification app.

Estimate a 50th-order regularized ARX model.

1 In the Polynomial Models dialog box, click Regularization.

2 In the Regularization Options dialog box, select TC from the Regularization
Kernel drop-down list.

1-50
 Estimate Regularized ARX Model Using System Identification App

Specifying this option automatically determines regularization constants using the

TC regularization kernel. To learn more, see the arxRegul reference page.

Click Close to close the dialog box.

3 In the Name field in the Polynomial Models dialog box, type arx0500reg.
4 Click Estimate.

A model arx0500reg is added to the System Identification app.

Compare the unregularized and regularized model outputs to estimation data.

Select the Model output check box in the System Identification app.

1-51
1 Choosing Your System Identification Approach

The Measured and simulated model output plot shows that both the models have an 84%
fit with the data.

Determine if regularization leads to parameter values with less variance.

Because the model fit to the estimation data is similar with and without using
regularization, compare the impulse response of the ARX models with the impulse
responses of trueSys, the system used to collect the estimation data.

1 Click the trueSys icon in the model board of the System Identification app.

1-52
 Estimate Regularized ARX Model Using System Identification App

2 Select the Transient resp check box to open the Transient Response plot window.

By default, the plot shows the step response.

3 In the Transient response plot window, select Options > Impulse response to
change to plot to display the impulse response.
4 Select Options > Show 99% confidence intervals to plot the confidence intervals.

1-53
1 Choosing Your System Identification Approach

The plot shows that the impulse response of the unregularized model arx0500 is far
off from the true system and has huge uncertainties.

To get a closer look at the model fits to the data and the variances, magnify a portion
of the plot.

1-54
 Estimate Regularized ARX Model Using System Identification App

The fit of the regularized ARX model arx0500reg closely matches the impulse
response of the true system and the variance is greatly reduced as compared to the
unregularized model.

Related Examples
• “Regularized Identification of Dynamic Systems”

More About
• “Regularized Estimates of Model Parameters” on page 1-41

1-55
1-56
 2

Data Import and Processing

• “Supported Data” on page 2-3

• “Ways to Obtain Identification Data” on page 2-5
• “Ways to Prepare Data for System Identification” on page 2-6
• “Requirements on Data Sampling” on page 2-8
• “Representing Data in MATLAB Workspace” on page 2-9
• “Import Time-Domain Data into the App” on page 2-16
• “Import Frequency-Domain Data into the App” on page 2-19
• “Import Data Objects into the App” on page 2-25
• “Specifying the Data Sampling Interval” on page 2-28
• “Specify Estimation and Validation Data in the App” on page 2-30
• “Preprocess Data Using Quick Start” on page 2-32
• “Create Data Sets from a Subset of Signal Channels” on page 2-33
• “Create Multiexperiment Data Sets in the App” on page 2-35
• “Managing Data in the App” on page 2-42
• “Representing Time- and Frequency-Domain Data Using iddata Objects” on page
2-50
• “Create Multiexperiment Data at the Command Line” on page 2-60
• “Dealing with Multi-Experiment Data and Merging Models” on page 2-63
• “Managing iddata Objects” on page 2-78
• “Representing Frequency-Response Data Using idfrd Objects” on page 2-83
• “Analyzing Data Quality” on page 2-90
• “Selecting Subsets of Data” on page 2-100
• “Handling Missing Data and Outliers” on page 2-104
• “Handling Offsets and Trends in Data” on page 2-108
• “How to Detrend Data Using the App” on page 2-111
2 Data Import and Processing

• “How to Detrend Data at the Command Line” on page 2-112

• “Resampling Data” on page 2-114
• “Resampling Data Using the App” on page 2-120
• “Resampling Data at the Command Line” on page 2-121
• “Filtering Data” on page 2-123
• “How to Filter Data Using the App” on page 2-125
• “How to Filter Data at the Command Line” on page 2-128
• “Generate Data Using Simulation” on page 2-131
• “Transforming Between Time- and Frequency-Domain Data” on page 2-137
• “Manipulating Complex-Valued Data” on page 2-145

2-2
 Supported Data

Supported Data
System Identification Toolbox software supports estimation of linear models from both
time- and frequency-domain data. For nonlinear models, this toolbox supports only time-
domain data. For more information, see “Supported Models for Time- and Frequency-
Domain Data” on page 1-32.

The data can have single or multiple inputs and outputs, and can be either real or
complex.

Your time-domain data should be sampled at discrete and uniformly spaced time instants
to obtain an input sequence
u={u(T),u(2T),...,u(NT)}

and a corresponding output sequence

y={y(T),y(2T),...,y(NT)}

u(t) and y(t) are the values of the input and output signals at time t, respectively.

This toolbox supports modeling both single- or multiple-channel input-output data or

time-series data.

Supported Data Description

Time-domain I/O data One or more input variables u(t) and one or more
output variables y(t), sampled as a function of time.
Time-domain data can be either real or complex
Time-series data Contains one or more outputs y(t) and no measured
input. Can be time-domain or frequency-domain
data.
Frequency-domain data Fourier transform of the input and output time-
domain signals. The data is the set of input and
output signals in frequency domain; the frequency
grid need not be uniform.
Frequency-response data Complex frequency-response values for a linear
system characterized by its transfer function G,
measurable directly using a spectrum analyzer.
Also called frequency function data. Represented by
frd or idfrd objects. The data sample time may be
zero or nonzero. The frequency vector need not be
uniformly spaced.

2-3
2 Data Import and Processing

Note: If your data is complex valued, see “Manipulating Complex-Valued Data” on page
2-145 for information about supported operations for complex data.

2-4
 Ways to Obtain Identification Data

Ways to Obtain Identification Data

You can obtain identification data by:

• Measuring input and output signals from a physical system.

Your data must capture the important system dynamics, such as dominant time
constants. After measuring the signals, organize the data into variables, as described
in “Representing Data in MATLAB Workspace” on page 2-9. Then, import it in
the System Identification app or represent it as a data object for estimating models at
the command line.
• Generating an input signal with desired characteristics, such as a random Gaussian
or binary signal or a sinusoid, using idinput. Then, generate an output signal using
this input to simulate a model with known coefficients. For more information, see
“Generate Data Using Simulation” on page 2-131.

Using input/output data thus generated helps you study the impact of input signal
characteristics and noise on estimation.
• Logging signals from Simulink® models.

This technique is useful when you want to replace complex components in your model
with identified models to speed up simulations or simplify control design tasks.
For more information on how to log signals, see “Export Signal Data Using Signal
Logging” in the Simulink documentation.

2-5
2 Data Import and Processing

Ways to Prepare Data for System Identification

Before you can perform any task in this toolbox, your data must be in the MATLAB
workspace. You can import the data from external data files or manually create
data arrays at the command line. For more information about importing data, see
“Representing Data in MATLAB Workspace” on page 2-9.

The following tasks help to prepare your data for identifying models from data:

Represent data for system identification

You can represent data in the format of this toolbox by doing one of the following:

• For working in the app, import data into the System Identification app.

See “Represent Data”.

• For working at the command line, create an iddata or idfrd object.

For time-domain or frequency-domain data, see “Representing Time- and Frequency-

Domain Data Using iddata Objects” on page 2-50.

For frequency-response data, see “Representing Frequency-Response Data Using idfrd

Objects” on page 2-83.
• To simulate data with and without noise, see “Generate Data Using Simulation” on
page 2-131.

Analyze data quality

You can analyze your data by doing either of the following:

• Plotting data to examine both time- and frequency-domain behavior.

See “Analyzing Data Quality” on page 2-90.

• Using the advice command to analyze the data for the presence of constant offsets
and trends, delay, possible feedback, and signal excitation levels.

See “How to Analyze Data Using the advice Command” on page 2-98.

2-6
 Preprocess data

Preprocess data
Review the data characteristics for any of the following features to determine if there is a
need for preprocessing:

• Missing or faulty values (also known as outliers). For example, you might see
gaps that indicate missing data, values that do not fit with the rest of the data, or
noninformative values.

See “Handling Missing Data and Outliers” on page 2-104.

• Offsets and drifts in signal levels (low-frequency disturbances).

See “Handling Offsets and Trends in Data” on page 2-108 for information about
subtracting means and linear trends, and “Filtering Data” on page 2-123 for
information about filtering.
• High-frequency disturbances above the frequency interval of interest for the system
dynamics.

See “Resampling Data” on page 2-114 for information about decimating and
interpolating values, and “Filtering Data” on page 2-123 for information about
filtering.

Select a subset of your data

You can use data selection as a way to clean the data and exclude parts with noisy or
missing information. You can also use data selection to create independent data sets for
estimation and validation.

To learn more about selecting data, see “Selecting Subsets of Data” on page 2-100.

Combine data from multiple experiments

You can combine data from several experiments into a single data set. The model you
estimate from a data set containing several experiments describes the average system
that represents these experiments.

To learn more about creating multiple-experiment data sets, see “Create

Multiexperiment Data Sets in the App” on page 2-35 or “Create Multiexperiment
Data at the Command Line” on page 2-60.

2-7
2 Data Import and Processing

Requirements on Data Sampling

A sampling interval is the time between successive data samples. It is sometimes also
referred to as sampling time or sample time.

The System Identification app only supports uniformly sampled data.

The System Identification Toolbox product provides limited support for nonuniformly
sampled data. For more information about specifying uniform and nonuniform time
vectors, see “Constructing an iddata Object for Time-Domain Data” on page 2-50.

2-8
 Representing Data in MATLAB Workspace

Representing Data in MATLAB Workspace

In this section...
“Time-Domain Data Representation” on page 2-9
“Time-Series Data Representation” on page 2-10
“Frequency-Domain Data Representation” on page 2-11

Time-Domain Data Representation

Time-domain data consists of one or more input variables u(t) and one or more output
variables y(t), sampled as a function of time. If there is no input variable, see “Time-
Series Data Representation” on page 2-10.

You must organize time-domain input/output data in the following format:

• For single-input/single-output (SISO) data, the sampled data values must be double
column vectors.
• For multi-input/multi-output (MIMO) data with Nu inputs and Ny outputs, and Ns
number of data samples (measurements):

• The input data must be an Ns-by-Nu matrix

• The output data must be an Ns-by-Ny matrix

To use time-domain data for identification, you must know the sampling interval. If you
are working with uniformly sampled data, use the actual sampling interval from your
experiment. Each data value is assigned a time instant, which is calculated from the
start time and sampling interval. You can work with nonuniformly sampled data only at
the command line by specifying a vector of time instants using the SamplingInstants
property of iddata, as described in “Constructing an iddata Object for Time-Domain
Data” on page 2-50.

For continuous-time models, you must also know the input intersample behavior, such as
zero-order hold and first-order-hold.

For more information about importing data into MATLAB, see “Data Import and
Export”.

After you have the variables in the MATLAB workspace, import them into the System
Identification app or create a data object for working at the command line. For more

2-9
2 Data Import and Processing

information, see “Import Time-Domain Data into the App” on page 2-16 and
“Representing Time- and Frequency-Domain Data Using iddata Objects” on page
2-50.

Time-Series Data Representation

Time-series data is time-domain or frequency-domain data that consist of one or more
outputs y(t) with no corresponding input. For more information on how to obtain
identification data, see “Ways to Obtain Identification Data” on page 2-5.

You must organize time-series data in the following format:

• For single-input/single-output (SISO) data, the output data values must be a column
vector.
• For data with Ny outputs, the output is an Ns-by-Ny matrix, where Ns is the number of
output data samples (measurements).

To use time-series data for identification, you also need the sampling interval. If you
are working with uniformly sampled data, use the actual sampling interval from your
experiment. Each data value is assigned a sample time, which is calculated from the
start time and the sampling interval. If you are working with nonuniformly sampled
data at the command line, you can specify a vector of time instants using the iddata
SamplingInstants property, as described in “Constructing an iddata Object for Time-
Domain Data” on page 2-50. Note that model estimation cannot be performed using
non-uniformly sampled data.

For more information about importing data into the MATLAB workspace, see “Data
Import and Export”.

After you have the variables in the MATLAB workspace, import them into the System
Identification app or create a data object for working at the command line. For more
information, see “Import Time-Domain Data into the App” on page 2-16 and
“Representing Time- and Frequency-Domain Data Using iddata Objects” on page
2-50.

For information about estimating time-series model parameters, see “Time-Series Model
Identification”.

2-10
 Representing Data in MATLAB Workspace

Frequency-Domain Data Representation

Frequency-domain data consists of either transformed input and output time-domain
signals or system frequency response sampled as a function of the independent variable
frequency.

• “Frequency-Domain Input/Output Signal Representation” on page 2-11

• “Frequency-Response Data Representation” on page 2-13

Frequency-Domain Input/Output Signal Representation

• “What Is Frequency-Domain Input/Output Signal?” on page 2-11

• “How to Represent Frequency-Domain Data in MATLAB” on page 2-12

What Is Frequency-Domain Input/Output Signal?

Frequency-domain data is the Fourier transform of the input and output time-domain
signals. For continuous-time signals, the Fourier transform over the entire time axis is
defined as follows:

•
Y (iw) = Ú y(t) e-iwt dt
-•
•
-iwt
U ( iw) = Ú u(t)e dt
-•

In the context of numerical computations, continuous equations are replaced by their

discretized equivalents to handle discrete data values. For a discrete-time system with
a sampling interval T, the frequency-domain output Y(eiw) and input U(eiw) is the time-
discrete Fourier transform (TDFT):

N
Y ( eiwT ) = T Â y(kT )e-iwkT
k =1

In this example, k = 1,2,...,N, where N is the number of samples in the sequence.

Note: This form only discretizes the time. The frequency is continuous.

2-11
2 Data Import and Processing

In practice, the Fourier transform cannot be handled for all continuous frequencies
and you must specify a finite number of frequencies. The discrete Fourier transform
(DFT) of time-domain data for N equally spaced frequencies between 0 and the sampling
frequency 2π/N is:

N
Y ( eiwnT ) = Â y(kT )e-iw kT
n

k= 1
2pn
wn = n = 0,1, 2,… , N - 1
T

The DFT is useful because it can be calculated very efficiently using the fast Fourier
transform (FFT) method. Fourier transforms of the input and output data are complex
numbers.

For more information on how to obtain identification data, see “Ways to Obtain
Identification Data” on page 2-5.
How to Represent Frequency-Domain Data in MATLAB

You must organize frequency-domain data in the following format:

• Input and output

• For single-input/single-output (SISO) data:

•
The input data must be a column vector containing the values u eiw kT
( )
•
The output data must be a column vector containing the values y eiw kT
( )
k=1, 2, ..., Nf, where Nf is the number of frequencies.
• For multi-input/multi-output data with Nu inputs, Ny outputs and Nf frequency
measurements:

• The input data must be an Nf-by-Nu matrix

• The output data must be an Nf-by-Ny matrix
• Frequencies

• Must be a column vector.

2-12
 Representing Data in MATLAB Workspace

For more information about importing data into the MATLAB workspace, see “Data
Import and Export”.

After you have the variables in the MATLAB workspace, import them into the System
Identification app or create a data object for working at the command line. For more
information, see “Importing Frequency-Domain Input/Output Signals into the App” on
page 2-19 and “Representing Time- and Frequency-Domain Data Using iddata
Objects” on page 2-50.

Frequency-Response Data Representation

• “What Is Frequency-Response Data?” on page 2-13

• “How to Represent Frequency-Response Data in MATLAB” on page 2-14

What Is Frequency-Response Data?

Frequency-response data, also called frequency-function data, consists of complex

frequency-response values for a linear system characterized by its transfer function G.
Frequency-response data tells you how the system handles sinusoidal inputs. You can
measure frequency-response data values directly using a spectrum analyzer, for example,
which provides a compact representation of the input-output relationship (compared to
storing input and output independently).

The transfer function G is an operator that takes the input u of a linear system to the
output y:

y = Gu

For a continuous-time system, the transfer function relates the Laplace transforms of the
input U(s) and output Y(s):

Y ( s) = G ( s)U ( s)

In this case, the frequency function G(iw) is the transfer function evaluated on the
imaginary axis s=iw.

For a discrete-time system sampled with a time interval T, the transfer function relates
the Z-transforms of the input U(z) and output Y(z):

Y ( z) = G ( z)U ( z)

2-13
2 Data Import and Processing

In this case, the frequency function G(eiwT) is the transfer function G(z) evaluated on
the unit circle. The argument of the frequency function G(eiwT) is scaled by the sampling
interval T to make the frequency function periodic with the sampling frequency 2 p T .

When the input to the system is a sinusoid of a specific frequency, the output is also
a sinusoid with the same frequency. The amplitude of the output is G times the
amplitude of the input. The phase of the shifted from the input by j = arg G . G is
evaluated at the frequency of the input sinusoid.

Frequency-response data represents a (nonparametric) model of the relationship between

the input and the outputs as a function of frequency. You might use such a model, which
consists of a table or plot of values, to study the system frequency response. However,
this model is not suitable for simulation and prediction. You should create parametric
model from the frequency-response data.

For more information on how to obtain identification data, see “Ways to Obtain
Identification Data” on page 2-5.
How to Represent Frequency-Response Data in MATLAB

You can represent frequency-response data in two ways:

• Complex-values G(eiω) , for given frequencies ω

• Amplitude G and phase shift j = arg G values

You can import both the formats directly in the System Identification app. At the
command line, you must represent complex data using an frd or idfrd object. If the
data is in amplitude and phase format, convert it to complex frequency-response vector
using h(ω) = A(ω)ejϕ(ω).

You must organize frequency-response data in the following format:

Rrequency-Response For Single-Input Single-Output (SISO) For Multi-Input Multi-Output (MIMO) Data
Data Representation Data
Complex Values • Frequency function must be a • Frequency function must be an Ny-
column vector. by-Nu-by-Nf array, where Nu is the
• Frequency values must be a number of inputs, Ny is the number
column vector.

2-14
 Representing Data in MATLAB Workspace

Rrequency-Response For Single-Input Single-Output (SISO) For Multi-Input Multi-Output (MIMO) Data
Data Representation Data
of outputs, and Nf is the number of
frequency measurements.
• Frequency values must be a column
vector.
Amplitude and phase • Amplitude and phase must • Amplitude and phase must each be
shift values each be a column vector. an Ny-by-Nu-by-Nf array, where Nu
• Frequency values must be a is the number of inputs, Ny is the
column vector. number of outputs, and Nf is the
number of frequency measurements.
• Frequency values must be a column
vector.

For more information about importing data into the MATLAB workspace, see “Data
Import and Export”.

After you have the variables in the MATLAB workspace, import them into the System
Identification app or create a data object for working at the command line. For more
information about importing data into the app, see “Importing Frequency-Response
Data into the App” on page 2-21. To learn more about creating a data object, see
“Representing Frequency-Response Data Using idfrd Objects” on page 2-83.

2-15
2 Data Import and Processing

Import Time-Domain Data into the App

Before you can import time-domain data into the System Identification app, you must
import the data into the MATLAB workspace, as described in “Time-Domain Data
Representation” on page 2-9.

Note: Your time-domain data must be sampled at equal time intervals. The input and
output signals must have the same number of data samples.

To import data into the app:

1 Type the following command in the MATLAB Command Window to open the app:
systemIdentification
2 In the System Identification app window, select Import data > Time domain data.
This action opens the Import Data dialog box.

3 Specify the following options:

Note: For time series, only import the output signal and enter [] for the input.

• Input — Enter the MATLAB variable name (column vector or matrix) or a

MATLAB expression that represents the input data. The expression must
evaluate to a column vector or matrix.
• Output — Enter the MATLAB variable name (column vector or matrix) or a
MATLAB expression that represents the output data. The expression must
evaluate to a column vector or matrix.
• Data name — Enter the name of the data set, which appears in the System
Identification app window after the import operation is completed.

2-16
 Import Time-Domain Data into the App

• Starting time — Enter the starting value of the time axis for time plots.
• Sampling interval — Enter the actual sampling interval in the experiment. For
more information about this setting, see “Specifying the Data Sampling Interval”
on page 2-28.

Tip The System Identification Toolbox product uses the sampling interval during
model estimation and to set the horizontal axis on time plots. If you transform
a time-domain signal to a frequency-domain signal, the Fourier transforms are
computed as discrete Fourier transforms (DFTs) using this sampling interval.
4 (Optional) In the Data Information area, click More to expand the dialog box and
enter the following settings:

Input Properties

• InterSample — This options specifies the behavior of the input signals between
samples during data acquisition. It is used when transforming models from
discrete-time to continuous-time and when resampling the data.

• zoh (zero-order hold) indicates that the input was piecewise-constant during
data acquisition.
• foh (first-order hold) indicates that the output was piecewise-linear during
data acquisition.
• bl (bandwidth-limited behavior) specifies that the continuous-time input
signal has zero power above the Nyquist frequency (equal to the inverse of the
sampling interval).

Note: See the d2c and c2d reference pages for more information about
transforming between discrete-time and continuous-time models.
• Period — Enter Inf to specify a nonperiodic input. If the underlying time-
domain data was periodic over an integer number of periods, enter the period of
the input signal.

Note: If your data is periodic, always include a whole number of periods for model
estimation.

Channel Names

2-17
2 Data Import and Processing

• Input — Enter a string to specify the name of one or more input channels.

Tip Naming channels helps you to identify data in plots. For multivariable input-
output signals, you can specify the names of individual Input and Output
channels, separated by commas.
• Output — Enter a string to specify the name of one or more output channels.

Physical Units of Variables

• Input — Enter a string to specify the input units.

Tip When you have multiple inputs and outputs, enter a comma-separated list of
Input and Output units corresponding to each channel.
• Output — Enter a string to specify the output units.

Notes — Enter comments about the experiment or the data. For example, you might
enter the experiment name, date, and a description of experimental conditions.
Models you estimate from this data inherit your data notes.
5 Click Import. This action adds a new data icon to the System Identification app
window.
6 Click Close to close the Import Data dialog box.

2-18
 Import Frequency-Domain Data into the App

Import Frequency-Domain Data into the App

In this section...
“Importing Frequency-Domain Input/Output Signals into the App” on page 2-19
“Importing Frequency-Response Data into the App” on page 2-21

Importing Frequency-Domain Input/Output Signals into the App

Frequency-domain data consists of Fourier transforms of time-domain data (a function of
frequency).

Before you can import frequency-domain data into the System Identification app, you
must import the data into the MATLAB workspace, as described in “Frequency-Domain
Input/Output Signal Representation” on page 2-11.

Note: The input and output signals must have the same number of data samples.

To import data into the app:

1 Type the following command in the MATLAB Command Window to open the app:

systemIdentification
2 In the System Identification app window, select Import data > Freq. domain
data. This action opens the Import Data dialog box.
3 Specify the following options:

• Input — Enter the MATLAB variable name (column vector or matrix) or a

MATLAB expression that represents the input data. The expression must
evaluate to a column vector or matrix.
• Output — Enter the MATLAB variable name (column vector or matrix) or a
MATLAB expression that represents the output data. The expression must
evaluate to a column vector or matrix.
• Frequency — Enter the MATLAB variable name of a vector or a MATLAB
expression that represents the frequencies. The expression must evaluate to a
column vector.

2-19
2 Data Import and Processing

The frequency vector must have the same number of rows as the input and
output signals.
• Data name — Enter the name of the data set, which appears in the System
Identification app window after the import operation is completed.
• Frequency unit — Enter Hz for Hertz or keep the rad/s default value.
• Sampling interval — Enter the actual sampling interval in the experiment.
For continuous-time data, enter 0. For more information about this setting, see
“Specifying the Data Sampling Interval” on page 2-28.
4 (Optional) In the Data Information area, click More to expand the dialog box and
enter the following optional settings:

Input Properties

• InterSample — This options specifies the behavior of the input signals between
samples during data acquisition. It is used when transforming models from
discrete-time to continuous-time and when resampling the data.

• zoh (zero-order hold) indicates that the input was piecewise-constant during
data acquisition.
• foh (first-order hold) indicates that the output was piecewise-linear during
data acquisition.
• bl (bandwidth-limited behavior) specifies that the continuous-time input
signal has zero power above the Nyquist frequency (equal to the inverse of the
sampling interval).

Note: See the d2c and c2d reference page for more information about
transforming between discrete-time and continuous-time models.
• Period — Enter Inf to specify a nonperiodic input. If the underlying time-
domain data was periodic over an integer number of periods, enter the period of
the input signal.

Note: If your data is periodic, always include a whole number of periods for model
estimation.

Channel Names

2-20
 Import Frequency-Domain Data into the App

• Input — Enter a string to specify the name of one or more input channels.

Tip Naming channels helps you to identify data in plots. For multivariable input
and output signals, you can specify the names of individual Input and Output
channels, separated by commas.
• Output — Enter a string to specify the name of one or more output channels.

Physical Units of Variables

• Input — Enter a string to specify the input units.

Tip When you have multiple inputs and outputs, enter a comma-separated list of
Input and Output units corresponding to each channel.
• Output — Enter a string to specify the output units.

Notes — Enter comments about the experiment or the data. For example, you might
enter the experiment name, date, and a description of experimental conditions.
Models you estimate from this data inherit your data notes.
5 Click Import. This action adds a new data icon to the System Identification app
window.
6 Click Close to close the Import Data dialog box.

Importing Frequency-Response Data into the App

• “Prerequisite” on page 2-21
• “Importing Complex-Valued Frequency-Response Data” on page 2-22
• “Importing Amplitude and Phase Frequency-Response Data” on page 2-23

Prerequisite

Before you can import frequency-response data into the System Identification app, you
must import the data into the MATLAB workspace, as described in “Frequency-Response
Data Representation” on page 2-13.

2-21
2 Data Import and Processing

Importing Complex-Valued Frequency-Response Data

To import frequency-response data consisting of complex-valued frequency values at

specified frequencies:

1 Type the following command in the MATLAB Command Window to open the app:

systemIdentification
2 In the System Identification app window, select Import data > Freq. domain
data. This action opens the Import Data dialog box.
3 In the Data Format for Signals list, select Freq. Function (Complex).
4 Specify the following options:

• Freq. Func. — Enter the MATLAB variable name or a MATLAB expression

that represents the complex frequency-response data G(eiw).
• Frequency — Enter the MATLAB variable name of a vector or a MATLAB
expression that represents the frequencies. The expression must evaluate to a
column vector.
• Data name — Enter the name of the data set, which appears in the System
Identification app window after the import operation is completed.
• Frequency unit — Enter Hz for Hertz or keep the rad/s default value.
• Sampling interval — Enter the actual sampling interval in the experiment.
For continuous-time data, enter 0. For more information about this setting, see
“Specifying the Data Sampling Interval” on page 2-28.
5 (Optional) In the Data Information area, click More to expand the dialog box and
enter the following optional settings:

Channel Names

• Input — Enter a string to specify the name of one or more input channels.

Tip Naming channels helps you to identify data in plots. For multivariable input
and output signals, you can specify the names of individual Input and Output
channels, separated by commas.
• Output — Enter a string to specify the name of one or more output channels.

Physical Units of Variables

2-22
 Import Frequency-Domain Data into the App

• Input — Enter a string to specify the input units.

Tip When you have multiple inputs and outputs, enter a comma-separated list of
Input and Output units corresponding to each channel.
• Output — Enter a string to specify the output units.

Notes — Enter comments about the experiment or the data. For example, you might
enter the experiment name, date, and a description of experimental conditions.
Models you estimate from this data inherit your data notes.
6 Click Import. This action adds a new data icon to the System Identification app
window.
7 Click Close to close the Import Data dialog box.

Importing Amplitude and Phase Frequency-Response Data

To import frequency-response data consisting of amplitude and phase values at specified

frequencies:

1 Type the following command in the MATLAB Command Window to open the app:
systemIdentification
2 In the System Identification app window, select Import data > Freq. domain
data. This action opens the Import Data dialog box.
3 In the Data Format for Signals list, select Freq. Function (Amp/Phase).
4 Specify the following options:

• Amplitude — Enter the MATLAB variable name or a MATLAB expression that

represents the amplitude G .
• Phase (deg) — Enter the MATLAB variable name or a MATLAB expression that
represents the phase j = arg G .
• Frequency — Enter the MATLAB variable name of a vector or a MATLAB
expression that represents the frequencies. The expression must evaluate to a
column vector.
• Data name — Enter the name of the data set, which appears in the System
Identification app window after the import operation is completed.
• Frequency unit — Enter Hz for Hertz or keep the rad/s default value.

2-23
2 Data Import and Processing

• Sampling interval — Enter the actual sampling interval in the experiment.

For continuous-time data, enter 0. For more information about this setting, see
“Specifying the Data Sampling Interval” on page 2-28.
5 (Optional) In the Data Information area, click More to expand the dialog box and
enter the following optional settings:

Channel Names

• Input — Enter a string to specify the name of one or more input channels.

Tip Naming channels helps you to identify data in plots. For multivariable input
and output signals, you can specify the names of individual Input and Output
channels, separated by commas.
• Output — Enter a string to specify the name of one or more output channels.

Physical Units of Variables

• Input — Enter a string to specify the input units.

Tip When you have multiple inputs and outputs, enter a comma-separated list of
Input and Output units corresponding to each channel.
• Output — Enter a string to specify the output units.

Notes — Enter comments about the experiment or the data. For example, you might
enter the experiment name, date, and a description of experimental conditions.
Models you estimate from this data inherit your data notes.
6 Click Import. This action adds a new data icon to the System Identification app
window.
7 Click Close to close the Import Data dialog box.

2-24
 Import Data Objects into the App

Import Data Objects into the App

You can import the System Identification Toolbox iddata and idfrd data objects into
the System Identification app.

Before you can import a data object into the System Identification app, you must create
the data object in the MATLAB workspace, as described in “Representing Time- and
Frequency-Domain Data Using iddata Objects” on page 2-50 or “Representing
Frequency-Response Data Using idfrd Objects” on page 2-83.

Note: You can also import a Control System Toolbox frd object. Importing an frd object
converts it to an idfrd object.

Select Import data > Data object to open the Import Data dialog box.

Import iddata, idfrd, or frd data object in the MATLAB workspace.

To import a data object into the app:

1 Type the following command in the MATLAB Command Window to open the app:

systemIdentification
2 In the System Identification app window, select Import data > Data object.

This action opens the Import Data dialog box. IDDATA or IDFRD/FRD is already
selected in the Data Format for Signals list.
3 Specify the following options:

• Object — Enter the name of the MATLAB variable that represents the data
object in the MATLAB workspace. Press Enter.

2-25
2 Data Import and Processing

• Data name — Enter the name of the data set, which appears in the System
Identification app window after the import operation is completed.
• (Only for time-domain iddata object) Starting time — Enter the starting value
of the time axis for time plots.
• (Only for frequency domain iddata or idfrd object) Frequency unit — Enter
the frequency unit for response plots.
• Sampling interval — Enter the actual sampling interval in the experiment. For
more information about this setting, see “Specifying the Data Sampling Interval”
on page 2-28.

Tip The System Identification Toolbox product uses the sampling interval during
model estimation and to set the horizontal axis on time plots. If you transform
a time-domain signal to a frequency-domain signal, the Fourier transforms are
computed as discrete Fourier transforms (DFTs) using this sampling interval.
4 (Optional) In the Data Information area, click More to expand the dialog box and
enter the following optional settings:

(Only for iddata object) Input Properties

• InterSample — This options specifies the behavior of the input signals between
samples during data acquisition. It is used when transforming models from
discrete-time to continuous-time and when resampling the data.

• zoh (zero-order hold) indicates that the input was piecewise-constant during
data acquisition.
• foh (first-order hold) indicates that the input was piecewise-linear during
data acquisition.
• bl (bandwidth-limited behavior) specifies that the continuous-time input
signal has zero power above the Nyquist frequency (equal to the inverse of the
sampling interval).

Note: See the d2c and c2d reference page for more information about
transforming between discrete-time and continuous-time models.
• Period — Enter Inf to specify a nonperiodic input. If the underlying time-
domain data was periodic over an integer number of periods, enter the period of
the input signal.

2-26
 Import Data Objects into the App

Note: If your data is periodic, always include a whole number of periods for model
estimation.

Channel Names

• Input — Enter a string to specify the name of one or more input channels.

Tip Naming channels helps you to identify data in plots. For multivariable input
and output signals, you can specify the names of individual Input and Output
channels, separated by commas.
• Output — Enter a string to specify the name of one or more output channels.

Physical Units of Variables

• Input — Enter a string to specify the input units.

Tip When you have multiple inputs and outputs, enter a comma-separated list of
Input and Output units corresponding to each channel.
• Output — Enter a string to specify the output units.

Notes — Enter comments about the experiment or the data. For example, you might
enter the experiment name, date, and a description of experimental conditions.
Models you estimate from this data inherit your data notes.
5 Click Import. This action adds a new data icon to the System Identification app
window.
6 Click Close to close the Import Data dialog box.

2-27
2 Data Import and Processing

Specifying the Data Sampling Interval

When you import data into the app, you must specify the data sampling interval.

The sampling interval is the time between successive data samples in your experiment
and must be the numerical time interval at which your data is sampled in any units. For
example, enter 0.5 if your data was sampled every 0.5 s, and enter 1 if your data was
sampled every 1 s.

You can also use the sampling interval as a flag to specify continuous-time data. When
importing continuous-time frequency domain or frequency-response data, set the
Sampling interval to 0.

The sampling interval is used during model estimation. For time-domain data, the
sampling interval is used together with the start time to calculate the sampling time
instants. When you transform time-domain signals to frequency-domain signals (see the
fft reference page), the Fourier transforms are computed as discrete Fourier transforms
(DFTs) for this sampling interval. In addition, the sampling instants are used to set the
horizontal axis on time plots.

2-28
 Specifying the Data Sampling Interval

Sampling Interval in the Import Data dialog box

2-29
2 Data Import and Processing

Specify Estimation and Validation Data in the App

You should use different data sets to estimate and validate your model for best validation
results.

In the System Identification app, Working Data refers to estimation data. Similarly,
Validation Data refers to the data set you use to validate a model. For example, when
you plot the model output, the input to the model is the input signal from the validation
data set. This plot compares model output to the measured output in the validation data
set. Selecting Model resids performs residual analysis using the validation data.

To specify Working Data, drag and drop the corresponding data icon into the Working
Data rectangle, as shown in the following figure. Similarly, to specify Validation
Data, drag and drop the corresponding data icon into the Validation Data rectangle.
Alternatively, right-click the icon to open the Data/model Info dialog box. Select the Use
as Working Data or Use as Validation Data and click Apply to specify estimation
and validation data, respectively.

2-30
 Specify Estimation and Validation Data in the App

Drag and drop estimation data set

Drag and drop validation data set

2-31
2 Data Import and Processing

Preprocess Data Using Quick Start

As a preprocessing shortcut for time-domain data, select Preprocess > Quick start to
simultaneously perform the following four actions:

• Subtract the mean value from each channel.

Note: For information about when to subtract mean values from the data, see
“Handling Offsets and Trends in Data” on page 2-108.
• Split data into two parts.
• Specify the first part as estimation data for models (or Working Data).
• Specify the second part as Validation Data.

2-32
 Create Data Sets from a Subset of Signal Channels

Create Data Sets from a Subset of Signal Channels

You can create a new data set in the System Identification app by extracting subsets of
input and output channels from an existing data set.

To create a new data set from selected channels:

1 In the System Identification app, drag the icon of the data from which you want to
select channels to the Working Data rectangle.
2 Select Preprocess > Select channels to open the Select Channels dialog box.

The Inputs list displays the input channels and the Outputs list displays the
output channels in the selected data set.
3 In the Inputs list, select one or more channels in any of following ways:

• Select one channel by clicking its name.

• Select adjacent channels by pressing the Shift key while clicking the first and
last channel names.
• Select nonadjacent channels by pressing the Ctrl key while clicking each channel
name.

2-33
2 Data Import and Processing

Tip To exclude input channels and create time-series data, clear all selections by
holding down the Ctrl key and clicking each selection. To reset selections, click
Revert.
4 In the Outputs list, select one or more channels in any of following ways:

• Select one channel by clicking its name.

• Select adjacent channels by pressing the Shift key while clicking the first and
last channel names.
• Select nonadjacent channels by pressing the Ctrl key while clicking each channel
name.

Tip To reset selections, click Revert.

5 In the Data name field, type the name of the new data set. Use a name that is
unique in the Data Board.
6 Click Insert to add the new data set to the Data Board in the System Identification
app.
7 Click Close.

2-34
 Create Multiexperiment Data Sets in the App

Create Multiexperiment Data Sets in the App

In this section...
“Why Create Multiexperiment Data?” on page 2-35
“Limitations on Data Sets” on page 2-35
“Merging Data Sets” on page 2-35
“Extracting Specific Experiments from a Multiexperiment Data Set into a New Data
Set” on page 2-39

Why Create Multiexperiment Data?

You can create a time-domain or frequency-domain data set in the System Identification
app that includes several experiments. Identifying models for multiexperiment data
results in an average model.

Experiments can mean data that was collected during different sessions, or portions
of the data collected during a single session. In the latter situation, you can create
multiexperiment data by splitting a single data set into multiple segments that exclude
corrupt data, and then merge the good data segments.

Limitations on Data Sets

You can only merge data sets that have all of the following characteristics:

• Same number of input and output channels.

• Different names. The name of each data set becomes the experiment name in the
merged data set.
• Same input and output channel names.
• Same data domain (that is, time-domain data or frequency-domain data only).

Merging Data Sets

You can merge data sets using the System Identification app.

For example, suppose that you want to combine the data sets tdata, tdata2, tdata3,
tdata4 shown in the following figure.

2-35
2 Data Import and Processing

App Contains Four Data Sets to Merge

To merge data sets in the app:

1 In the Operations area, select <--Preprocess > Merge experiments from the
drop-down menu to open the Merge Experiments dialog box.

2-36
 Create Multiexperiment Data Sets in the App

2 In the System Identification app window, drag a data set icon to the Merge
Experiments dialog box, to the drop them here to be merged rectangle.

The name of the data set is added to the List of sets. Repeat for each data set you
want to merge.

2-37
2 Data Import and Processing

tdata and tdata2 to Be Merged

Tip To empty the list, click Revert.

3 In the Data name field, type the name of the new data set. This name must be
unique in the Data Board.
4 Click Insert to add the new data set to the Data Board in the System Identification
app window.

2-38
 Create Multiexperiment Data Sets in the App

Data Board Now Contains tdatam with Merged Experiments

5 Click Close to close the Merge Experiments dialog box.

Tip To get information about a data set in the System Identification app, right-click the
data icon to open the Data/model Info dialog box.

Extracting Specific Experiments from a Multiexperiment Data Set into a

New Data Set
When a data set already consists of several experiments, you can extract one or more of
these experiments into a new data set, using the System Identification app.

For example, suppose that tdatam consists of four experiments.

To create a new data set that includes only the first and third experiments in this data
set:

1 In the System Identification app window, drag and drop the tdatam data icon to the
Working Data rectangle.

2-39
2 Data Import and Processing

tdatam Is Set to Working Data

2 In the Operations area, select Preprocess > Select experiments from the drop-
down menu to open the Select Experiment dialog box.
3 In the Experiments list, select one or more data sets in either of the following ways:

• Select one data set by clicking its name.

• Select adjacent data sets by pressing the Shift key while clicking the first and
last names.
• Select nonadjacent data sets by pressing the Ctrl key while clicking each name.

2-40
 Create Multiexperiment Data Sets in the App

4 In the Data name field, type the name of the new data set. This name must be
unique in the Data Board.
5 Click Insert to add the new data set to the Data Board in the System Identification
app.
6 Click Close to close the Select Experiment dialog box.

2-41
2 Data Import and Processing

Managing Data in the App

In this section...
“Viewing Data Properties” on page 2-42
“Renaming Data and Changing Display Color” on page 2-43
“Distinguishing Data Types” on page 2-46
“Organizing Data Icons” on page 2-46
“Deleting Data Sets” on page 2-47
“Exporting Data to the MATLAB Workspace” on page 2-48

Viewing Data Properties

You can get information about each data set in the System Identification app by right-
clicking the corresponding data icon.

The Data/model Info dialog box opens. This dialog box describes the contents and the
properties of the corresponding data set. It also displays any associated notes and the
command-line equivalent of the operations you used to create this data.

Tip To view or modify properties for several data sets, keep this window open and right-
click each data set in the System Identification app. The Data/model Info dialog box
updates as you select each data set.

2-42
 Managing Data in the App

Data object
description

History of
syntax that
created this
object

To displays the data properties in the MATLAB Command Window, click Present.

Renaming Data and Changing Display Color

You can rename data and change its display color by double-clicking the data icon in the
System Identification app.

The Data/model Info dialog box opens. This dialog box describes both the contents
and the properties of the data. The object description area displays the syntax of the
operations you used to create the data in the app.

The Data/model Info dialog box also lets you rename the data by entering a new name in
the Data name field.

2-43
2 Data Import and Processing

You can also specify a new display color using three RGB values in the Color field.
Each value is between 0 to 1 and indicates the relative presence of red, green, and blue,
respectively. For more information about specifying default data color, see “Customizing
the System Identification App” on page 14-15.

Tip As an alternative to using three RGB values, you can enter any one of the following
letters in single quotes:
'y' 'r' 'b' 'c' 'g' 'm' 'k'

These strings represent yellow, red, blue, cyan, green, magenta, and black, respectively.

2-44
 Managing Data in the App

Specify name
for data set
Specify color
used to display
data set

Information About the Data

You can enter comments about the origin and state of the data in the Diary And Notes
area. For example, you might want to include the experiment name, date, and the
description of experimental conditions. When you estimate models from this data, these
notes are associated with the models.

Clicking Present display portions of this information in the MATLAB Command

Window.

2-45
2 Data Import and Processing

Distinguishing Data Types

The background color of a data icon is color-coded, as follows:

• White background represents time-domain data.

• Blue background represents frequency-domain data.
• Yellow background represents frequency-response data.

Time-domain data

Frequency-domain
data

Frequency-response
data

Colors Representing Type of Data

Organizing Data Icons

You can rearrange data icons in the System Identification app by dragging and dropping
the icons to empty Data Board rectangles in the app.

Note: You cannot drag and drop a data icon into the model area on the right.

When you need additional space for organizing data or model icons, select Options
> Extra model/data board in the System Identification app. This action opens an
extra session window with blank rectangles for data and models. The new window is an
extension of the current session and does not represent a new session.

2-46
 Managing Data in the App

Tip When you import or create data sets and there is insufficient space for the icons, an
additional session window opens automatically.

You can drag and drop data between the main System Identification app and any extra
session windows.

Type comments in the Notes field to describe the data sets. When you save a session, as
described in “Saving, Merging, and Closing Sessions” on page 14-6, all additional
windows and notes are also saved.

Deleting Data Sets

To delete data sets in the System Identification app, drag and drop the corresponding
icon into Trash. You can also use the Delete key on your keyboard to move items to the
Trash. Moving items to Trash does not permanently delete these items.

Note: You cannot delete a data set that is currently designated as Working Data or
Validation Data. You must first specify a different data set in the System Identification
app to be Working Data or Validation Data, as described in “Specify Estimation and
Validation Data in the App” on page 2-30.

To restore a data set from Trash, drag its icon from Trash to the Data or Model Board
in the System Identification app window. You can view the Trash contents by double-
clicking the Trash icon.

2-47
2 Data Import and Processing

Note: You must restore data to the Data Board; you cannot drag data icons to the Model
Board.

To permanently delete all items in Trash, select Options > Empty trash.

Exiting a session empties the Trash automatically.

Exporting Data to the MATLAB Workspace

The data you create in the System Identification app is not available in the MATLAB
workspace until you export the data set. Exporting to the MATLAB workspace is
necessary when you need to perform an operation on the data that is only available at
the command line.

To export a data set to the MATLAB workspace, do one of the following:

• Drag and drop the corresponding icon to the To Workspace rectangle.

• Right-click the icon to open the Data/model Info dialog box. Click Export.

When you export data to the MATLAB workspace, the resulting variables have the same
name as in the System Identification app. For example, the following figure shows how to
export the time-domain data object datad.

2-48
 Managing Data in the App

Exporting Data to the MATLAB Workspace

In this example, the MATLAB workspace contains a variable named data after export.

2-49
2 Data Import and Processing

Representing Time- and Frequency-Domain Data Using iddata

Objects
In this section...
“iddata Constructor” on page 2-50
“iddata Properties” on page 2-52
“Select Data Channels, I/O Data and Experiments in iddata Objects” on page 2-55
“Increasing Number of Channels or Data Points of iddata Objects” on page 2-58

iddata Constructor
• “Requirements for Constructing an iddata Object” on page 2-50
• “Constructing an iddata Object for Time-Domain Data” on page 2-50
• “Constructing an iddata Object for Frequency-Domain Data” on page 2-52

Requirements for Constructing an iddata Object

To construct an iddata object, you must have already imported data into the MATLAB
workspace, as described in “Representing Data in MATLAB Workspace” on page 2-9.

Constructing an iddata Object for Time-Domain Data

Use the following syntax to create a time-domain iddata object data:

data = iddata(y,u,Ts)

You can also specify additional properties, as follows:

data = iddata(y,u,Ts,'Property1',Value1,...,'PropertyN',ValueN)

For more information about accessing object properties, see “Properties”.

In this example, Ts is the sampling time, or the time interval, between successive data
samples. For uniformly sampled data, Ts is a scalar value equal to the sampling interval
of your experiment. The default time unit is seconds, but you can set it to a new value
using the TimeUnit property. For more information about iddata time properties, see
“Modifying Time and Frequency Vectors” on page 2-78.

2-50
 Representing Time- and Frequency-Domain Data Using iddata Objects

For nonuniformly sampled data, specify Ts as [], and set the value of the
SamplingInstants property as a column vector containing individual time values. For
example:

data = iddata(y,u,Ts,[],'SamplingInstants',TimeVector)

Where TimeVector represents a vector of time values.

Note: You can modify the property SamplingInstants by setting it to a new vector
with the length equal to the number of data samples.

To represent time-series data, use the following syntax:

ts_data = iddata(y,[],Ts)

where y is the output data, [] indicates empty input data, and Ts is the sampling
interval.

The following example shows how to create an iddata object using single-input/single-
output (SISO) data from dryer2.mat. The input and output each contain 1000 samples
with the sampling interval of 0.08 second.

load dryer2 % Load input u2 and output y2.

data = iddata(y2,u2,0.08) % Create iddata object.

data =

Time domain data set with 1000 samples.

Sample time: 0.08 seconds

Outputs Unit (if specified)

y1

Inputs Unit (if specified)

u1

The default channel name 'y1' is assigned to the first and only output channel.
When y2 contains several channels, the channels are assigned default names
'y1','y2','y2',...,'yn'. Similarly, the default channel name 'u1' is assigned

2-51
2 Data Import and Processing

to the first and only input channel. For more information about naming channels, see
“Naming, Adding, and Removing Data Channels” on page 2-80.

Constructing an iddata Object for Frequency-Domain Data

Frequency-domain data is the Fourier transform of the input and output signals at
specific frequency values. To represent frequency-domain data, use the following syntax
to create the iddata object:

data = iddata(y,u,Ts,'Frequency',w)

'Frequency' is an iddata property that specifies the frequency values w, where w is

the frequency column vector that defines the frequencies at which the Fourier transform
values of y and u are computed. Ts is the time interval between successive data samples
in seconds for the original time-domain data. w, y, and u have the same number of rows.

Note: You must specify the frequency vector for frequency-domain data.

For more information about iddata time and frequency properties, see “Modifying Time
and Frequency Vectors” on page 2-78.

To specify a continuous-time system, set Ts to 0.

You can specify additional properties when you create the iddata object, as follows:

data = iddata(y,u,Ts,'Property1',Value1,...,'PropertyN',ValueN)

For more information about accessing object properties, see “Properties”.

iddata Properties
To view the properties of the iddata object, use the get command. For example, type
the following commands at the prompt:
load dryer2 % Load input u2 and output y2
data = iddata(y2,u2,0.08); % Create iddata object
get(data) % Get property values of data

ans =

2-52
 Representing Time- and Frequency-Domain Data Using iddata Objects

Domain: 'Time'
Name: ''
OutputData: [1000x1 double]
y: 'Same as OutputData'
OutputName: {'y1'}
OutputUnit: {''}
InputData: [1000x1 double]
u: 'Same as InputData'
InputName: {'u1'}
InputUnit: {''}
Period: Inf
InterSample: 'zoh'
Ts: 0.0800
Tstart: []
SamplingInstants: [1000x0 double]
TimeUnit: 'seconds'
ExperimentName: 'Exp1'
Notes: {}
UserData: []

For a complete description of all properties, see the iddata reference page.

You can specify properties when you create an iddata object using the constructor
syntax:

data = iddata(y,u,Ts,'Property1',Value1,...,'PropertyN',ValueN)

To change property values for an existing iddata object, use the set command or dot
notation. For example, to change the sampling interval to 0.05, type the following at the
prompt:

set(data,'Ts',0.05)

or equivalently:

data.ts = 0.05

Property names are not case sensitive. You do not need to type the entire property name
if the first few letters uniquely identify the property.

Tip You can use data.y as an alternative to data.OutputData to access the output
values, or use data.u as an alternative to data.InputData to access the input values.

2-53
2 Data Import and Processing

An iddata object containing frequency-domain data includes frequency-specific

properties, such as Frequency for the frequency vector and Units for frequency units
(instead of Tstart and SamplingIntervals).

To view the property list, type the following command sequence at the prompt:
% Load input u2 and output y2
load dryer2;
% Create iddata object
data = iddata(y2,u2,0.08);
% Take the Fourier transform of the data
% transforming it to frequency domain
data = fft(data)
% Get property values of data
get(data)

data =

Frequency domain data set with responses at 501 frequencies,

ranging from 0 to 39.27 rad/seconds
Sample time: 0.08 seconds

Outputs Unit (if specified)

y1

Inputs Unit (if specified)

u1

ans =

Domain: 'Frequency'
Name: ''
OutputData: [501x1 double]
y: 'Same as OutputData'
OutputName: {'y1'}
OutputUnit: {''}
InputData: [501x1 double]
u: 'Same as InputData'
InputName: {'u1'}
InputUnit: {''}
Period: Inf
InterSample: 'zoh'
Ts: 0.0800

2-54
 Representing Time- and Frequency-Domain Data Using iddata Objects

FrequencyUnit: 'rad/TimeUnit'
Frequency: [501x1 double]
TimeUnit: 'seconds'
ExperimentName: 'Exp1'
Notes: {}
UserData: []

Select Data Channels, I/O Data and Experiments in iddata Objects

• “Subreferencing Input and Output Data” on page 2-55
• “Subreferencing Data Channels” on page 2-56
• “Subreferencing Experiments” on page 2-57

Subreferencing Input and Output Data

Subreferencing data and its properties lets you select data values and assign new data
and property values.

Use the following general syntax to subreference specific data values in iddata objects:

data(samples,outputchannels,inputchannels,experimentname)

In this syntax, samples specify one or more sample indexes, outputchannels and
inputchannels specify channel indexes or channel names, and experimentname
specifies experiment indexes or names.

For example, to retrieve samples 5 through 30 in the iddata object data and store them
in a new iddata object data_sub, use the following syntax:

data_sub = data(5:30)

You can also use logical expressions to subreference data. For example, to retrieve all
data values from a single-experiment data set that fall between sample instants 1.27
and 9.3 in the iddata object data and assign them to data_sub, use the following
syntax:
data_sub = data(data.sa>1.27&data.sa<9.3)

Note: You do not need to type the entire property name. In this example, sa in data.sa
uniquely identifies the SamplingInstants property.

2-55
2 Data Import and Processing

You can retrieve the input signal from an iddata object using the following commands:

u = get(data,'InputData')

or

data.InputData

or

data.u % u is the abbreviation for InputData

Similarly, you can retrieve the output data using

data.OutputData

or

data.y % y is the abbreviation for OutputData

Subreferencing Data Channels

Use the following general syntax to subreference specific data channels in iddata
objects:

data(samples,outputchannels,inputchannels,experiment)

In this syntax, samples specify one or more sample indexes, outputchannels and
inputchannels specify channel indexes or channel names, and experimentname
specifies experiment indexes or names.

To specify several channel names, you must use a cell array of name strings.

For example, suppose the iddata object data contains three output channels (named
y1, y2, and y3), and four input channels (named u1, u2, u3, and u4). To select all data
samples in y3, u1, and u4, type the following command at the prompt:

% Use a cell array to reference

% input channels 'u1' and 'u4'
data_sub = data(:,'y3',{'u1','u4'})

or equivalently

% Use channel indexes 1 and 4

2-56
 Representing Time- and Frequency-Domain Data Using iddata Objects

% to reference the input channels

data_sub = data(:,3,[1 4])

Tip Use a colon (:) to specify all samples or all channels, and the empty matrix ([]) to
specify no samples or no channels.

If you want to create a time-series object by extracting only the output data from an
iddata object, type the following command:

data_ts = data(:,:,[])

You can assign new values to subreferenced variables. For example, the following
command assigns the first 10 values of output channel 1 of data to values in samples
101 through 110 in the output channel 2 of data1. It also assigns the values in samples
101 through 110 in the input channel 3 of data1 to the first 10 values of input channel 1
of data.

data(1:10,1,1) = data1(101:110,2,3)

Subreferencing Experiments

Use the following general syntax to subreference specific experiments in iddata objects:

data(samples,outputchannels,inputchannels,experimentname)

In this syntax, samples specify one or more sample indexes, outputchannels and
inputchannels specify channel indexes or channel names, and experimentname
specifies experiment indexes or names.

When specifying several experiment names, you must use a cell array of name strings.
The iddata object stores experiments name in the ExperimentName property.

For example, suppose the iddata object data contains five experiments with default
names, Exp1, Exp2, Exp3, Exp4, and Exp5. Use the following syntax to subreference the
first and fifth experiment in data:

data_sub = data(:,:,:,{'Exp1','Exp5'}) % Using experiment name

or

data_sub = data(:,:,:,[1 5]) % Using experiment index

2-57
2 Data Import and Processing

Tip Use a colon (:) to denote all samples and all channels, and the empty matrix ([]) to
specify no samples and no channels.

Alternatively, you can use the getexp command. The following example shows how to
subreference the first and fifth experiment in data:

data_sub = getexp(data,{'Exp1','Exp5'}) % Using experiment name

or

data_sub = getexp(data,[1 5]) % Using experiment index

The following example shows how to retrieve the first 100 samples of output channels 2
and 3 and input channels 4 to 8 of Experiment 3:

dat(1:100,[2,3],[4:8],3)

Increasing Number of Channels or Data Points of iddata Objects

• “iddata Properties Storing Input and Output Data” on page 2-58
• “Horizontal Concatenation” on page 2-58
• “Vertical Concatenation” on page 2-59

iddata Properties Storing Input and Output Data

The InputData iddata property stores column-wise input data, and the OutputData
property stores column-wise output data. For more information about accessing iddata
properties, see “iddata Properties” on page 2-52.

Horizontal Concatenation

Horizontal concatenation of iddata objects creates a new iddata object that appends
all InputData information and all OutputData. This type of concatenation produces
a single object with more input and output channels. For example, the following syntax
performs horizontal concatenation on the iddata objects data1,data2,...,dataN:

data = [data1,data2,...,dataN]

This syntax is equivalent to the following longer syntax:

data.InputData =

2-58
 Representing Time- and Frequency-Domain Data Using iddata Objects

[data1.InputData,data2.InputData,...,dataN.InputData]
data.OutputData =
[data1.OutputData,data2.OutputData,...,dataN.OutputData]

For horizontal concatenation, data1,data2,...,dataN must have the same number of

samples and experiments , and the sameTs and Tstart values.

The channels in the concatenated iddata object are named according to the following
rules:

• Combining default channel names. If you concatenate iddata objects with

default channel names, such as u1 and y1, channels in the new iddata object are
automatically renamed to avoid name duplication.
• Combining duplicate input channels. If data1,data2,...,dataN have input
channels with duplicate user-defined names, such that dataK contains channel
names that are already present in dataJ with J < K, the dataK channels are
ignored.
• Combining duplicate output channels. If data1,data2,...,dataN have input
channels with duplicate user-defined names, only the output channels with unique
names are added during the concatenation.

Vertical Concatenation

Vertical concatenation of iddata objects creates a new iddata object that vertically
stacks the input and output data values in the corresponding data channels. The
resulting object has the same number of channels, but each channel contains more
data points. For example, the following syntax creates a data object such that its total
number of samples is the sum of the samples in data1,data2,...,dataN.

data = [data1;data2;... ;dataN]

This syntax is equivalent to the following longer syntax:

data.InputData =
[data1.InputData;data2.InputData;...;dataN.InputData]
data.OutputData =
[data1.OutputData;data2.OutputData;...;dataN.OutputData]

For vertical concatenation, data1,data2,...,dataN must have the same number of

input channels, output channels, and experiments.

2-59
2 Data Import and Processing

Create Multiexperiment Data at the Command Line

In this section...
“Why Create Multiexperiment Data Sets?” on page 2-60
“Limitations on Data Sets” on page 2-60
“Entering Multiexperiment Data Directly” on page 2-60
“Merging Data Sets” on page 2-61
“Adding Experiments to an Existing iddata Object” on page 2-61

Why Create Multiexperiment Data Sets?

You can create iddata objects that contain several experiments. Identifying models for
an iddata object with multiple experiments results in an average model.

In the System Identification Toolbox product, experiments can either mean data collected
during different sessions, or portions of the data collected during a single session. In the
latter situation, you can create a multiexperiment iddata object by splitting the data
from a single session into multiple segments to exclude bad data, and merge the good
data portions.

Note: The idfrd object does not support the iddata equivalent of multiexperiment
data.

Limitations on Data Sets

You can only merge data sets that have all of the following characteristics:

• Same number of input and output channels.

• Same input and output channel names.
• Same data domain (that is, time-domain data or frequency-domain data).

Entering Multiexperiment Data Directly

To construct an iddata object that includes N data sets, you can use this syntax:

2-60
 Create Multiexperiment Data at the Command Line

data = iddata(y,u,Ts)

where y, u, and Ts are 1-by-N cell arrays containing data from the different experiments.
Similarly, when you specify Tstart, Period, InterSample, and SamplingInstants
properties of the iddata object, you must assign their values as 1-by-N cell arrays.

Merging Data Sets

Create a multiexperiment iddata object by merging iddata objects, where each
contains data from a single experiment or is a multiexperiment data set. For example,
you can use the following syntax to merge data:

load iddata1 % Loads iddata object z1

load iddata3 % Loads iddata object z3
z = merge(z1,z3) % Merges experiments z1 and z3 into the iddata object z

z =

Time domain data set containing 2 experiments.

Experiment Samples Sample Time

Exp1 300 0.1
Exp2 300 1

Outputs Unit (if specified)

y1

Inputs Unit (if specified)

u1

These commands create an iddata object that contains two experiments, where the
experiments are assigned default names 'Exp1' and 'Exp2', respectively.

Adding Experiments to an Existing iddata Object

You can add experiments individually to an iddata object as an alternative approach to
merging data sets.

For example, to add the experiments in the iddata object dat4 to data, use the
following syntax:

2-61
2 Data Import and Processing

data(:,:,:,'Run4') = dat4

This syntax explicitly assigns the experiment name 'Run4' to the new experiment. The
Experiment property of the iddata object stores experiment names.

For more information about subreferencing experiments in a multiexperiment data set,

see “Subreferencing Experiments” on page 2-57.

2-62
 Dealing with Multi-Experiment Data and Merging Models

Dealing with Multi-Experiment Data and Merging Models

This example shows how to deal with multiple experiments and merging models when
working with System Identification Toolbox™ for estimating and refining models.

Introduction

The analysis and estimation functions in System Identification Toolbox let you work with
multiple batches of data. Essentially, if you have performed multiple experiments and
recorded several input-output datasets, you can group them up into a single IDDATA
object and use them with any estimation routine.

In some cases, you may want to "split up" your (single) measurement dataset to remove
portions where the data quality is not good. For example, portion of data may be
unusable due to external disturbance or a sensor failure. In those cases, each good
portion of data may be separated out and then combined into a single multi-experiment
IDDATA object.

For example, let us look at the dataset iddemo8.mat:

load iddemo8

The name of the data object is dat, and let us view it.

dat
plot(dat)

dat =

Time domain data set with 1000 samples.

Sample time: 1 seconds

Outputs Unit (if specified)

y1

Inputs Unit (if specified)

u1

2-63
2 Data Import and Processing

We see that there are some problems with the output around sample 250-280 and around
samples 600 to 650. These might have been sensor failures.

Therefore split the data into three separate experiments and put then into a multi-
experiment data object:

d1 = dat(1:250);
d2 = dat(281:600);
d3 = dat(651:1000);
d = merge(d1,d2,d3) % merge lets you create multi-exp IDDATA object

d =

2-64
 Dealing with Multi-Experiment Data and Merging Models

Time domain data set containing 3 experiments.

Experiment Samples Sample Time

Exp1 250 1
Exp2 320 1
Exp3 350 1

Outputs Unit (if specified)

y1

Inputs Unit (if specified)

u1

The different experiments can be given other names, for example:

d.exp = {'Period 1';'Day 2';'Phase 3'}

d =

Time domain data set containing 3 experiments.

Experiment Samples Sample Time

Period 1 250 1
Day 2 320 1
Phase 3 350 1

Outputs Unit (if specified)

y1

Inputs Unit (if specified)

u1

To examine it, use plot, as in plot(d).

Performing Estimation Using Multi-Experiment Data

As mentioned before, all model estimation routines accept multi-experiment data and
take into account that they are recorded at different periods. Let us use the two first
experiments for estimation and the third one for validation:

de = getexp(d,[1,2]); % subselection is done using the command GETEXP

2-65
2 Data Import and Processing

dv = getexp(d,'Phase 3'); % using numbers or names.

m1 = arx(de,[2 2 1]);
m2 = n4sid(de,2);
m3 = armax(de,[2 2 2 1]);
compare(dv,m1,m2,m3)

Compare also accepts multiple experiments:

compare(d,m1,m2,m3) %generates 3 tabbed plots

2-66
 Dealing with Multi-Experiment Data and Merging Models

Also, spa, etfe, resid, predict, sim operate in the same way for multi-experiment
data, as they do for single experiment data.

Merging Models After Estimation

There is another way to deal with separate data sets: a model can be computed for each
set, and then the models can be merged:

m4 = armax(getexp(de,1),[2 2 2 1]);
m5 = armax(getexp(de,2),[2 2 2 1]);
m6 = merge(m4,m5); % m4 and m5 are merged into m6

This is conceptually the same as computing m from the merged set de, but it is not
numerically the same. Working on de assumes that the signal-to-noise ratios are

2-67
2 Data Import and Processing

(about) the same in the different experiments, while merging separate models makes
independent estimates of the noise levels. If the conditions are about the same for the
different experiments, it is more efficient to estimate directly on the multi-experiment
data.

We can check the models m3 and m6 that are both ARMAX models obtained on the same
data in two different ways:

[m3.a;m6.a]
[m3.b;m6.b]
[m3.c;m6.c]
compare(dv,m3,m6)

ans =

1.0000 -1.5037 0.7009

1.0000 -1.5024 0.7001

ans =

0 1.0079 0.4973
0 1.0073 0.4989

ans =

1.0000 -0.9747 0.1580

1.0000 -0.9753 0.1585

2-68
 Dealing with Multi-Experiment Data and Merging Models

Case Study: Concatenating Vs. Merging Independent Datasets

We now turn to another situation. Let us consider two data sets generated by the system
m0. The system is given by:
m0

m0 =
Discrete-time identified state-space model:
x(t+Ts) = A x(t) + B u(t) + K e(t)
y(t) = C x(t) + D u(t) + e(t)

A =
x1 x2 x3

2-69
2 Data Import and Processing

x1 0.5296 -0.476 0.1238

x2 -0.476 -0.09743 0.1354
x3 0.1238 0.1354 -0.8233

B =
u1 u2
x1 -1.146 -0.03763
x2 1.191 0.3273
x3 0 0

C =
x1 x2 x3
y1 -0.1867 -0.5883 -0.1364
y2 0.7258 0 0.1139

D =
u1 u2
y1 1.067 0
y2 0 0

K =
y1 y2
x1 0 0
x2 0 0
x3 0 0

Sample time: 1 seconds

Parameterization:
STRUCTURED form (some fixed coefficients in A, B, C).
Feedthrough: on some input channels
Disturbance component: none
Number of free coefficients: 23
Use "idssdata", "getpvec", "getcov" for parameters and their uncertainties.

Status:
Created by direct construction or transformation. Not estimated.

The data sets that have been collected are z1 and z2, obtained from m0 with different
inputs, noise and initial conditions. These datasets are obtained from iddemo8.mat that
was loaded earlier.
pause off

First data set:

2-70
 Dealing with Multi-Experiment Data and Merging Models

plot(z1) %generates a separate plot for each I/O pair if pause is on; showing only the

The second set:

plot(z2) %generates a separate plot for each I/O pair if pause is on; showing only the

2-71
2 Data Import and Processing

If we just concatenate the data we obtained:

zzl = [z1;z2]
plot(zzl)

pause on

zzl =

Time domain data set with 400 samples.

Sample time: 1 seconds

Outputs Unit (if specified)

y1

2-72
 Dealing with Multi-Experiment Data and Merging Models

y2

Inputs Unit (if specified)

u1
u2

A discrete-time state-space model can be obtained by using ssest:

ml = ssest(zzl,3,'Ts',1, 'Feedthrough', [true, false]);

Compare the bode response for models m0 and ml:

clf

2-73
2 Data Import and Processing

bode(m0,ml)
legend('show')

This is not a very good model, as observed from the four Bode plots above.

Now, instead treat the two data sets as different experiments:

zzm = merge(z1,z2)
% The model for this data can be estimated as before (watching progress this time)
mm = ssest(zzm,3,'Ts',1,'Feedthrough',[true, false], ssestOptions('Display', 'on'));

zzm =

2-74
 Dealing with Multi-Experiment Data and Merging Models

Time domain data set containing 2 experiments.

Experiment Samples Sample Time

Exp1 200 1
Exp2 200 1

Outputs Unit (if specified)

y1
y2

Inputs Unit (if specified)

u1
u2

Let us compare the Bode plots of the true system (blue)

the model from concatenated data (green) and the model from the
merged data set (red):

clf
bode(m0,'b',ml,'g',mm,'r')
legend('show')

2-75
2 Data Import and Processing

The merged data give a better model, as observed from the plot above.

Conclusions

In this example we analyzed how to use multiple data sets together for estimation of
one model. This technique is useful when you have multiple datasets from independent
experiment runs or when you segment data into multiple sets to remove bad segments.
Multiple experiments can be packaged into a single IDDATA object, which is then usable
for all estimation and analysis requirements. This technique works for both time and
frequency domain iddata.

It is also possible to merge models after estimation. This technique can be used to
"average out" independently estimated models. If the noise characteristics on multiple

2-76
 Dealing with Multi-Experiment Data and Merging Models

datasets are different, merging models after estimation works better than merging the
datasets themselves before estimation.

Additional Information

For more information on identification of dynamic systems with System Identification

Toolbox visit the System Identification Toolbox product information page.

2-77
2 Data Import and Processing

Managing iddata Objects

In this section...
“Modifying Time and Frequency Vectors” on page 2-78
“Naming, Adding, and Removing Data Channels” on page 2-80
“Subreferencing iddata Objects” on page 2-82
“Concatenating iddata Objects” on page 2-82

Modifying Time and Frequency Vectors

The iddata object stores time-domain data or frequency-domain data and has several
properties that specify the time or frequency values. To modify the time or frequency
values, you must change the corresponding property values.

Note: You can modify the property SamplingInstants by setting it to a new vector
with the length equal to the number of data samples. For more information, see
“Constructing an iddata Object for Time-Domain Data” on page 2-50.

The following tables summarize time-vector and frequency-vector properties,

respectively, and provides usage examples. In each example, data is an iddata object.

Note: Property names are not case sensitive. You do not need to type the entire property
name if the first few letters uniquely identify the property.

iddata Time-Vector Properties

Property Description Syntax Example
Ts Sampling time interval. To set the sampling interval to
0.05:
• For a single experiment, Ts
is a scalar value. set(data,'ts',0.05)
• For multiexperiement data or
with Ne experiments, Ts
is a 1-by-Ne cell array, data.ts = 0.05
and each cell contains the

2-78
 Managing iddata Objects

Property Description Syntax Example

sampling interval of the
corresponding experiment.
Tstart Starting time of the To change starting time of the
experiment. first data sample to 24:

• For a single experiment, Ts data.Tstart = 24

is a scalar value.
Time units are set by the
• For multiexperiement data property TimeUnit.
with Ne experiments, Ts
is a 1-by-Ne cell array,
and each cell contains the
sampling interval of the
corresponding experiment.
SamplingInstants Time values in the time vector, To retrieve the time vector for
computed from the properties iddata object data, use:
Tstart and Ts.
get(data,'sa')
• For a single experiment,
SamplingInstants is an To plot the input data as a
N-by-1 vector. function of time:
• For multiexperiement plot(data.sa,data.u)
data with Ne experiments,
this property is a 1-
by-Ne cell array, and Note: sa is the first two letters
each cell contains the of the SamplingInstants
sampling instants of the property that uniquely
corresponding experiment. identifies this property.
TimeUnit Unit of time. Specify To change the unit of the time
as one of the following: vector to milliseconds:
'nanoseconds',
'microseconds', data.ti = 'milliseconds'
'milliseconds',
'seconds', 'minutes',
'hours', 'days', 'weeks',
'months', and 'years'.

iddata Frequency-Vector Properties

2-79
2 Data Import and Processing

Property Description Syntax Example

Frequency Frequency values at which To specify 100 frequency
the Fourier transforms of the values in log space, ranging
signals are defined. between 0.1 and 100, use the
following syntax:
• For a single experiment,
Frequency is a scalar data.freq =
value. logspace(-1,2,100)

• For multiexperiement
data with Ne experiments,
Frequency is a 1-by-Ne
cell array, and each cell
contains the frequencies
of the corresponding
experiment.
FrequencyUnit Unit of Frequency. Specify Set the frequency unit to Hz:
as one of the following: be
one of the following: 'rad/ data.FrequencyUnit = 'Hz'
TimeUnit', 'cycles/ Note that changing the
TimeUnit', 'rad/s', 'Hz', frequency unit does not scale
'kHz', 'MHz', 'GHz', and, the frequency vector. For a
'rpm'. Default: ‘rad/ proper translation of units, use
TimeUnit’ chgFreqUnit.

For multi-experiement data

with Ne experiments, Units is
a 1-by-Ne cell array, and each
cell contains the frequency
unit for each experiment.

Naming, Adding, and Removing Data Channels

• “What Are Input and Output Channels?” on page 2-81
• “Naming Channels” on page 2-81
• “Adding Channels” on page 2-81
• “Modifying Channel Data” on page 2-82

2-80
 Managing iddata Objects

What Are Input and Output Channels?

A multivariate system might contain several input variables or several output variables,
or both. When an input or output signal includes several measured variables, these
variables are called channels.

Naming Channels

The iddata properties InputName and OutputName store the channel names for the
input and output signals. When you plot the data, you use channel names to select the
variable displayed on the plot. If you have multivariate data, it is helpful to assign a
name to each channel that describes the measured variable. For more information about
selecting channels on a plot, see “Selecting Measured and Noise Channels in Plots” on
page 14-14.

You can use the set command to specify the names of individual channels. For example,
suppose data contains two input channels (voltage and current) and one output channel
(temperature). To set these channel names, use the following syntax:

set(data,'InputName',{'Voltage','Current'},
'OutputName','Temperature')

Tip You can also specify channel names as follows:

data.una = {'Voltage','Current')
data.yna = 'Temperature'

una is equivalent to the property InputName, and yna is equivalent to OutputName.

If you do not specify channel names when you create the iddata object, the
toolbox assigns default names. By default, the output channels are named
'y1','y2',...,'yn', and the input channels are named 'u1','u2',...,'un'.

Adding Channels

You can add data channels to an iddata object.

For example, consider an iddata object named data that contains an input signal
with four channels. To add a fifth input channel, stored as the vector Input5, use the
following syntax:

2-81
2 Data Import and Processing

data.u(:,5) = Input5;

Input5 must have the same number of rows as the other input channels. In this
example, data.u(:,5) references all samples as (indicated by :) of the input signal u
and sets the values of the fifth channel. This channel is created when assigning its value
to Input5.

You can also combine input channels and output channels of several iddata objects into
one iddata object using concatenation. For more information, see “Increasing Number of
Channels or Data Points of iddata Objects” on page 2-58.

Modifying Channel Data

After you create an iddata object, you can modify or remove specific input and output
channels, if needed. You can accomplish this by subreferencing the input and output
matrices and assigning new values.

For example, suppose the iddata object data contains three output channels (named
y1, y2, and y3), and four input channels (named u1, u2, u3, and u4). To replace data
such that it only contains samples in y3, u1, and u4, type the following at the prompt:

data = data(:,3,[1 4])

The resulting data object contains one output channel and two input channels.

Subreferencing iddata Objects

See “Select Data Channels, I/O Data and Experiments in iddata Objects” on page 2-55.

Concatenating iddata Objects

See “Increasing Number of Channels or Data Points of iddata Objects” on page 2-58.

2-82
 Representing Frequency-Response Data Using idfrd Objects

Representing Frequency-Response Data Using idfrd Objects

In this section...
“idfrd Constructor” on page 2-83
“idfrd Properties” on page 2-84
“Select I/O Channels and Data in idfrd Objects” on page 2-85
“Adding Input or Output Channels in idfrd Objects” on page 2-86
“Managing idfrd Objects” on page 2-88
“Operations That Create idfrd Objects” on page 2-88

idfrd Constructor
The idfrd represents complex frequency-response data. Before you can create an
idfrd object, you must import your data as described in “Frequency-Response Data
Representation” on page 2-13.

Note: The idfrd object can only encapsulate one frequency-response data set. It does not
support the iddata equivalent of multiexperiment data.

Use the following syntax to create the data object fr_data:

fr_data = idfrd(response,f,Ts)

Suppose that ny is the number of output channels, nu is the number of input channels,
and nf is a vector of frequency values. response is an ny-by-nu-by-nf 3-D array. f is the
frequency vector that contains the frequencies of the response.Ts is the sampling time,
which is used when measuring or computing the frequency response. If you are working
with a continuous-time system, set Ts to 0.

response(ky,ku,kf), where ky, ku, and kf reference the kth output, input, and
frequency value, respectively, is interpreted as the complex-valued frequency response
from input ku to output ky at frequency f(kf).

Note: When you work at the command line, you can only create idfrd objects from
complex values of G(eiw). For a SISO system, response can be a vector.

2-83
2 Data Import and Processing

You can specify object properties when you create the idfrd object using the constructor
syntax:
fr_data = idfrd(response,f,Ts,
'Property1',Value1,...,'PropertyN',ValueN)

idfrd Properties
To view the properties of the idfrd object, you can use the get command. The following
example shows how to create an idfrd object that contains 100 frequency-response
values with a sampling time interval of 0.1 s and get its properties:
f = logspace(-1,1,100);
[mag, phase] = bode(idtf([1 .2],[1 2 1 1]),f);
response = mag.*exp(1j*phase*pi/180);
fr_data = idfrd(response,f,0.1);
get(fr_data)

FrequencyUnit: 'rad/TimeUnit'
Report: [1x1 idresults.frdest]
SpectrumData: []
CovarianceData: []
NoiseCovariance: []
InterSample: {'zoh'}
ResponseData: [1x1x100 double]
ioDelay: 0
InputDelay: 0
OutputDelay: 0
Ts: 0.1000
TimeUnit: 'seconds'
InputName: {''}
InputUnit: {''}
InputGroup: [1x1 struct]
OutputName: {''}
OutputUnit: {''}
OutputGroup: [1x1 struct]
Name: ''
Notes: {}
UserData: []
SamplingGrid: [1x1 struct]
Frequency: [100x1 double]

For a complete description of all idfrd object properties, see the idfrd reference page.

2-84
 Representing Frequency-Response Data Using idfrd Objects

To change property values for an existing idfrd object, use the set command or dot
notation. For example, to change the name of the idfrd object, type the following
command sequence at the prompt:
fr_data.Name = 'DC_Converter';

Select I/O Channels and Data in idfrd Objects

You can reference specific data values in the idfrd object using the following syntax:
fr_data(outputchannels,inputchannels)

Reference specific channels by name or by channel index.

Tip Use a colon (:) to specify all channels, and use the empty matrix ([]) to specify no
channels.

For example, the following command references frequency-response data from input
channel 3 to output channel 2:
fr_data(2,3)

You can also access the data in specific channels using channel names. To list multiple
channel names, use a cell array. For example, to retrieve the power output, and the
voltage and speed inputs, use the following syntax:
fr_data('power',{'voltage','speed'})

To retrieve only the responses corresponding to frequency values between 200 and 300,
use the following command:
fr_data_sub = fselect(fr_data,[200:300])

You can also use logical expressions to subreference data. For example, to retrieve
all frequency-response values between frequencies 1.27 and 9.3 in the idfrd object
fr_data, use the following syntax:
fr_data_sub = fselect(fr_data,fr_data.f>1.27&fr_data.f<9.3)

Tip Use end to reference the last sample number in the data. For example,
data(77:end).

2-85
2 Data Import and Processing

Note: You do not need to type the entire property name. In this example, f in
fr_data.f uniquely identifies the Frequency property of the idfrd object.

Adding Input or Output Channels in idfrd Objects

• “About Concatenating idfrd Objects” on page 2-86
• “Horizontal Concatenation of idfrd Objects” on page 2-86
• “Vertical Concatenation of idfrd Objects” on page 2-87
• “Concatenating Noise Spectrum Data of idfrd Objects” on page 2-87

About Concatenating idfrd Objects

The horizontal and vertical concatenation of idfrd objects combine information in the
ResponseData properties of these objects. ResponseData is an ny-by-nu-by-nf array
that stores the response of the system, where ny is the number of output channels, nu is
the number of input channels, and nf is a vector of frequency values (see “Properties”).

Horizontal Concatenation of idfrd Objects

The following syntax creates a new idfrd object data that contains the horizontal
concatenation of data1,data2,...,dataN:
data = [data1,data2,...,dataN]

data contains the frequency responses from all of the inputs in

data1,data2,...,dataN to the same outputs. The following diagram is a graphical
representation of horizontal concatenation of frequency-response data. The (j,i,:)
vector of the resulting response data represents the frequency response from the ith
input to the jth output at all frequencies.

u1 Data 1 y1 Data 2 y1
2-by-2-by-nf u3 2-by-1-by-nf
u2 y2 y2

u1 Horizonal Concatenation y1
Combined of Data 1 and Data 2 Same
inputs u2 outputs
2-by-3-by-nf y2
u3

2-86
 Representing Frequency-Response Data Using idfrd Objects

Note: Horizontal concatenation of idfrd objects requires that they have the same
outputs and frequency vectors. If the output channel names are different and their
dimensions are the same, the concatenation operation resets the output names to their
default values.

Vertical Concatenation of idfrd Objects

The following syntax creates a new idfrd object data that contains the vertical
concatenation of data1,data2,...,dataN:

data = [data1;data2;... ;dataN]

The resulting idfrd object data contains the frequency responses from the same inputs
in data1,data2,...,dataN to all the outputs. The following diagram is a graphical
representation of vertical concatenation of frequency-response data. The (j,i,:) vector
of the resulting response data represents the frequency response from the ith input to
the jth output at all frequencies.

u1 Data 1 y1 u1 Data 1 y3
u2 2-by-2-by-nf y2 u2 1-by-2-by-nf

y1
Same u1 Vertical Concatenation Combined
of Data 1 and Data 2 y2
inputs u2 outputs
3-by-2-by-nf y3

Note: Vertical concatenation of idfrd objects requires that they have the same inputs
and frequency vectors. If the input channel names are different and their dimensions are
the same, the concatenation operation resets the input names to their default values.

Concatenating Noise Spectrum Data of idfrd Objects

When the SpectrumData property of individual idfrd objects is not empty, horizontal
and vertical concatenation handle SpectrumData, as follows.

2-87
2 Data Import and Processing

In case of horizontal concatenation, there is no meaningful way to combine the

SpectrumData of individual idfrd objects and the resulting SpectrumData property
is empty. An empty property results because each idfrd object has its own set of noise
channels, where the number of noise channels equals the number of outputs. When
the resulting idfrd object contains the same output channels as each of the individual
idfrd objects, it cannot accommodate the noise data from all the idfrd objects.

In case of vertical concatenation, the toolbox concatenates individual noise models

diagonally. The following shows that data.SpectrumData is a block diagonal matrix of
the power spectra and cross spectra of the output noise in the system:

Ê data1 .s
Á 0 ˆ˜
data.s = Á O ˜
ÁÁ ˜
Ë 0 dataN .s ˜¯

s in data.s is the abbreviation for the SpectrumData property name.

Managing idfrd Objects

• “Subreferencing idfrd Objects” on page 2-88
• “Concatenating idfrd Objects” on page 2-88

Subreferencing idfrd Objects

See “Select I/O Channels and Data in idfrd Objects” on page 2-85.

Concatenating idfrd Objects

See “Adding Input or Output Channels in idfrd Objects” on page 2-86.

Operations That Create idfrd Objects

The following operations create idfrd objects:

• Constructing idfrd objects.

• Estimating nonparametric models using etfe, spa, and spafdr. For more
information, see “Identifying Frequency-Response Models” on page 3-8.
• Converting the Control System Toolbox frd object. For more information, see “Using
Identified Models for Control Design Applications” on page 12-2.

2-88
 Representing Frequency-Response Data Using idfrd Objects

• Converting any linear dynamic system using the idfrd command.

For example:
sys_idpoly = idpoly([1 2 1],[0 2],'Ts',1);
G = idfrd(sys_idpoly,linspace(0,pi,128))

G =
IDFRD model.
Contains Frequency Response Data for 1 output(s) and 1 input(s), and the spectra for
Response data and disturbance spectra are available at 128 frequency points, ranging

Sample time: 1 seconds

Status:
Created by direct construction or transformation. Not estimated.

2-89
2 Data Import and Processing

Analyzing Data Quality

In this section...
“Is Your Data Ready for Modeling?” on page 2-90
“Plotting Data in the App Versus at the Command Line” on page 2-91
“How to Plot Data in the App” on page 2-91
“How to Plot Data at the Command Line” on page 2-96
“How to Analyze Data Using the advice Command” on page 2-98

Is Your Data Ready for Modeling?

Before you start estimating models from data, you should check your data for the
presence of any undesirable characteristics. For example, you might plot the data to
identify drifts and outliers. You plot analysis might lead you to preprocess your data
before model estimation.

The following data plots are available in the toolbox:

• Time plot — Shows data values as a function of time.

Tip You can infer time delays from time plots, which are required inputs to most
parametric models. A time delay is the time interval between the change in input and
the corresponding change in output.
• Spectral plot — Shows a periodogram that is computed by taking the absolute squares
of the Fourier transforms of the data, dividing by the number of data points, and
multiplying by the sampling interval.
• Frequency-response plot — For frequency-response data, shows the amplitude and
phase of the frequency-response function on a Bode plot. For time- and frequency-
domain data, shows the empirical transfer function estimate (see etfe) .

See Also

“How to Analyze Data Using the advice Command” on page 2-98

“Ways to Prepare Data for System Identification” on page 2-6

2-90
 Analyzing Data Quality

Plotting Data in the App Versus at the Command Line

The plots you create using the System Identification app provide options that are specific
to the System Identification Toolbox product, such as selecting specific channel pairs
in a multivariate signals or converting frequency units between Hertz and radians per
second. For more information, see “How to Plot Data in the App” on page 2-91.

The plots you create using the plot commands, such as plot, and bode are displayed in
the standard MATLAB Figure window, which provides options for formatting, saving,
printing, and exporting plots to a variety of file formats. To learn about plotting at the
command line, see “How to Plot Data at the Command Line” on page 2-96. For more
information about working with Figure window, see “Graphics”.

How to Plot Data in the App

• “How to Plot Data in the App” on page 2-91
• “Manipulating a Time Plot” on page 2-92
• “Manipulating Data Spectra Plot” on page 2-94
• “Manipulating a Frequency Function Plot” on page 2-95

How to Plot Data in the App

After importing data into the System Identification app, as described in “Represent
Data”, you can plot the data.

To create one or more plots, select the corresponding check box in the Data Views area
of the System Identification app.

An active data icon has a thick line in the icon, while an inactive data set has a thin line.
Only active data sets appear on the selected plots. To toggle including and excluding data
on a plot, click the corresponding icon in the System Identification app. Clicking the data
icon updates any plots that are currently open.

When you have several data sets, you can view different input-output channel pair
by selecting that pair from the Channel menu. For more information about selecting
different input and output pairs, see “Selecting Measured and Noise Channels in Plots”
on page 14-14.

2-91
2 Data Import and Processing

Thick lines indicate

active data sets
included in plots.

All three available

data plots are
selected.

In this example, data and dataff are active and appear on the three selected plots.

To close a plot, clear the corresponding check box in the System Identification app.

Tip To get information about working with a specific plot, select a help topic from the
Help menu in the plot window.

Manipulating a Time Plot

The Time plot only shows time-domain data. In this example, data1 is displayed on
the time plot because, of the three data sets, it is the only one that contains time-domain
input and output.

2-92
 Analyzing Data Quality

Time Plot of data1

The following table summarizes options that are specific to time plots, which you can
select from the plot window menus. For general information about working with System
Identification Toolbox plots, see “Working with Plots” on page 14-11.

Time Plot Options

Action Command
Toggle input display between piece-wise Select Style > Staircase input for zero-
continuous (zero-order hold) and linear order hold or Style > Regular input for
interpolation (first-order hold) between first-order hold.
samples.

Note: This option only affects the display

and not the intersample behavior specified
when importing the data.

2-93
2 Data Import and Processing

Manipulating Data Spectra Plot

The Data spectra plot shows a periodogram or a spectral estimate of data1 and
data3fd.

The periodogram is computed by taking the absolute squares of the Fourier transforms
of the data, dividing by the number of data points, and multiplying by the sampling
interval. The spectral estimate for time-domain data is a smoothed spectrum calculated
using spa. For frequency-domain data, the Data spectra plot shows the square of the
absolute value of the actual data, normalized by the sampling interval.

The top axes show the input and the bottom axes show the output. The vertical axis of
each plot is labeled with the corresponding channel name.

Periodograms of data1 and data3fd

Data Spectra Plot Options

2-94
 Analyzing Data Quality

Action Command
Toggle display between periodogram and Select Options > Periodogram or
spectral estimate. Options > Spectral analysis.
Change frequency units. Select Style > Frequency (rad/s) or Style
> Frequency (Hz).
Toggle frequency scale between linear and Select Style > Linear frequency scale or
logarithmic. Style > Log frequency scale.
Toggle amplitude scale between linear and Select Style > Linear amplitude scale or
logarithmic. Style > Log amplitude scale.

Manipulating a Frequency Function Plot

For time-domain data, the Frequency function plot shows the empirical transfer
function estimate (etfe). For frequency-domain data, the plot shows the ratio of output
to input data.

The frequency-response plot shows the amplitude and phase plots of the corresponding
frequency response. For more information about frequency-response data, see
“Frequency-Response Data Representation” on page 2-13.

2-95
2 Data Import and Processing

Frequency Functions of data1 and data3fd

Frequency Function Plot Options

Action Command
Change frequency units. Select Style > Frequency (rad/s) or Style
> Frequency (Hz).
Toggle frequency scale between linear and Select Style > Linear frequency scale or
logarithmic. Style > Log frequency scale.
Toggle amplitude scale between linear and Select Style > Linear amplitude scale or
logarithmic. Style > Log amplitude scale.

How to Plot Data at the Command Line

The following table summarizes the commands available for plotting time-domain,
frequency-domain, and frequency-response data.

Commands for Plotting Data

2-96
 Analyzing Data Quality

Command Description Example

bode, bodeplot For frequency-response data only. To plot idfrd data:
Shows the magnitude and phase
of the frequency response on a bode(idfrd_data)
logarithmic frequency scale of a or:
Bode plot. bodeplot(idfrd_data)
plot The type of plot corresponds to To plot iddata or idfrd data:
the type of data. For example,
plotting time-domain data plot(data)
generates a time plot, and
plotting frequency-response data
generates a frequency-response
plot.

When plotting time- or frequency-

domain inputs and outputs, the
top axes show the output and the
bottom axes show the input.

All plot commands display the data in the standard MATLAB Figure window. For more
information about working with the Figure window, see “Graphics”.

To plot portions of the data, you can subreference specific samples (see “Select Data
Channels, I/O Data and Experiments in iddata Objects” on page 2-55 and “Select I/O
Channels and Data in idfrd Objects” on page 2-85. For example:
plot(data(1:300))

For time-domain data, to plot only the input data as a function of time, use the following
syntax:
plot(data(:,[],:)

When data.intersample = 'zoh', the input is piece-wise constant between sampling

points on the plot. For more information about properties, see the iddata reference page.

You can generate plots of the input data in the time domain using:
plot(data.SamplingInstants,data.u)

To plot frequency-domain data, you can use the following syntax:

2-97
2 Data Import and Processing

semilogx(data.Frequency,abs(data.u))

When you specify to plot a multivariable iddata object, each input-output combination is
displayed one at a time in the same MATLAB Figure window. You must press Enter to
update the Figure window and view the next channel combination. To cancel the plotting
operation, press Ctrl+C.

Tip To plot specific input and output channels, use plot(data(:,ky,ku)), where ky
and ku are specific output and input channel indexes or names. For more information
about subreferencing channels, see “Subreferencing Data Channels” on page 2-56.

To plot several iddata sets d1,...,dN, use plot(d1,...,dN). Input-output channels

with the same experiment name, input name, and output name are always plotted in the
same plot.

How to Analyze Data Using the advice Command

You can use the advice command to analyze time- or frequency- domain data before
estimating a model. The resulting report informs you about the possible need to
preprocess the data and identifies potential restrictions on the model accuracy. You
should use these recommendations in combination with plotting the data and validating
the models estimated from this data.

Note: advice does not support frequency-response data.

Before applying the advice command to your data, you must have represented your data
as an iddata object. For more information, see “Representing Time- and Frequency-
Domain Data Using iddata Objects” on page 2-50.

If you are using the System Identification app, you must export your data to the
MATLAB workspace before you can use the advice command on this data. For more
information about exporting data, see “Exporting Models from the App to the MATLAB
Workspace” on page 14-10.

Use the following syntax to get advice about an iddata object data:

advice(data)

2-98
 Analyzing Data Quality

For more information about the advice syntax, see the advice reference page.

Advice provide guidance for these kinds of questions:

• Does it make sense to remove constant offsets and linear trends from the data?
• What are the excitation levels of the signals and how does this affects the model
orders?
• Is there an indication of output feedback in the data? When feedback is present in the
system, only prediction-error methods work well for estimating closed-loop data.
• Is there an indication of nonlinearity in the process that generated the data?

See Also

advice

delayest

detrend

feedback

pexcit

2-99
2 Data Import and Processing

Selecting Subsets of Data

In this section...
“Why Select Subsets of Data?” on page 2-100
“Extract Subsets of Data Using the App” on page 2-100
“Extract Subsets of Data at the Command Line” on page 2-102

Why Select Subsets of Data?

You can use data selection to create independent data sets for estimation and validation.

You can also use data selection as a way to clean the data and exclude parts with noisy
or missing information. For example, when your data contains missing values, outliers,
level changes, and disturbances, you can select one or more portions of the data that are
suitable for identification and exclude the rest.

If you only have one data set and you want to estimate linear models, you should split
the data into two portions to create two independent data sets for estimation and
validation, respectively. Splitting the data is selecting parts of the data set and saving
each part independently.

You can merge several data segments into a single multiexperiment data set and identify
an average model. For more information, see “Create Data Sets from a Subset of Signal
Channels” on page 2-33 or “Representing Time- and Frequency-Domain Data Using
iddata Objects” on page 2-50.

Note: Subsets of the data set must contain enough samples to adequately represent the
system, and the inputs must provide suitable excitation to the system.

Selecting potions of frequency-domain data is equivalent to filtering the data. For more
information about filtering, see “Filtering Data” on page 2-123.

Extract Subsets of Data Using the App

• “Ways to Select Data in the App” on page 2-101
• “Selecting a Range for Time-Domain Data” on page 2-101
• “Selecting a Range of Frequency-Domain Data” on page 2-102

2-100
 Selecting Subsets of Data

Ways to Select Data in the App

You can use System Identification app to select ranges of data on a time-domain or
frequency-domain plot. Selecting data in the frequency domain is equivalent to passband-
filtering the data.

After you select portions of the data, you can specify to use one data segment for
estimating models and use the other data segment for validating models. For more
information, see “Specify Estimation and Validation Data in the App” on page 2-30.

Note: Selecting <--Preprocess > Quick start performs the following actions
simultaneously:

• Remove the mean value from each channel.

• Split the data into two parts.

• Specify the first part as estimation data (or Working Data).
• Specify the second part as Validation Data.

Selecting a Range for Time-Domain Data

You can select a range of data values on a time plot and save it as a new data set in the
System Identification app.

Note: Selecting data does not extract experiments from a data set containing
multiple experiments. For more information about multiexperiment data, see “Create
Multiexperiment Data Sets in the App” on page 2-35.

To extract a subset of time-domain data and save it as a new data set:

1 Import time-domain data into the System Identification app, as described in “Create
Data Sets from a Subset of Signal Channels” on page 2-33.
2 Drag the data set you want to subset to the Working Data area.
3 If your data contains multiple I/O channels, in the Channel menu, select the
channel pair you want to view. The upper plot corresponds to the input signal, and
the lower plot corresponds to the output signal.

2-101
2 Data Import and Processing

Although you view only one I/O channel pair at a time, your data selection is applied
to all channels in this data set.
4 Select the data of interest in either of the following ways:

• Graphically — Draw a rectangle on either the input-signal or the output-signal

plot with the mouse to select the desired time interval. Your selection appears on
both plots regardless of the plot on which you draw the rectangle. The Time span
and Samples fields are updated to match the selected region.
• By specifying the Time span — Edit the beginning and the end times in seconds.
The Samples field is updated to match the selected region. For example:

28.5 56.8
• By specifying the Samples range — Edit the beginning and the end indices of the
sample range. The Time span field is updated to match the selected region. For
example:

342 654

Note: To clear your selection, click Revert.

5 In the Data name field, enter the name of the data set containing the selected data.
6 Click Insert. This action saves the selection as a new data set and adds it to the
Data Board.
7 To select another range, repeat steps 4 to 6.

Selecting a Range of Frequency-Domain Data

Selecting a range of values in frequency domain is equivalent to filtering the data. For
more information about data filtering, see “Filtering Frequency-Domain or Frequency-
Response Data in the App” on page 2-126.

Extract Subsets of Data at the Command Line

Selecting ranges of data values is equivalent to subreferencing the data.

For more information about subreferencing time-domain and frequency-domain data, see
“Select Data Channels, I/O Data and Experiments in iddata Objects” on page 2-55.

2-102
 Selecting Subsets of Data

For more information about subreferencing frequency-response data, see “Select I/O
Channels and Data in idfrd Objects” on page 2-85.

2-103
2 Data Import and Processing

Handling Missing Data and Outliers

In this section...
“Handling Missing Data” on page 2-104
“Handling Outliers” on page 2-105
“Extract and Model Specific Data Segments” on page 2-106
“See Also” on page 2-107

Handling Missing Data

Data acquisition failures sometimes result in missing measurements both in the input
and the output signals. When you import data that contains missing values using the
MATLAB Import Wizard, these values are automatically set to NaN. NaN serves as a
flag for nonexistent or undefined data. When you plot data on a time-plot that contains
missing values, gaps appear on the plot where missing data exists.

You can use misdata to estimate missing values. This command linearly interpolates
missing values to estimate the first model. Then, it uses this model to estimate the
missing data as parameters by minimizing the output prediction errors obtained from
the reconstructed data. You can specify the model structure you want to use in the
misdata argument or estimate a default-order model using the n4sid method. For more
information, see the misdata reference page.

Note: You can only use misdata on time-domain data stored in an iddata object.
For more information about creating iddata objects, see “Representing Time- and
Frequency-Domain Data Using iddata Objects” on page 2-50.

For example, suppose y and u are output and input signals that contain NaNs. This data
is sampled at 0.2 s. The following syntax creates a new iddata object with these input
and output signals.
dat = iddata(y,u,0.2) % y and u contain NaNs
% representing missing data

Apply the misdata command to the new data object. For example:

dat1 = misdata(dat);

2-104
 Handling Missing Data and Outliers

plot(dat,dat1) % Check how the missing data

% was estimated on a time plot

Handling Outliers
Malfunctions can produce errors in measured values, called outliers. Such outliers
might be caused by signal spikes or by measurement malfunctions. If you do not remove
outliers from your data, this can adversely affect the estimated models.

To identify the presence of outliers, perform one of the following tasks:

• Before estimating a model, plot the data on a time plot and identify values that
appear out of range.
• After estimating a model, plot the residuals and identify unusually large values. For
more information about plotting residuals, see “Residual Analysis” on page 10-23.
Evaluate the original data that is responsible for large residuals. For example, for the
model Model and validation data Data, you can use the following commands to plot
the residuals:

% Compute the residuals

E = resid(Model,Data)
% Plot the residuals
plot(E)

Next, try these techniques for removing or minimizing the effects of outliers:

• Extract the informative data portions into segments and merge them into one
multiexperiment data set (see “Extract and Model Specific Data Segments” on page
2-106). For more information about selecting and extracting data segments, see
“Selecting Subsets of Data” on page 2-100.

Tip The inputs in each of the data segments must be consistently exciting the system.
Splitting data into meaningful segments for steady-state data results in minimum
information loss. Avoid making data segments too small.
• Manually replace outliers with NaNs and then use the misdata command to
reconstruct flagged data. This approach treats outliers as missing data and is
described in “Handling Missing Data” on page 2-104. Use this method when your
data contains several inputs and outputs, and when you have difficulty finding
reliable data segments in all variables.

2-105
2 Data Import and Processing

• Remove outliers by prefiltering the data for high-frequency content because outliers
often result from abrupt changes. For more information about filtering, see “Filtering
Data” on page 2-123.

Note: The estimation algorithm can handle outliers by assigning a smaller weight
to outlier data. A robust error criterion applies an error penalty that is quadratic for
small and moderate prediction errors, and is linear for large prediction errors. Because
outliers produce large prediction errors, this approach gives a smaller weight to the
corresponding data points during model estimation. Set the ErrorThreshold estimation
option (see Advanced.ErrorThreshold in, for example, polyestOptions) to a
nonzero value to activate the correction for outliers in the estimation algorithm.

Extract and Model Specific Data Segments

This example shows how to create a multi-experiment, time-domain data set by merging
only the accurate data segments and ignoring the rest.

Assume that the data has poor or no measurements for some sample ranges (for
example 341–499). You cannot simply concatenate the good data segments because the
transients at the connection points compromise the model. Instead, you must create a
multiexperiment iddata object, where each experiment corresponds to a good segment of
data, as follows:

% Plot the data in a MATLAB Figure window

plot(data)

% Create multiexperiment data set

% by merging data segments
datam = merge(data(1:340),...
data(500:897),...
data(1001:1200),...
data(1550:2000));

% Model the multiexperiment data set

% using "experiments" 1, 2, and 4
m = n4sid(getexp(datam,[1,2,4]))

% Validate the model by comparing its output to

% the output data of experiment 3
compare(getexp(datam,3),m)

2-106
 Handling Missing Data and Outliers

See Also
To learn more about the theory of handling missing data and outliers, see the chapter
on preprocessing data in System Identification: Theory for the User, Second Edition, by
Lennart Ljung, Prentice Hall PTR, 1999.

2-107
2 Data Import and Processing

Handling Offsets and Trends in Data

In this section...
“When to Detrend Data” on page 2-108
“Alternatives for Detrending Data in App or at the Command-Line” on page 2-109
“Next Steps After Detrending” on page 2-110

When to Detrend Data

Detrending is removing means, offsets, or linear trends from regularly sampled time-
domain input-output data signals. This data processing operation helps you estimate
more accurate linear models because linear models cannot capture arbitrary differences
between the input and output signal levels. The linear models you estimate from
detrended data describe the relationship between the change in input signals and the
change in output signals.

For steady-state data, you should remove mean values and linear trends from both input
and output signals.

For transient data, you should remove physical-equilibrium offsets measured prior to the
excitation input signal.

Remove one linear trend or several piecewise linear trends when the levels drift during
the experiment. Signal drift is considered a low-frequency disturbance and can result in
unstable models.

You should not detrend data before model estimation when you want:

• Linear models that capture offsets essential for describing important system
dynamics. For example, when a model contains integration behavior, you could
estimate a low-order transfer function (process model) from nondetrended data. For
more information, see “Identifying Process Models”.
• Nonlinear black-box models, such as nonlinear ARX or Hammerstein-Wiener models.
For more information, see “Nonlinear Model Identification”.

Tip When signals vary around a large signal level, you can improve computational
accuracy of nonlinear models by detrending the signal means.

2-108
 Examples

• Nonlinear ODE parameters (nonlinear grey-box models). For more information, see
“Estimating Nonlinear Grey-Box Models”.

To simulate or predict the linear model response at the system operating conditions,
you can restore the removed trend to the simulated or predicted model output using the
retrend command.

For more information about handling drifts in the data, see the chapter on preprocessing
data in System Identification: Theory for the User, Second Edition, by Lennart Ljung,
Prentice Hall PTR, 1999.

Examples
“How to Detrend Data Using the App” on page 2-111

“How to Detrend Data at the Command Line” on page 2-112

Alternatives for Detrending Data in App or at the Command-Line

You can detrend data using the System Identification app and at the command line using
the detrend command.

Both the app and the command line let you subtract the mean values and one linear
trend from steady-state time-domain signals.

However, the detrend command provides the following additional functionality (not
available in the app):

• Subtracting piecewise linear trends at specified breakpoints. A breakpoint is a time

value that defines the discontinuities between successive linear trends.
• Subtracting arbitrary offsets and linear trends from transient data signals.
• Saving trend information to a variable so that you can apply it to multiple data sets.

To learn how to detrend data, see:

• “How to Detrend Data Using the App” on page 2-111

• “How to Detrend Data at the Command Line” on page 2-112

As an alternative to detrending data beforehand, you can specify the offsets levels as
estimation options and use them directly with the estimation command.

2-109
2 Data Import and Processing

For example, suppose your data has an input offset, u0, and an output offset, y0. There
are two ways to perform a linear model estimation (say, a transfer function model
estimation) using this data:

• Using detrend:
T=getTrend(data)
T.InputOffset = u0;
T.OutputOffset = y0;
datad = detrend(data, T);

model = tfest(datad, np);

• Specify offsets as estimation options:
opt = tfestOptions('InputOffset',u0, 'OutputOffset', y0);

model = tfest(data, np, opt)

The advantage of this approach is that there is a record of offset levels in the model
in model.Report.OptionsUsed. The limitation of this approach is that it cannot
handle linear trends, which can only be removed from the data by using detrend.

Next Steps After Detrending

After detrending your data, you might do the following:

• Perform other data preprocessing operations. See “Ways to Prepare Data for System
Identification” on page 2-6.
• Estimate a linear model. See “Linear Model Identification”.

2-110
 How to Detrend Data Using the App

How to Detrend Data Using the App

Before you can perform this task, you must have regularly-sampled, steady-state time-
domain data imported into the System Identification app. See “Import Time-Domain
Data into the App” on page 2-16). For transient data, see “How to Detrend Data at the
Command Line” on page 2-112.

Tip You can use the shortcut Preprocess > Quick start to perform several operations:
remove the mean value from each signal, split data into two halves, specify the first
half as model estimation data (or Working Data), and specify the second half as model
Validation Data.

1 In the System Identification app, drag the data set you want to detrend to the
Working Data rectangle.
2 Detrend the data.

• To remove linear trends, select Preprocess > Remove trends.

• To remove mean values from each input and output data signal, select
Preprocess > Remove means.

More About
“Handling Offsets and Trends in Data” on page 2-108

2-111
2 Data Import and Processing

How to Detrend Data at the Command Line

In this section...
“Detrending Steady-State Data” on page 2-112
“Detrending Transient Data” on page 2-112
“See Also” on page 2-113

Detrending Steady-State Data

Before you can perform this task, you must have time-domain data as an iddata object.
See “Representing Time- and Frequency-Domain Data Using iddata Objects” on page
2-50.

Note: If you plan to estimate models from this data, your data must be regularly
sampled.

Use the detrend command to remove the signal means or linear trends:

[data_d,T]=detrend(data,Type)

where data is the data to be detrended. The second input argument Type=0 removes
signal means or Type=1 removes linear trends. data_d is the detrended data. T is
a TrendInfo object that stores the values of the subtracted offsets and slopes of the
removed trends.

More About
“Handling Offsets and Trends in Data” on page 2-108

Detrending Transient Data

Before you can perform this task, you must have

• Time-domain data as an iddata object. See “Representing Time- and Frequency-

Domain Data Using iddata Objects” on page 2-50.

2-112
 More About

Note: If you plan to estimate models from this data, your data must be regularly
sampled.
• Values of the offsets you want to remove from the input and output data. If you do not
know these values, visually inspect a time plot of your data. For more information, see
“How to Plot Data at the Command Line” on page 2-96.

1 Create a default object for storing input-output offsets that you want to remove from
the data.
T = getTrend(data)

where T is a TrendInfo object.

2 Assign offset values to T.

T.InputOffset=I_value;
T.OutputOffset=O_value;

where I_value is the input offset value, and O_value is the input offset value.
3 Remove the specified offsets from data.

data_d = detrend(data,T)

where the second input argument T stores the offset values as its properties.

More About
“Handling Offsets and Trends in Data” on page 2-108

See Also
detrend

TrendInfo

2-113
2 Data Import and Processing

Resampling Data
In this section...
“What Is Resampling?” on page 2-114
“Resampling Data Without Aliasing Effects” on page 2-115
“See Also” on page 2-119

What Is Resampling?
Resampling data signals in the System Identification Toolbox product applies an
antialiasing (lowpass) FIR filter to the data and changes the sampling rate of the signal
by decimation or interpolation.

If your data is sampled faster than needed during the experiment, you can decimate
it without information loss. If your data is sampled more slowly than needed, there
is a possibility that you miss important information about the dynamics at higher
frequencies. Although you can resample the data at a higher rate, the resampled values
occurring between measured samples do not represent new measured information about
your system. Instead of resampling, repeat the experiment using a higher sampling rate.

Tip You should decimate your data when it contains high-frequency noise outside the
frequency range of the system dynamics.

Resampling takes into account how the data behaves between samples, which you specify
when you import the data into the System Identification app (zero-order or first-order
hold). For more information about the data properties you specify before importing the
data, see “Represent Data”.

You can resample data using the System Identification app or the resample command.
You can only resample time-domain data at uniform time intervals.

Examples
“Resampling Data Using the App” on page 2-120

“Resampling Data at the Command Line” on page 2-121

2-114
 Examples

Resampling Data Without Aliasing Effects

Typically, you decimate a signal to remove the high-frequency contributions that result
from noise from the total energy. Ideally, you want to remove the energy contribution due
to noise and preserve the energy density of the signal.

The command resample performs the decimation without aliasing effects. This
command includes a factor of T to normalize the spectrum and preserve the energy
density after decimation. For more information about spectrum normalization, see
“Spectrum Normalization” on page 3-13.

If you use manual decimation instead of resample—by picking every fourth sample from
the signal, for example—the energy contributions from higher frequencies are folded
back into the lower frequencies("aliasing"). Because the total signal energy is preserved
by this operation and this energy must now be squeezed into a smaller frequency range,
the amplitude of the spectrum at each frequency increases. Thus, the energy density of
the decimated signal is not constant.

This example shows how resample avoids folding effects.

Construct a fourth-order moving-average process.

m0 = idpoly(1,[ ],[1 1 1 1]);

m0 is a time-series model with no inputs.

Generate error signal.

e = idinput(2000,'rgs');

Simulate the output using the error signal.

sim_opt = simOptions('AddNoise',true,'NoiseData',e);
y = sim(m0,zeros(2000,0),sim_opt);
y = iddata(y,[],1);

Estimate the signal spectrum.

g1 = spa(y);

Estimate the spectrum of the modified signal including every fourth sample of the
original signal. This command automatically sets Ts to 4.
g2 = spa(y(1:4:2000));

2-115
2 Data Import and Processing

Plot the frequency response to view folding effects.

h = spectrumplot(g1,g2,g1.Frequency);
opt = getoptions(h);
opt.FreqScale='linear';
opt.FreqUnits='Hz';
setoptions(h,opt);

Estimate the spectrum after prefiltering that does not introduce folding effects.

g3 = spa(resample(y,1,4));
figure
spectrumplot(g1,g3,g1.Frequency,opt);

2-116
 Examples

Use resample to decimate the signal before estimating the spectrum and plot the
frequency response.

g3 = spa(resample(y,1,4));
figure
spectrumplot(g1,g3,g1.Frequency,opt);

2-117
2 Data Import and Processing

The plot shows that the estimated spectrum of the resampled signal has the same
amplitude as the original spectrum. Thus, there is no indication of folding effects when
you use resample to eliminate aliasing.

Examples
“Resampling Data Using the App” on page 2-120

“Resampling Data at the Command Line” on page 2-121

2-118
 Examples

See Also
For a detailed discussion about handling disturbances, see the chapter on preprocessing
data in System Identification: Theory for the User, Second Edition, by Lennart Ljung,
Prentice Hall PTR, 1999.

2-119
2 Data Import and Processing

Resampling Data Using the App

Use the System Identification app to resample time-domain data. To specify additional
options, such as the prefilter order, see “Resampling Data at the Command Line” on page
2-121.

The System Identification app uses idresamp to interpolate or decimate the data. For
more information about this command, type help idresamp at the prompt.

To create a new data set by resampling the input and output signals:

1 Import time-domain data into the System Identification app, as described in “Create
Data Sets from a Subset of Signal Channels” on page 2-33.
2 Drag the data set you want to resample to the Working Data area.
3 In the Resampling factor field, enter the factor by which to multiply the current
sampling interval:

• For decimation (fewer samples), enter a factor greater than 1 to increase the
sampling interval by this factor.
• For interpolation (more samples), enter a factor less than 1 to decrease the
sampling interval by this factor.

Default = 1.
4 In the Data name field, type the name of the new data set. Choose a name that is
unique in the Data Board.
5 Click Insert to add the new data set to the Data Board in the System Identification
Toolbox window.
6 Click Close to close the Resample dialog box.

More About
“Resampling Data” on page 2-114

2-120
 Resampling Data at the Command Line

Resampling Data at the Command Line

Use resample to decimate and interpolate time-domain iddata objects. You can specify
the order of the antialiasing filter as an argument.

Note: resample uses the Signal Processing Toolbox™ command, when this toolbox
is installed on your computer. If this toolbox is not installed, use idresamp instead.
idresamp only lets you specify the filter order, whereas resample also lets you specify
filter coefficients and the design parameters of the Kaiser window.

To create a new iddata object datar by resampling data, use the following syntax:
datar = resample(data,P,Q,filter_order)

In this case, P and Q are integers that specify the new sampling interval: the new
sampling interval is Q/P times the original one. You can also specify the order of the
resampling filter as a fourth argument filter_order, which is an integer (default is
10). For detailed information about resample, see the corresponding reference page.

For example, resample(data,1,Q) results in decimation with the sampling interval

modified by a factor Q.

The next example shows how you can increase the sampling rate by a factor of 1.5 and
compare the signals:
plot(u)
ur = resample(u,3,2);
plot(u,ur)

When the Signal Processing Toolbox product is not installed, using resample calls
idresamp instead.

idresamp uses the following syntax:

datar = idresamp(data,R,filter_order)

In this case, R=Q/P, which means that data is interpolated by a factor P and then
decimated by a factor Q. To learn more about idresamp, type help idresamp.

The data.InterSample property of the iddata object is taken into account during
resampling (for example, first-order hold or zero-order hold). For more information, see
“iddata Properties” on page 2-52.

2-121
2 Data Import and Processing

More About
“Resampling Data” on page 2-114

2-122
 Filtering Data

Filtering Data
In this section...
“Supported Filters” on page 2-123
“Choosing to Prefilter Your Data” on page 2-123
“See Also” on page 2-124

Supported Filters
You can filter the input and output signals through a linear filter before estimating a
model in the System Identification app or at the command line. How you want to handle
the noise in the system determines whether it is appropriate to prefilter the data.

The filter available in the System Identification app is a fifth-order (passband)

Butterworth filter. If you need to specify a custom filter, use the idfilt command.

Examples
“How to Filter Data Using the App” on page 2-125

“How to Filter Data at the Command Line” on page 2-128

Choosing to Prefilter Your Data

Prefiltering data can help remove high-frequency noise or low-frequency disturbances
(drift). The latter application is an alternative to subtracting linear trends from the data,
as described in “Handling Offsets and Trends in Data” on page 2-108.

In addition to minimizing noise, prefiltering lets you focus your model on specific
frequency bands. The frequency range of interest often corresponds to a passband over
the breakpoints on a Bode plot. For example, if you are modeling a plant for control-
design applications, you might prefilter the data to specifically enhance frequencies
around the desired closed-loop bandwidth.

Prefiltering the input and output data through the same filter does not change the input-
output relationship for a linear system. However, prefiltering does change the noise
characteristics and affects the estimated model of the system.

2-123
2 Data Import and Processing

To get a reliable noise model, avoid prefiltering the data. Instead, set the Focus property
of the estimation algorithm to Simulation.

Note: When you prefilter during model estimation, the filtered data is used to only model
the input-to-output dynamics. However, the disturbance model is calculated from the
unfiltered data.

Examples
“How to Filter Data Using the App” on page 2-125

“How to Filter Data at the Command Line” on page 2-128

See Also
To learn how to filter data during linear model estimation instead, you can set the Focus
property of the estimation algorithm to Filter and specify the filter characteristics.

For more information about prefiltering data, see the chapter on preprocessing data in
System Identification: Theory for the User, Second Edition, by Lennart Ljung, Prentice
Hall PTR, 1999.

For practical examples of prefiltering data, see the section on posttreatment of data in
Modeling of Dynamic Systems, by Lennart Ljung and Torkel Glad, Prentice Hall PTR,
1994.

2-124
 How to Filter Data Using the App

How to Filter Data Using the App

In this section...
“Filtering Time-Domain Data in the App” on page 2-125
“Filtering Frequency-Domain or Frequency-Response Data in the App” on page 2-126

Filtering Time-Domain Data in the App

The System Identification app lets you filter time-domain data using a fifth-order
Butterworth filter by enhancing or selecting specific passbands.

To create a filtered data set:

1 Import time-domain data into the System Identification app, as described in

“Represent Data”.
2 Drag the data set you want to filter to the Working Data area.
3 Select <--Preprocess > Filter. By default, this selection shows a periodogram of the
input and output spectra (see the etfe reference page).

Note: To display smoothed spectral estimates instead of the periodogram, select

Options > Spectral analysis. This spectral estimate is computed using spa and
your previous settings in the Spectral Model dialog box. To change these settings,
select <--Estimate > Spectral model in the System Identification app, and specify
new model settings.
4 If your data contains multiple input/output channels, in the Channel menu, select
the channel pair you want to view. Although you view only one channel pair at a
time, the filter applies to all input/output channels in this data set.
5 Select the data of interest using one of the following ways:

• Graphically — Draw a rectangle with the mouse on either the input-signal or

the output-signal plot to select the desired frequency interval. Your selection is
displayed on both plots regardless of the plot on which you draw the rectangle.
The Range field is updated to match the selected region. If you need to clear your
selection, right-click the plot.
• Specify the Range — Edit the beginning and the end frequency values.

2-125
2 Data Import and Processing

For example:

8.5 20.0 (rad/s).

Tip To change the frequency units from rad/s to Hz, select Style > Frequency
(Hz). To change the frequency units from Hz to rad/s, select Style > Frequency
(rad/s).
6 In the Range is list, select one of the following:

• Pass band — Allows data in the selected frequency range.

• Stop band — Excludes data in the selected frequency range.
7 Click Filter to preview the filtered results. If you are satisfied, go to step 8.
Otherwise, return to step 5.
8 In the Data name field, enter the name of the data set containing the selected data.
9 Click Insert to save the selection as a new data set and add it to the Data Board.
10 To select another range, repeat steps 5 to 9.

More About
“Filtering Data” on page 2-123

Filtering Frequency-Domain or Frequency-Response Data in the App

For frequency-domain and frequency-response data, filtering is equivalent to selecting
specific data ranges.

To select a range of data in frequency-domain or frequency-response data:

1 Import data into the System Identification app, as described in “Represent Data”.
2 Drag the data set you want you want to filter to the Working Data area.
3 Select <--Preprocess > Select range. This selection displays one of the following
plots:

• Frequency-domain data — Plot shows the absolute of the squares of the input and
output spectra.

2-126
 More About

• Frequency-response data — Top axes show the frequency response magnitude

equivalent to the ratio of the output to the input, and the bottom axes show the
ratio of the input signal to itself, which has the value of 1 at all frequencies.
4 If your data contains multiple input/output channels, in the Channel menu, select
the channel pair you want to view. Although you view only one channel pair at a
time, the filter applies to all input/output channels in this data set.
5 Select the data of interest using one of the following ways:

• Graphically — Draw a rectangle with the mouse on either the input-signal or

the output-signal plot to select the desired frequency interval. Your selection is
displayed on both plots regardless of the plot on which you draw the rectangle.
The Range field is updated to match the selected region.

If you need to clear your selection, right-click the plot.

• Specify the Range — Edit the beginning and the end frequency values.

For example:

8.5 20.0 (rad/s).

Tip If you need to change the frequency units from rad/s to Hz, select Style >
Frequency (Hz). To change the frequency units from Hz to rad/s, select Style >
Frequency (rad/s).
6 In the Range is list, select one of the following:

• Pass band — Allows data in the selected frequency range.

• Stop band — Excludes data in the selected frequency range.
7 In the Data name field, enter the name of the data set containing the selected data.
8 Click Insert. This action saves the selection as a new data set and adds it to the
Data Board.
9 To select another range, repeat steps 5 to 8.

More About
“Filtering Data” on page 2-123

2-127
2 Data Import and Processing

How to Filter Data at the Command Line

In this section...
“Simple Passband Filter” on page 2-128
“Defining a Custom Filter” on page 2-129
“Causal and Noncausal Filters” on page 2-130

Simple Passband Filter

Use idfilt to apply passband and other custom filters to a time-domain or a frequency-
domain iddata object.

In general, you can specify any custom filter. Use this syntax to filter an iddata object
data using the filter called filter:
fdata = idfilt(data,filter)

In the simplest case, you can specify a passband filter for time-domain data using the
following syntax:
fdata = idfilt(data,[wl wh])

In this case, w1 and wh represent the low and high frequencies of the passband,
respectively.

You can specify several passbands, as follows:

filter=[[w1l,w1h];[ w2l,w2h]; ....;[wnl,wnh]]

The filter is an n-by-2 matrix, where each row defines a passband in radians per second.

To define a stopband between ws1 and ws2, use

filter = [0 ws1; ws2 Nyqf]

where, Nyqf is the Nyquist frequency.

For time-domain data, the passband filtering is cascaded Butterworth filters of specified
order. The default filter order is 5. The Butterworth filter is the same as butter in the
Signal Processing Toolbox product. For frequency-domain data, select the indicated
portions of the data to perform passband filtering.

2-128
 More About

More About
“Filtering Data” on page 2-123

Defining a Custom Filter

Use idfilt to apply passband and other custom filters to a time-domain or a frequency-
domain iddata object.

In general, you can specify any custom filter. Use this syntax to filter an iddata object
data using the filter called filter:
fdata = idfilt(data,filter)

You can define a general single-input/single-output (SISO) system for filtering time-
domain or frequency-domain data. For frequency-domain only, you can specify the
(nonparametric) frequency response of the filter.

You use this syntax to filter an iddata object data using a custom filter specified by
filter:
fdata = idfilt(data,filter)

filter can be also any of the following:

filter = idm
filter = {num,den}
filter = {A,B,C,D}

idm is a SISO identified linear model or LTI object. For more information about LTI
objects, see the Control System Toolbox documentation.

{num,den} defines the filter as a transfer function as a cell array of numerator and
denominator filter coefficients.

{A,B,C,D} is a cell array of SISO state-space matrices.

Specifically for frequency-domain data, you specify the frequency response of the filter:
filter = Wf

Here, Wf is a vector of real or complex values that define the filter frequency response,
where the inputs and outputs of data at frequency data.Frequency(kf) are

2-129
2 Data Import and Processing

multiplied by Wf(kf). Wf is a column vector with the length equal to the number of
frequencies in data.

When data contains several experiments, Wf is a cell array with the length equal to the
number of experiments in data.

More About
“Filtering Data” on page 2-123

Causal and Noncausal Filters

For time-domain data, the filtering is causal by default. Causal filters typically introduce
a phase shift in the results. To use a noncausal zero-phase filter (corresponding to
filtfilt in the Signal Processing Toolbox product), specify a third argument in
idfilt:

fdata = idfilt(data,filter,'noncausal')

For frequency-domain data, the signals are multiplied by the frequency response of the
filter. With the filters defined as passband filters, this calculation gives ideal, zero-phase
filtering (“brick wall filters”). Frequencies that have been assigned zero weight by the
filter (outside the passband or via frequency response) are removed.

When you apply idfilt to an idfrd data object, the data is first converted to a
frequency-domain iddata object (see “Transforming Between Frequency-Domain and
Frequency-Response Data” on page 2-143). The result is an iddata object.

More About
“Filtering Data” on page 2-123

2-130
 Generate Data Using Simulation

Generate Data Using Simulation

In this section...
“Commands for Generating Data Using Simulation” on page 2-131
“Create Periodic Input Data” on page 2-132
“Generate Output Data Using Simulation” on page 2-134
“Simulating Data Using Other MathWorks Products” on page 2-136

Commands for Generating Data Using Simulation

You can generate input data and then use it with a model to create output data.

Simulating output data requires that you have a model with known coefficients.
For more information about commands for constructing models, see “Commands for
Constructing Model Structures” on page 1-22.

To generate input data, use idinput to construct a signal with the desired
characteristics, such as a random Gaussian or binary signal or a sinusoid. idinput
returns a matrix of input values.

The following table lists the commands you can use to simulate output data. For more
information about these commands, see the corresponding reference pages.

Commands for Generating Data

Command Description Example

idinput Constructs a signal with u = iddata([],...
the desired characteristics, idinput(400,'rbs',[0 0.3]));
such as a random
Gaussian or binary signal
or a sinusoid, and returns
a matrix of input values.
sim Simulates response data To simulate the model output y for a given
based on existing linear input, use the following command:
or nonlinear parametric
model in the MATLAB y = sim(m,data)
workspace.

2-131
2 Data Import and Processing

Command Description Example

m is the model object name, and data is input
data matrix or iddata object.

Create Periodic Input Data

This example shows how to create a periodic random Gaussian input signal using
idinput.

Create a periodic input for one input and consisting of five periods, where each period is
300 samples.

per_u = idinput([300 1 5]);

Create an iddata object using the periodic input and leaving the output empty.

u = iddata([],per_u,'Period',.300);

View the data characteristics in time- and frequency-domain.

% Plot data in time-domain.

plot(u);
% Plot the spectrum.
spectrum(spa(u));

2-132
 Generate Data Using Simulation

(Optional) Simulate model output using the data.

% Construct a polynomial model.

m0 =idpoly([1 -1.5 0.7],[0 1 0.5]);
% Simulate model output with Gaussian noise.
sim_opt = simOptions('AddNoise',true);
sim(m0,u,sim_opt);

2-133
2 Data Import and Processing

Generate Output Data Using Simulation

This example shows how to generate output data by simulating a model using an input
signal created using idinput.

You use the generated data to estimate a model of the same order as the model used to
generate the data. Then, you check how closely both models match to understand the
effects of input data characteristics and noise on the estimation.

Create an ARMAX model with known coefficients.

A = [1 -1.2 0.7];
B = {[0 1 0.5 0.1],[0 1.5 -0.5],[0 -0.1 0.5 -0.1]};

2-134
 Generate Data Using Simulation

C = [1 0 0 0 0];
Ts = 1;
m0 = idpoly(A,B,C,'Ts',1);

The leading zeros in the B matrix indicate the input delay (nk), which is 1 for each input
channel.

Construct a pseudorandom binary input data.

u = idinput([255,3],'prbs');

Simulate model output with noise using the input data.

y = sim(m0,u,simOptions('AddNoise',true));

Represent the simulation data as an iddata object.

iodata = iddata(y,u,m0.Ts);

(Optional) Estimate a model of the same order as m0 using iodata.

na = 2;
nb = [3 2 3];
nc = 4;
nk = [1 1 1];
me = armax(iodata,[na,nb,nc,nk]);

Use bode(m0,me) and compare(iodata,me) to check how closely me and m0 match.

compare(iodata,me);

2-135
2 Data Import and Processing

Simulating Data Using Other MathWorks Products

You can also simulate data using the Simulink and Signal Processing Toolbox software.
Data simulated outside the System Identification Toolbox product must be in the
MATLAB workspace as double matrices. For more information about simulating models
using the Simulink software, see “Simulating Identified Model Output in Simulink”.

2-136
 Transforming Between Time- and Frequency-Domain Data

Transforming Between Time- and Frequency-Domain Data

In this section...
“Transforming Data Domain in the App” on page 2-137
“Transforming Data Domain at the Command Line” on page 2-141

Transforming Data Domain in the App

• “Transforming Time-Domain Data” on page 2-137
• “Transforming Frequency-Domain Data” on page 2-138
• “Transforming Frequency-Response Data” on page 2-139
• “See Also” on page 2-141

Transforming Time-Domain Data

In the System Identification app, time-domain data has an icon with a white background.
You can transform time-domain data to frequency-domain or frequency-response data.
The frequency values of the resulting frequency vector range from 0 to the Nyquist
frequency fS = p Ts , where Ts is the sampling interval.

Transforming from time-domain to frequency-response data is equivalent to estimating a

model from the data using the spafdr method.

1 In the System Identification app, drag the icon of the data you want to transform to
the Working Data rectangle.
2 In the Operations area, select <--Preprocess > Transform data in the drop-down
menu to open the Transform Data dialog box.
3 In the Transform to list, select one of the following:

• Frequency Function — Create a new idfrd object using the spafdr method.
Go to step 4.

2-137
2 Data Import and Processing

• Frequency Domain Data — Create a new iddata object using the fft method.
Go to step 6.
4 In the Frequency Spacing list, select the spacing of the frequencies at which the
frequency function is estimated:

• linear — Uniform spacing of frequency values between the endpoints.

• logarithmic — Base-10 logarithmic spacing of frequency values between the
endpoints.
5 In the Number of Frequencies field, enter the number of frequency values.
6 In the Name of new data field, type the name of the new data set. This name must
be unique in the Data Board.
7 Click Transform to add the new data set to the Data Board in the System
Identification app.
8 Click Close to close the Transform Data dialog box.

Transforming Frequency-Domain Data

In the System Identification app, frequency-domain data has an icon with a green
background. You can transform frequency-domain data to time-domain or frequency-
response (frequency-function) data.

Transforming from time-domain or frequency-domain data to frequency-response data is

equivalent to estimating a nonparametric model of the data using the spafdr method.

2-138
 Transforming Between Time- and Frequency-Domain Data

1 In the System Identification app, drag the icon of the data you want to transform to
the Working Data rectangle.
2 Select <--Preprocess > Transform data.
3 In the Transform to list, select one of the following:

• Frequency Function — Create a new idfrd object using the spafdr method.
Go to step 4.
• Time Domain Data — Create a new iddata object using the ifft (inverse fast
Fourier transform) method. Go to step 6.
4 In the Frequency Spacing list, select the spacing of the frequencies at which the
frequency function is estimated:

• linear — Uniform spacing of frequency values between the endpoints.

• logarithmic — Base-10 logarithmic spacing of frequency values between the
endpoints.
5 In the Number of Frequencies field, enter the number of frequency values.
6 In the Name of new data field, type the name of the new data set. This name must
be unique in the Data Board.
7 Click Transform to add the new data set to the Data Board in the System
Identification app.
8 Click Close to close the Transform Data dialog box.

Transforming Frequency-Response Data

In the System Identification app, frequency-response data has an icon with a yellow
background. You can transform frequency-response data to frequency-domain data
(iddata object) or to frequency-response data with a different frequency resolution.

When you select to transform single-input/single-output (SISO) frequency-response

data to frequency-domain data, the toolbox creates outputs that equal the frequency
responses, and inputs equal to 1. Therefore, the ratio between the Fourier transform
of the output and the Fourier transform of the input is equal to the system frequency
response.

For the multiple-input case, the toolbox transforms the frequency-response data to
frequency-domain data as if each input contributes independently to the entire output
of the system and then combines information. For example, if a system has three inputs,
u1, u2, and u3 and two frequency samples, the input matrix is set to:

2-139
2 Data Import and Processing

È1 0 0˘
Í ˙
Í1 0 0˙
Í0 1 0˙
Í ˙
Í0 1 0˙
Í0 0 1˙
Í ˙
ÍÎ0 0 1 ˙˚

In general, for nu inputs and ns samples (the number of frequencies), the input matrix
has nu columns and (ns ◊ nu) rows.

Note: To create a separate experiment for the response from each input, see
“Transforming Between Frequency-Domain and Frequency-Response Data” on page
2-143.

When you transform frequency-response data by changing its frequency resolution, you
can modify the number of frequency values by changing between linear or logarithmic
spacing. You might specify variable frequency spacing to increase the number of data
points near the system resonance frequencies, and also make the frequency vector
coarser in the region outside the system dynamics. Typically, high-frequency noise
dominates away from frequencies where interesting system dynamics occur. The System
Identification app lets you specify logarithmic frequency spacing, which results in a
variable frequency resolution.

Note: The spafdr command lets you lets you specify any variable frequency resolution.

1 In the System Identification app, drag the icon of the data you want to transform to
the Working Data rectangle.
2 Select <--Preprocess > Transform data.
3 In the Transform to list, select one of the following:

• Frequency Domain Data — Create a new iddata object. Go to step 6.

• Frequency Function — Create a new idfrd object with different resolution
(number and spacing of frequencies) using the spafdr method. Go to step 4.

2-140
 Transforming Between Time- and Frequency-Domain Data

4 In the Frequency Spacing list, select the spacing of the frequencies at which the
frequency function is estimated:

• linear — Uniform spacing of frequency values between the endpoints.

• logarithmic — Base-10 logarithmic spacing of frequency values between the
endpoints.
5 In the Number of Frequencies field, enter the number of frequency values.
6 In the Name of new data field, type the name of the new data set. This name must
be unique in the Data Board.
7 Click Transform to add the new data set to the Data Board in the System
Identification app.
8 Click Close to close the Transform Data dialog box.

See Also

For a description of time-domain, frequency-domain, and frequency-response data, see

“Representing Data in MATLAB Workspace” on page 2-9.

To learn how to transform data at the command line instead of the app, see
“Transforming Data Domain at the Command Line” on page 2-141.

Transforming Data Domain at the Command Line

• “Supported Data Transformations” on page 2-141
• “Transforming Between Time and Frequency Domain” on page 2-142
• “Transforming Between Frequency-Domain and Frequency-Response Data” on page
2-143
• “See Also” on page 2-144

Supported Data Transformations

The following table shows the different ways you can transform data from one data
domain to another. If the transformation is supported for a given row and column
combination in the table, the method used by the software is listed in the cell at their
intersection.

2-141
2 Data Import and Processing

Original Data Format To Time Domain To Frequency Domain To Frequency Function

(iddata object) (iddata object) (idfrd object)
Time Domain N/A. Yes, using fft. Yes, using etfe,
(iddata object) spa, or spafdr.
Frequency Yes, using ifft. N/A. Yes, using etfe,
Domain spa, or spafdr.
(iddata object)
Frequency No. Yes. Calculation Yes. Calculates a
Function creates frequency- frequency function
(idfrd object) domain iddata with different
object that has the resolution (number
same ratio between and spacing of
output and input as frequencies) using
the original idfrd spafdr.
object's response
data.

Transforming Between Time and Frequency Domain

The iddata object stores time-domain or frequency-domain data. The following table
summarizes the commands for transforming data between time and frequency domains.

Command Description Syntax Example

fft Transforms time-domain data To transform time-domain
to the frequency domain. iddata object t_data to
frequency-domain iddata
You can specify N, the number object f_data with N frequency
of frequency values. points, use:
f_data =
fft(t_data,N)
ifft Transforms frequency-domain To transform frequency-
data to the time domain. domainiddata object f_data
Frequencies are linear and to time-domain iddata object
equally spaced. t_data, use:

t_data =
ifft(f_data)

2-142
 Transforming Between Time- and Frequency-Domain Data

Transforming Between Frequency-Domain and Frequency-Response Data

You can transform frequency-response data to frequency-domain data (iddata object).

The idfrd object represents complex frequency-response of the system at different
frequencies. For a description of this type of data, see “Frequency-Response Data
Representation” on page 2-13.

When you select to transform single-input/single-output (SISO) frequency-response

data to frequency-domain data, the toolbox creates outputs that equal the frequency
responses, and inputs equal to 1. Therefore, the ratio between the Fourier transform
of the output and the Fourier transform of the input is equal to the system frequency
response.

For information about changing the frequency resolution of frequency-response data to

a new constant or variable (frequency-dependent) resolution, see the spafdr reference
page. You might use this feature to increase the number of data points near the system
resonance frequencies and make the frequency vector coarser in the region outside the
system dynamics. Typically, high-frequency noise dominates away from frequencies
where interesting system dynamics occur.

Note: You cannot transform an idfrd object to a time-domain iddata object.

To transform an idfrd object with the name idfrdobj to a frequency-domain iddata

object, use the following syntax:

dataf = iddata(idfrdobj)

The resulting frequency-domain iddata object contains values at the same frequencies
as the original idfrd object.

For the multiple-input case, the toolbox represents frequency-response data as if each
input contributes independently to the entire output of the system and then combines
information. For example, if a system has three inputs, u1, u2, and u3 and two frequency
samples, the input matrix is set to:

2-143
2 Data Import and Processing

È1 0 0˘
Í ˙
Í1 0 0˙
Í0 1 0˙
Í ˙
Í0 1 0˙
Í0 0 1˙
Í ˙
ÍÎ0 0 1 ˙˚

In general, for nu inputs and ns samples, the input matrix has nu columns and (ns ◊ nu)
rows.

If you have ny outputs, the transformation operation produces an output matrix has ny
columns and (ns ◊ nu) rows using the values in the complex frequency response G(iw)
matrix (ny-by-nu-by-ns). In this example, y1 is determined by unfolding G(1,1,:),
G(1,2,:), and G(1,3,:) into three column vectors and vertically concatenating
these vectors into a single column. Similarly, y2 is determined by unfolding G(2,1,:),
G(2,2,:), and G(2,3,:) into three column vectors and vertically concatenating these
vectors.

If you are working with multiple inputs, you also have the option of storing the
contribution by each input as an independent experiment in a multiexperiment data set.
To transform an idfrd object with the name idfrdobj to a multiexperiment data set
datf, where each experiment corresponds to each of the inputs in idfrdobj

datf = iddata(idfrdobj,'me')

In this example, the additional argument 'me' specifies that multiple experiments are
created.

By default, transformation from frequency-response to frequency-domain data strips

away frequencies where the response is inf or NaN. To preserve the entire frequency
vector, use datf = iddata(idfrdobj,'inf'). For more information, type help
idfrd/iddata.

See Also

Transforming from time-domain or frequency-domain data to frequency-response data is

equivalent to creating a frequency-response model from the data. For more information,
see “Identifying Frequency-Response Models” on page 3-8.

2-144
 Manipulating Complex-Valued Data

Manipulating Complex-Valued Data

In this section...
“Supported Operations for Complex Data” on page 2-145
“Processing Complex iddata Signals at the Command Line” on page 2-145

Supported Operations for Complex Data

System Identification Toolbox estimation algorithms support complex data. For example,
the following estimation commands estimate complex models from complex data: ar,
armax, arx, bj, ivar, iv4, oe, pem, spa, tfest, ssest, and n4sid.

Model transformation routines, such as freqresp and zpkdata, work for complex-
valued models. However, they do not provide pole-zero confidence regions. For complex
models, the parameter variance-covariance information refers to the complex-valued
parameters and the accuracy of the real and imaginary is not computed separately.

The display commands compare and plot also work with complex-valued data
and models. To plot the real and imaginary parts of the data separately, use
plot(real(data)) and plot(imag(data)), respectively.

Processing Complex iddata Signals at the Command Line

If the iddata object data contains complex values, you can use the following commands
to process the complex data and create a new iddata object.

Command Description
abs(data) Absolute value of complex signals in iddata object.
angle(data) Phase angle (in radians) of each complex signals in iddata
object.
complex(data) For time-domain data, this command makes the iddata object
complex—even when the imaginary parts are zero. For frequency-
domain data that only stores the values for nonnegative
frequencies, such that realdata(data)=1, it adds signal values
for negative frequencies using complex conjugation.
imag(data) Selects the imaginary parts of each signal in iddata object.

2-145
2 Data Import and Processing

Command Description
isreal(data) 1 when data (time-domain or frequency-domain) contains only
real input and output signals, and returns 0 when data (time-
domain or frequency-domain) contains complex signals.
real(data) Real part of complex signals in iddata object.
realdata(data) Returns a value of 1 when data is a real-valued, time-domain
signal, and returns 0 otherwise.

For example, suppose that you create a frequency-domain iddata object Datf by
applying fft to a real-valued time-domain signal to take the Fourier transform of the
signal. The following is true for Datf:

isreal(Datf) = 0
realdata(Datf) = 1

2-146
 3

Linear Model Identification

• “Black-Box Modeling” on page 3-3

• “Identifying Frequency-Response Models” on page 3-8
• “Identifying Impulse-Response Models” on page 3-16
• “Identifying Process Models” on page 3-24
• “Identifying Input-Output Polynomial Models” on page 3-40
• “Identifying Transfer Function Models” on page 3-70
• “Refining Linear Parametric Models” on page 3-86
• “Refine ARMAX Model with Initial Parameter Guesses at Command Line” on page
3-89
• “Refine Initial ARMAX Model at Command Line” on page 3-91
• “Extracting Numerical Model Data” on page 3-93
• “Transforming Between Discrete-Time and Continuous-Time Representations” on
page 3-96
• “Continuous-Discrete Conversion Methods” on page 3-100
• “Effect of Input Intersample Behavior on Continuous-Time Models” on page 3-110
• “Transforming Between Linear Model Representations” on page 3-114
• “Subreferencing Models” on page 3-116
• “Concatenating Models” on page 3-120
• “Merging Models” on page 3-124
• “Building and Estimating Process Models Using System Identification Toolbox™” on
page 3-125
• “Determining Model Order and Delay” on page 3-151
• “Model Structure Selection: Determining Model Order and Input Delay” on page
3-152
• “Frequency Domain Identification: Estimating Models Using Frequency Domain
Data” on page 3-167
3 Linear Model Identification

• “Building Structured and User-Defined Models Using System Identification

Toolbox™” on page 3-193

3-2
 Black-Box Modeling

Black-Box Modeling
In this section...
“Selecting Black-Box Model Structure and Order” on page 3-3
“When to Use Nonlinear Model Structures?” on page 3-4
“Black-Box Estimation Example” on page 3-5

Selecting Black-Box Model Structure and Order

Black-box modeling is useful when your primary interest is in fitting the data regardless
of a particular mathematical structure of the model. The toolbox provides several linear
and nonlinear black-box model structures, which have traditionally been useful for
representing dynamic systems. These model structures vary in complexity depending on
the flexibility you need to account for the dynamics and noise in your system. You can
choose one of these structures and compute its parameters to fit the measured response
data.

Black-box modeling is usually a trial-and-error process, where you estimate the

parameters of various structures and compare the results. Typically, you start with the
simple linear model structure and progress to more complex structures. You might also
choose a model structure because you are more familiar with this structure or because
you have specific application needs.

The simplest linear black-box structures require the fewest options to configure:

• Transfer function, with a given number of poles and zeros.

• “Linear ARX model”, which is the simplest input-output polynomial model.
• “State-space model”, which you can estimate by specifying the number of model states

Estimation of some of these structures also uses noniterative estimation algorithms,

which further reduces complexity.

You can configure a model structure using the model order. The definition of model order
varies depending on the type of model you select. For example, if you choose a transfer
function representation, the model order is related to the number of poles and zeros. For
state-space representation, the model order corresponds to the number of states. In some
cases, such as for linear ARX and state-space model structures, you can estimate the
model order from the data.

3-3
3 Linear Model Identification

If the simple model structures do not produce good models, you can select more complex
model structures by:

• Specifying a higher model order for the same linear model structure. Higher model
order increases the model flexibility for capturing complex phenomena. However,
unnecessarily high orders can make the model less reliable.
• Explicitly modeling the noise:

y(t)=Gu(t)+He(t)

where H models the additive disturbance by treating the disturbance as the output of
a linear system driven by a white noise source e(t).

Using a model structure that explicitly models the additive disturbance can help
to improve the accuracy of the measured component G. Furthermore, such a model
structure is useful when your main interest is using the model for predicting future
response values.
• Using a different linear model structure.

See “Linear Model Structures”.

• Using a nonlinear model structure.

Nonlinear models have more flexibility in capturing complex phenomena than linear
models of similar orders. See “Available Nonlinear Models”.

Ultimately, you choose the simplest model structure that provides the best fit to your
measured data. For more information, see “Estimating Linear Models Using Quick
Start”.

Regardless of the structure you choose for estimation, you can simplify the model for your
application needs. For example, you can separate out the measured dynamics (G) from
the noise dynamics (H) to obtain a simpler model that represents just the relationship
between y and u. You can also linearize a nonlinear model about an operating point.

When to Use Nonlinear Model Structures?

A linear model is often sufficient to accurately describe the system dynamics and, in
most cases, you should first try to fit linear models. If the linear model output does not
adequately reproduce the measured output, you might need to use a nonlinear model.

3-4
 Black-Box Modeling

You can assess the need to use a nonlinear model structure by plotting the response of
the system to an input. If you notice that the responses differ depending on the input
level or input sign, try using a nonlinear model. For example, if the output response to
an input step up is faster than the response to a step down, you might need a nonlinear
model.

Before building a nonlinear model of a system that you know is nonlinear, try
transforming the input and output variables such that the relationship between the
transformed variables is linear. For example, consider a system that has current and
voltage as inputs to an immersion heater, and the temperature of the heated liquid as
an output. The output depends on the inputs via the power of the heater, which is equal
to the product of current and voltage. Instead of building a nonlinear model for this two-
input and one-output system, you can create a new input variable by taking the product
of current and voltage and then build a linear model that describes the relationship
between power and temperature.

If you cannot determine variable transformations that yield a linear relationship between
input and output variables, you can use nonlinear structures such as Nonlinear ARX or
Hammerstein-Wiener models. For a list of supported nonlinear model structures and
when to use them, see “Available Nonlinear Models”.

Black-Box Estimation Example

You can use the System Identification app or commands to estimate linear and nonlinear
models of various structures. In most cases, you choose a model structure and estimate
the model parameters using a single command.

Consider the mass-spring-damper system, described in “About Dynamic Systems

and Models”. If you do not know the equation of motion of this system, you can use a
black-box modeling approach to build a model. For example, you can estimate transfer
functions or state-space models by specifying the orders of these model structures.

A transfer function is a ratio of polynomials:

G ( s) =
(b0 + b1 s + b2 s2 + ...)
(1 + f1s + f2s2 + ...)
For the mass-spring damper system, this transfer function is:

3-5
3 Linear Model Identification

1
G ( s) =
( ms 2
+ cs + k )

which is a system with no zeros and 2 poles.

In discrete-time, the transfer function of the mass-spring-damper system can be:

bz-1
G ( z -1 ) =
(1 + f1 z-1 + f2 z-2 )
where the model orders correspond to the number of coefficients of the numerator and
the denominator (nb = 1 and nf = 2) and the input-output delay equals the lowest order
exponent of z–1 in the numerator (nk = 1).

In continuous-time, you can build a linear transfer function model using the tfest
command:
m = tfest(data, 2, 0)

where data is your measured input-output data, represented as an iddata object and
the model order is the set of number of poles (2) and the number of zeros (0).

Similarly, you can build a discrete-time model Output Error structure using the following
command:
m = oe(data, [1 2 1])

The model order is [nb nf nk] = [1 2 1]. Usually, you do not know the model orders
in advance. You should try several model order values until you find the orders that
produce an acceptable model.

Alternatively, you can choose a state-space structure to represent the mass-spring-

damper system and estimate the model parameters using the ssest or the n4sid
command:
m = ssest(data, 2)

where order = 2 represents the number of states in the model.

In black-box modeling, you do not need the system’s equation of motion—only a guess of
the model orders.

3-6
 Black-Box Modeling

For more information about building models, see “Steps for Using the System
Identification App” and “Model Estimation Commands”.

3-7
3 Linear Model Identification

Identifying Frequency-Response Models

In this section...
“What Is a Frequency-Response Model?” on page 3-8
“Data Supported by Frequency-Response Models” on page 3-8
“Estimate Frequency-Response Models in the App” on page 3-9
“Estimate Frequency-Response Models at the Command Line” on page 3-10
“Selecting the Method for Computing Spectral Models” on page 3-11
“Controlling Frequency Resolution of Spectral Models ” on page 3-12
“Spectrum Normalization” on page 3-13

What Is a Frequency-Response Model?

You can estimate frequency-response models and visualize the responses on a Bode plot,
which shows the amplitude change and the phase shift as a function of the sinusoid
frequency.

The frequency-response function describes the steady-state response of a system to

sinusoidal inputs. For a linear system, a sinusoidal input of a specific frequency results
in an output that is also a sinusoid with the same frequency, but with a different
amplitude and phase. The frequency-response function describes the amplitude change
and phase shift as a function of frequency.

For a discrete-time system sampled with a time interval T, the frequency-response model
G(z) relates the Z-transforms of the input U(z) and output Y(z):

Y ( z) = G ( z)U ( z)

In other words, the frequency-response function, G(eiwT), is the Laplace transform of

the impulse response that is evaluated on the imaginary axis. The frequency-response
function is the transfer function G(z) evaluated on the unit circle.

The estimation result is an idfrd model, which stores the estimated frequency response
and its covariance.

Data Supported by Frequency-Response Models

You can estimate spectral analysis models from data with the following characteristics:

3-8
 Identifying Frequency-Response Models

• Complex or real data.

• Time- or frequency-domain iddata or idfrd data object. To learn more about
estimating time-series models, see “Time-Series Model Identification”.
• Single- or multiple-output data.

Estimate Frequency-Response Models in the App

You must have already imported your data into the app and performed any necessary
preprocessing operations. For more information, see “Data Preparation”.

To estimate frequency-response models in the System Identification app:

1 In the System Identification app, select Estimate > Spectral models to open the
Spectral Model dialog box.
2 In the Method list, select the spectral analysis method you want to use. For
information about each method, see “Selecting the Method for Computing Spectral
Models” on page 3-11.
3 Specify the frequencies at which to compute the spectral model in one of the
following ways:

• In the Frequencies field, enter either a vector of values, a MATLAB expression

that evaluates to a vector, or a variable name of a vector in the MATLAB
workspace. For example, logspace(-1,2,500).
• Use the combination of Frequency Spacing and Frequencies to construct the
frequency vector of values:

• In the Frequency Spacing list, select Linear or Logarithmic frequency

spacing.

Note: For etfe, only the Linear option is available.

• In the Frequencies field, enter the number of frequency points.

For time-domain data, the frequency ranges from 0 to the Nyquist frequency. For
frequency-domain data, the frequency ranges from the smallest to the largest
frequency in the data set.
4 In the Frequency Resolution field, enter the frequency resolution, as described in
“Controlling Frequency Resolution of Spectral Models ” on page 3-12. To use the
default value, enter default or, equivalently, the empty matrix [].

3-9
3 Linear Model Identification

5 In the Model Name field, enter the name of the correlation analysis model. The
model name should be unique in the Model Board.
6 Click Estimate to add this model to the Model Board in the System Identification
app.
7 In the Spectral Model dialog box, click Close.
8 To view the frequency-response plot, select the Frequency resp check box in the
System Identification app. For more information about working with this plot, see
“Frequency Response Plots” on page 10-38.
9 To view the estimated disturbance spectrum, select the Noise spectrum check box
in the System Identification app. For more information about working with this plot,
see “Noise Spectrum Plots” on page 10-46.
10 Validate the model after estimating it. For more information, see “Model Validation”.

To export the model to the MATLAB workspace, drag it to the To Workspace rectangle
in the System Identification app. You can retrieve the responses from the resulting
idfrd model object using the bode or nyquist command.

Estimate Frequency-Response Models at the Command Line

You can use the etfe, spa, and spafdr commands to estimate spectral models. The
following table provides a brief description of each command and usage examples.

The resulting models are stored as idfrd model objects. For detailed information about
the commands and their arguments, see the corresponding reference page.

Commands for Frequency Response

Command Description Usage

etfe Estimates an empirical To estimate a model m, use the following syntax:
transfer function using
Fourier analysis. m=etfe(data)

spa Estimates a frequency To estimate a model m, use the following syntax:

response with a fixed
frequency resolution m=spa(data)
using spectral analysis.
spafdr Estimates a frequency To estimate a model m, use the following syntax:
response with a variable
m=spafdr(data,R,w)

3-10
 Identifying Frequency-Response Models

Command Description Usage

frequency resolution where R is the resolution vector and w is the frequency
using spectral analysis. vector.

Validate the model after estimating it. For more information, see “Model Validation”.

Selecting the Method for Computing Spectral Models

This section describes how to select the method for computing spectral models in the
estimation procedures “Estimate Frequency-Response Models in the App” on page
3-9 and “Estimate Frequency-Response Models at the Command Line” on page
3-10.

You can choose from the following three spectral-analysis methods:

• etfe (Empirical Transfer Function Estimate)

For input-output data. This method computes the ratio of the Fourier transform of
the output to the Fourier transform of the input.

For time-series data. This method computes a periodogram as the normalized

absolute squares of the Fourier transform of the time series.

ETFE works well for highly resonant systems or narrowband systems. The drawback
of this method is that it requires linearly spaced frequency values, does not estimate
the disturbance spectrum, and does not provide confidence intervals. ETFE also
works well for periodic inputs and computes exact estimates at multiples of the
fundamental frequency of the input and their ratio.
• spa (SPectral Analysis)

This method is the Blackman-Tukey spectral analysis method, where windowed

versions of the covariance functions are Fourier transformed.
• spafdr (SPectral Analysis with Frequency Dependent Resolution)

This method is a variant of the Blackman-Tukey spectral analysis method with

frequency-dependent resolution. First, the algorithm computes Fourier transforms
of the inputs and outputs. Next, the products of the transformed inputs and outputs
with the conjugate input transform are smoothed over local frequency regions. The
widths of the local frequency regions can vary as a function of frequency. The ratio of
these averages computes the frequency-response estimate.

3-11
3 Linear Model Identification

Controlling Frequency Resolution of Spectral Models

• “What Is Frequency Resolution?” on page 3-12
• “Frequency Resolution for etfe and spa” on page 3-12
• “Frequency Resolution for spafdr” on page 3-13
• “etfe Frequency Resolution for Periodic Input” on page 3-13

This section supports the estimation procedures “Estimate Frequency-Response Models

in the App” on page 3-9 and “Estimate Frequency-Response Models at the Command
Line” on page 3-10.

What Is Frequency Resolution?

Frequency resolution is the size of the smallest frequency for which details in the
frequency response and the spectrum can be resolved by the estimate. A resolution of 0.1
rad/s means that the frequency response variations at frequency intervals at or below 0.1
rad/s are not resolved.

Note: Finer resolution results in greater uncertainty in the model estimate.

Specifying the frequency resolution for etfe and spa is different than for spafdr.

Frequency Resolution for etfe and spa

For etfe and spa, the frequency resolution is approximately equal to the following
value:

2p Ê radians ˆ
Á ˜
M Ë sampling interval ¯

M is a scalar integer that sets the size of the lag window. The value of M controls the
trade-off between bias and variance in the spectral estimate.

The default value of M for spa is good for systems without sharp resonances. For etfe,
the default value of M gives the maximum resolution.

A large value of M gives good resolution, but results in more uncertain estimates. If a
true frequency function has sharp peak, you should specify higher M values.

3-12
 Identifying Frequency-Response Models

Frequency Resolution for spafdr

In case of etfe and spa, the frequency response is defined over a uniform frequency
range, 0-Fs/2 radians per second, where Fs is the sampling frequency—equal to twice
the Nyquist frequency. In contrast, spafdr lets you increase the resolution in a specific
frequency range, such as near a resonance frequency. Conversely, you can make the
frequency grid coarser in the region where the noise dominates—at higher frequencies,
for example. Such customizing of the frequency grid assists in the estimation process by
achieving high fidelity in the frequency range of interest.

For spafdr, the frequency resolution around the frequency k is the value R(k). You can
enter R(k) in any one of the following ways:

• Scalar value of the constant frequency resolution value in radians per second.

Note: The scalar R is inversely related to the M value used for etfe and spa.
• Vector of frequency values the same size as the frequency vector.
• Expression using MATLAB workspace variables and evaluates to a resolution vector
that is the same size as the frequency vector.

The default value of the resolution for spafdr is twice the difference between
neighboring frequencies in the frequency vector.

etfe Frequency Resolution for Periodic Input

If the input data is marked as periodic and contains an integer number of periods
(data.Period is an integer), etfe computes the frequency response at frequencies
2 pk
(k
T Period ) where k = 1, 2,..., Period
.

For periodic data, the frequency resolution is ignored.

Spectrum Normalization
The spectrum of a signal is the square of the Fourier transform of the signal. The spectral
estimate using the commands spa, spafdr, and etfe is normalized by the sampling
interval T:
M
F y (w) = T Â Ry ( kT ) e-iwT WM ( k)
k= - M

3-13
3 Linear Model Identification

where WM(k) is the lag window, and M is the width of the lag window. The output
covariance Ry(kT) is given by the following discrete representation:

N
ˆ ( kT ) = 1
R y Â
N l =1
y( lT - kT) y(lT )

Because there is no scaling in a discrete Fourier transform of a vector, the purpose of T is

to relate the discrete transform of a vector to the physically meaningful transform of the
measured signal. This normalization sets the units of F y (w) as power per radians per
unit time, and makes the frequency units radians per unit time.

The scaling factor of T is necessary to preserve the energy density of the spectrum after
interpolation or decimation.

By Parseval's theorem, the average energy of the signal must equal the average energy in
the estimated spectrum, as follows:

1 p/T
Ey2 (t) = Ú
2 p - p/ T
F y (w) dw

S1 ∫ Ey2 (t)
1 p/T
S2 ∫ Ú F ( w) dw
2p -p / T y

To compare the left side of the equation (S1) to the right side (S2), enter the following
commands. In this code, phiy contains F y (w) between w = 0 and w = p T with the
frequency step given as follows:

Ê p ˆ
Á ˜
Ë T ◊ length(phiy) ¯

load iddata1
% Create a time-series iddata object.
y = z1(:,1,[]);
% Define sample interval from the data.
T = y.Ts;
% Estimate the frequency response.
sp = spa(y);

3-14
 Identifying Frequency-Response Models

% Remove spurious dimensions

phiy = squeeze(sp.spec);
% Compute average energy from the estimated energy spectrum, where S1 is
% scaled by T.
S1 = sum(phiy)/length(phiy)/T
% Compute average energy of the signal.
S2 = sum(y.y.^2)/size(y,1)

S1 =

19.2076

S2 =

19.4646

Thus, the average energy of the signal approximately equals the average energy in the
estimated spectrum.

3-15
3 Linear Model Identification

Identifying Impulse-Response Models

In this section...
“What Is Time-Domain Correlation Analysis?” on page 3-16
“Data Supported by Correlation Analysis” on page 3-16
“Estimate Impulse-Response Models Using System Identification App” on page 3-17
“Estimate Impulse-Response Models at the Command Line” on page 3-18
“Compute Response Values” on page 3-19
“Identify Delay Using Transient-Response Plots” on page 3-19
“Correlation Analysis Algorithm” on page 3-21

What Is Time-Domain Correlation Analysis?

Time-domain correlation analysis refers to non-parametric estimation of the impulse
response of dynamic systems as a finite impulse response (FIR) model from the data.
Correlation analysis assumes a linear system and does not require a specific model
structure.

Impulse response is the output signal that results when the input is an impulse and has
the following definition for a discrete model:

u(t) = 0 t>0
u(t) = 1 t=0

The response to an input u(t) is equal to the convolution of the impulse response, as
follows:

t
y( t) = Ú0 h ( t - z ) ◊ u(z)dz

Data Supported by Correlation Analysis

You can estimate impulse-response models from data with the following characteristics:

• Real or complex data.

• Single- or multiple-output data.

3-16
 Identifying Impulse-Response Models

• Time- or frequency-domain data with nonzero sampling time.

Time-domain data must be regularly sampled. You cannot use time-series data for
correlation analysis.

Estimate Impulse-Response Models Using System Identification App

Before you can perform this task, you must have:

• Imported data into the System Identification app. See “Import Time-Domain Data
into the App” on page 2-16. For supported data formats, see “Data Supported by
Correlation Analysis”.
• Performed any required data preprocessing operations. To improve the accuracy of
your model, you should detrend your data. See “Ways to Prepare Data for System
Identification” on page 2-6.

To estimate in the System Identification app using time-domain correlation analysis:

1 In the System Identification app, select Estimate > Correlation models to open
the Correlation Model dialog box.
2 In the Time span (s) field, specify a scalar value as the time interval over which
the impulse or step response is calculated. For a scalar time span T, the resulting
response is plotted from -T/4 to T.

Tip You can also enter a 2-D vector in the format [min_value max_value].
3 In the Order of whitening filter field, specify the filter order.

The prewhitening filter is determined by modeling the input as an autoregressive

process of order N. The algorithm applies a filter of the form A(q)u(t)=u_F(t). That
is, the input u(t) is subjected to an FIR filter A to produce the filtered signal u_F(t).
Prewhitening the input by applying a whitening filter before estimation might
improve the quality of the estimated impulse response g.

The order of the prewhitening filter, N, is the order of the A filter. N equals the
number of lags. The default value of N is 10, which you can also specify as [].
4 In the Model Name field, enter the name of the correlation analysis model. The
name of the model should be unique in the Model Board.
5 Click Estimate to add this model to the Model Board in the System Identification
app.

3-17
3 Linear Model Identification

6 In the Correlation Model dialog box, click Close.

Next Steps

• Export the model to the MATLAB workspace for further analysis by dragging it to the
To Workspace rectangle in the System Identification app.
• View the transient response plot by selecting the Transient resp check box in the
System Identification app. For more information about working with this plot and
selecting to view impulse- versus step-response, see “Impulse and Step Response
Plots” on page 10-31.

Estimate Impulse-Response Models at the Command Line

Before you can perform this task, you must have:

• Input/output or frequency-response data. See “Representing Time- and Frequency-

Domain Data Using iddata Objects” on page 2-50. For supported data formats, see
“Data Supported by Correlation Analysis”.
• Performed any required data preprocessing operations. If you use time-domain
data, you can detrend it before estimation. See “Ways to Prepare Data for System
Identification” on page 2-6.

Use impulseest to compute impulse response models. impulseest estimates a high-

order, noncausal FIR model using correlation analysis. The resulting models are stored
as idtf model objects and contain impulse-response coefficients in the model numerator.

To estimate the model m and plot the impulse or step response, use the following syntax:
m=impulseest(data,N);
impulse(m,Time);
step(m,Time);

where data is a single- or multiple-output iddata or idfrd object. N is a scalar value

specifying the order of the FIR system corresponding to the time range 0:Ts:(N-1)*Ts,
where Ts is the data sampling time.

You can also specify estimation options, such as regularizing kernel, pre-whitening
filter order and data offsets, using impulseestOptions and pass them as an input to
impulseest. For example:
opt = impulseestOptions('RegulKernel','TC'));

3-18
 Identifying Impulse-Response Models

m = impulseest(data,N,opt);

To view the confidence region for the estimated response, use impulseplot and
stepplot to create the plot. Then use showConfidence.

For example:
h = stepplot(m,Time);
showConfidence(h,3) % 3 std confidence region

Note: cra is an alternative method for computing impulse response from time-domain
data only.

Next Steps

• Perform model analysis. See “Validating Models After Estimation”.

Compute Response Values

You can use impulse and step commands with output arguments to get the numerical
impulse- and step-response vectors as a function of time, respectively.

To get the numerical response values:

1 Compute the FIR model by using impulseest, as described in “Estimate Impulse-

Response Models at the Command Line” on page 3-18.
2 Apply the following syntax on the resulting model:
% To compute impulse-response data
[y,t,~,ysd] = impulse(model)
% To compute step-response data
[y,t,~,ysd] = step(model)

where y is the response data, t is the time vector, and ysd is the standard deviations
of the response.

Identify Delay Using Transient-Response Plots

You can use transient-response plots to estimate the input delay, or dead time, of linear
systems. Input delay represents the time it takes for the output to respond to the input.

3-19
3 Linear Model Identification

In the System Identification app. To view the transient response plot, select the
Transient resp check box in the System Identification app. For example, the following
step response plot shows a time delay of about 0.25 s before the system responds to the
input.

Step Response Plot

At the command line. You can use impulseplot to plot the impulse response. The
time delay is equal to the first positive peak in the transient response magnitude that is
greater than the confidence region for positive time values.

For example, the following commands create an impulse-response plot with a 1-standard-
deviation confidence region:

load dry2
ze = dry2(1:500);
opt = impulseestOptions('RegulKernel','TC');
sys = impulseest(ze,40,opt);
h = impulseplot(sys);
showConfidence(h,1);

3-20
 Identifying Impulse-Response Models

The resulting figure shows that the first positive peak of the response magnitude, which
is greater than the confidence region for positive time values, occurs at 0.24 s.

Instead of using showConfidence, you can plot the confidence interval interactively, by
right-clicking on the plot and selecting Characteristics > Confidence Region.

Correlation Analysis Algorithm

Correlation analysis refers to methods that estimate the impulse response of a linear
model, without specific assumptions about model orders.

3-21
3 Linear Model Identification

The impulse response, g, is the system's output when the input is an impulse signal. The
output response to a general input, u(t), is obtained as the convolution with the impulse
response. In continuous time:

t
y( t) = Ú-• g (t ) u (t - t )dt
In discrete-time:

•
y (t ) = Â g (k) u ( t - k)
k= 1

The values of g(k) are the discrete time impulse response coefficients.

You can estimate the values from observed input-output data in several different ways.
impulseest estimates the first n coefficients using the least-squares method to obtain a
finite impulse response (FIR) model of order n.

Several important options are associated with the estimate:

• Prewhitening — The input can be pre-whitened by applying an input-whitening

filter of order PW to the data. This minimizes the effect of the neglected tail (k > n) of
the impulse response.

1 A filter of order PW is applied such that it whitens the input signal u:

1/A = A(u)e, where A is a polynomial and e is white noise.

2 The inputs and outputs are filtered using the filter:

uf = Au, yf = Ay
3 The filtered signals uf and yf are used for estimation.

You can specify prewhitening using the PW name-value pair argument of

impulseestOptions.
• Regularization — The least-squares estimate can be regularized. This means that
a prior estimate of the decay and mutual correlation among g(k) is formed and used
to merge with the information about g from the observed data. This gives an estimate
with less variance, at the price of some bias. You can choose one of the several kernels
to encode the prior estimate.

3-22
 Identifying Impulse-Response Models

This option is essential because, often, the model order n can be quite large. In
cases where there is no regularization, n can be automatically decreased to secure a
reasonable variance.

You can specify the regularizing kernel using the RegulKernel Name-Value pair
argument of impulseestOptions.
• Autoregressive Parameters — The basic underlying FIR model can be
complemented by NA autoregressive parameters, making it an ARX model.

n NA
y (t ) = Â g (k) u ( t - k) - Â ak y (t - k)
k= 1 k=1

This gives both better results for small n and allows unbiased estimates when
data are generated in closed loop. impulseest uses NA = 5 for t>0 and NA = 0 (no
autoregressive component) for t<0.
• Noncausal effects — Response for negative lags. It may happen that the data has
been generated partly by output feedback:

•
u(t) = Â h(k) y( t - k) + r (t )
k= 0

where h(k) is the impulse response of the regulator and r is a setpoint or

disturbance term. The existence and character of such feedback h can be
estimated in the same way as g, simply by trading places between y and u in
the estimation call. Using impulseest with an indication of negative delays,
mi = impulseest ( data, nk, nb), nk < 0 , returns a model mi with an impulse response

[ h(- nk), h(-nk - 1),..., h(0), g(1), g(2),..., g(nb + nk) ]

aligned so that it corresponds to lags [ nk, nk + 1,.., 0,1, 2,..., nb + nk] . This is achieved
because the input delay (InputDelay) of model mi is nk.

For a multi-input multi-output system, the impulse response g(k) is an ny-by-nu matrix,
where ny is the number of outputs and nu is the number of inputs. The i–j element of the
matrix g(k) describes the behavior of the ith output after an impulse in the jth input.

3-23
3 Linear Model Identification

Identifying Process Models

In this section...
“What Is a Process Model?” on page 3-24
“Data Supported by Process Models” on page 3-25
“Estimate Process Models Using the App” on page 3-25
“Estimate Process Models at the Command Line” on page 3-29
“Process Model Structure Specification” on page 3-35
“Estimating Multiple-Input, Multi-Output Process Models” on page 3-36
“Disturbance Model Structure for Process Models” on page 3-37
“Assigning Estimation Weightings” on page 3-37
“Specifying Initial Conditions for Iterative Estimation Algorithms” on page 3-38

What Is a Process Model?

The structure of a process model is a simple continuous-time transfer function that
describes linear system dynamics in terms of one or more of the following elements:

• Static gain Kp.

• One or more time constants Tpk. For complex poles, the time constant is called T —
w
equal to the inverse of the natural frequency—and the damping coefficient is z
(zeta).
• Process zero Tz.
• Possible time delay Td before the system output responds to the input (dead time).
• Possible enforced integration.

Process models are popular for describing system dynamics in many industries and apply
to various production environments. The advantages of these models are that they are
simple, support transport delay estimation, and the model coefficients have an easy
interpretation as poles and zeros.

You can create different model structures by varying the number of poles, adding an
integrator, or adding or removing a time delay or a zero. You can specify a first-, second-,
or third-order model, and the poles can be real or complex (underdamped modes).

3-24
 Identifying Process Models

For example, the following model structure is a first-order continuous-time process

model, where K is the static gain, Tp1 is a time constant, and Td is the input-to-output
delay:

Kp
G ( s) = e- sTd
1 + sTp1

Data Supported by Process Models

You can estimate low-order (up to third order), continuous-time transfer functions
using regularly sampled time- or frequency-domain iddata or idfrd data objects. The
frequency-domain data may have a zero sample time.

You must import your data into the MATLAB workspace, as described in “Data
Preparation”.

Estimate Process Models Using the App

Before you can perform this task, you must have

• Imported data into the System Identification app. See “Import Time-Domain Data
into the App” on page 2-16. For supported data formats, see “Data Supported by
Process Models” on page 3-25.
• Performed any required data preprocessing operations. If you need to model nonzero
offsets, such as when model contains integration behavior, do not detrend your data.
In other cases, to improve the accuracy of your model, you should detrend your data.
See “Ways to Prepare Data for System Identification” on page 2-6.

1 In the System Identification app, select Estimate > Process models to open the
Process Models dialog box.

3-25
3 Linear Model Identification

To learn more about the options in the dialog box, click Help.
2 If your model contains multiple inputs, select the input channel in the Input list.
This list only appears when you have multiple inputs. For more information, see
“Estimating Multiple-Input, Multi-Output Process Models” on page 3-36.
3 In the Model Transfer Function area, specify the model structure using the
following options:

• Under Poles, select the number of poles, and then select All real or
Underdamped.

Note: You need at least two poles to allow underdamped modes (complex-
conjugate pair).
• Select the Zero check box to include a zero, which is a numerator term other than
a constant, or clear the check box to exclude the zero.

3-26
 Identifying Process Models

• Select the Delay check box to include a delay, or clear the check box to exclude
the delay.
• Select the Integrator check box to include an integrator (self-regulating process),
or clear the check box to exclude the integrator.

The Parameter area shows as many active parameters as you included in the model
structure.

Note: By default, the model Name is set to the acronym that reflects the model
structure, as described in “Process Model Structure Specification” on page 3-35.
4 In the Initial Guess area, select Auto-selected to calculate the initial parameter
values for the estimation. The Initial Guess column in the Parameter table displays
Auto. If you do not have a good guess for the parameter values, Auto works better
than entering an ad hoc value.

5 (Optional) If you approximately know a parameter value, enter this value in the
Initial Guess column of the Parameter table. The estimation algorithm uses this
value as a starting point. If you know a parameter value exactly, enter this value in
the Initial Guess column, and also select the corresponding Known check box in
the table to fix its value.

3-27
3 Linear Model Identification

If you know the range of possible values for a parameter, enter these values into the
corresponding Bounds field to help the estimation algorithm.

For example, the following figure shows that the delay value Td is fixed at 2 s and is
not estimated.

6 In the Disturbance Model list, select one of the available options. For more
information about each option, see “Disturbance Model Structure for Process Models”
on page 3-37.
7 In the Focus list, select how to weigh the relative importance of the fit at different
frequencies. For more information about each option, see “Assigning Estimation
Weightings” on page 3-37.
8 In the Initial state list, specify how you want the algorithm to treat initial states.
For more information about the available options, see “Specifying Initial Conditions
for Iterative Estimation Algorithms” on page 3-38.

Tip If you get a bad fit, you might try setting a specific method for handling initial
states, rather than choosing it automatically.

3-28
 Identifying Process Models

9 In the Covariance list, select Estimate if you want the algorithm to compute
parameter uncertainties. Effects of such uncertainties are displayed on plots as
model confidence regions.

To omit estimating uncertainty, select None. Skipping uncertainty computation

might reduce computation time for complex models and large data sets.
10 In the Model Name field, edit the name of the model or keep the default. The name
of the model should be unique in the Model Board.
11 To view the estimation progress, select the Display Progress check box. This opens
a progress viewer window in which the estimation progress is reported.
12 Click Regularization to obtain regularized estimates of model parameters. Specify
the regularization constants in the Regularization Options dialog box. To learn more,
see “Regularized Estimates of Model Parameters”.
13 Click Estimate to add this model to the Model Board in the System Identification
app.
14 To stop the search and save the results after the current iteration has been
completed, click Stop Iterations. To continue iterations from the current model,
click the Continue button to assign current parameter values as initial guesses for
the next search.

Next Steps

• Validate the model by selecting the appropriate check box in the Model Views area
of the System Identification app. For more information about validating models, see
“Validating Models After Estimation” on page 10-2.
• Refine the model by clicking the Value —> Initial Guess button to assign current
parameter values as initial guesses for the next search, edit the Name field, and click
Estimate.
• Export the model to the MATLAB workspace for further analysis by dragging it to the
To Workspace rectangle in the System Identification app.

Estimate Process Models at the Command Line

• “Prerequisites” on page 3-30
• “Using procest to Estimate Process Models” on page 3-30
• “Estimate Process Models with Free Parameters” on page 3-31
• “Estimate Process Models with Fixed Parameters” on page 3-32

3-29
3 Linear Model Identification

Prerequisites

Before you can perform this task, you must have

• Input-output data as an iddata object or frequency response data as frd or idfrd

objects. See “Representing Time- and Frequency-Domain Data Using iddata Objects”
on page 2-50. For supported data formats, see “Data Supported by Process Models” on
page 3-25.
• Performed any required data preprocessing operations. When working with time
domain data, if you need to model nonzero offsets, such as when model contains
integration behavior, do not detrend your data. In other cases, to improve the
accuracy of your model, you should detrend your data. See “Ways to Prepare Data for
System Identification” on page 2-6.

Using procest to Estimate Process Models

You can estimate process models using the iterative estimation method procest that
minimizes the prediction errors to obtain maximum likelihood estimates. The resulting
models are stored as idproc model objects.

You can use the following general syntax to both configure and estimate process models:

m = procest(data,mod_struc,opt)

data is the estimation data and mod_struc is one of the following:

• A string that represents the process model structure, as described in “Process Model
Structure Specification” on page 3-35.
• A template idproc model. opt is an option set for configuring the estimation of
the process model, such as handling of initial conditions, input offset and numerical
search method.

Tip You do not need to construct the model object using idproc before estimation unless
you want to specify initial parameter guesses, minimum/maximum bounds, or fixed
parameter values, as described in “Estimate Process Models with Fixed Parameters” on
page 3-32.

For more information about validating a process model, see “Validating Models After
Estimation” on page 10-2.

3-30
 Identifying Process Models

You can use procest to refine parameter estimates of an existing process model, as
described in “Refining Linear Parametric Models” on page 3-86.

For detailed information, see procest and idproc.

Estimate Process Models with Free Parameters

This example shows how to estimate the parameters of a first-order process model:

This process has two inputs and the response from each input is estimated by a first-
order process model. All parameters are free to vary.
load co2data
% Sampling interval is 0.5 min (known)
Ts = 0.5;
% Split data set into estimation data ze and validation data zv
ze = iddata(Output_exp1,Input_exp1,Ts,...
'TimeUnit','min');
zv = iddata(Output_exp2,Input_exp2,Ts,...
'TimeUnit','min');

Estimate model with one pole, a delay, and a first-order disturbance component. The
data contains known offsets. Specify them using the InputOffset and OutputOffset
options.
opt = procestOptions;
opt.InputOffset = [170;50];
opt.OutputOffset = -45;
opt.Display = 'on';
opt.DisturbanceModel = 'arma1';
m = procest(ze,'p1d',opt)

m =
Process model with 2 inputs: y = G11(s)u1 + G12(s)u2
From input "u1" to output "y1":
Kp
G11(s) = ---------- * exp(-Td*s)
1+Tp1*s

Kp = 2.6542

3-31
3 Linear Model Identification

Tp1 = 0.15451
Td = 2.3185

From input "u2" to output "y1":

Kp
G12(s) = ---------- * exp(-Td*s)
1+Tp1*s

Kp = 9.9754
Tp1 = 2.0653
Td = 4.9195

An additive ARMA disturbance model exists for output "y1":

y = G u + (C/D)e

C(s) = s + 2.677
D(s) = s + 0.6237

Parameterization:
'P1D' 'P1D'
Number of free coefficients: 8
Use "getpvec", "getcov" for parameters and their uncertainties.

Status:
Estimated using PROCEST on time domain data "ze".
Fit to estimation data: 91.07% (prediction focus)
FPE: 2.431, MSE: 2.412

Use dot notation to get the value of any model parameter. For example, get the value of
dc gain parameter Kp .
m.Kp

ans =

2.6542 9.9754

Estimate Process Models with Fixed Parameters

This example shows how to estimate a process model with fixed parameters.

When you know the values of certain parameters in the model and want to estimate only
the values you do not know, you must specify the fixed parameters after creating the

3-32
 Identifying Process Models

idproc model object. Use the following commands to prepare the data and construct a
process model with one pole and a delay:
% Load data
load co2data
% Sampling interval is 0.5 min (known)
Ts = 0.5;
% Split data set into estimation data ze and validation data zv
ze = iddata(Output_exp1,Input_exp1,Ts,...
'TimeUnit','min');
zv = iddata(Output_exp2,Input_exp2,Ts,...
'TimeUnit','min');
mod = idproc({'p1d','p1d'},'TimeUnit','min')

mod =
Process model with 2 inputs: y = G11(s)u1 + G12(s)u2
From input 1 to output 1:
Kp
G11(s) = ---------- * exp(-Td*s)
1+Tp1*s

Kp = NaN
Tp1 = NaN
Td = NaN

From input 2 to output 1:

Kp
G12(s) = ---------- * exp(-Td*s)
1+Tp1*s

Kp = NaN
Tp1 = NaN
Td = NaN

Parameterization:
'P1D' 'P1D'
Number of free coefficients: 6
Use "getpvec", "getcov" for parameters and their uncertainties.

Status:
Created by direct construction or transformation. Not estimated.

The model parameters Kp , Tp1 , and Td are assigned NaN values, which means that the
parameters have not yet been estimated from the data.

3-33
3 Linear Model Identification

Use the Structure model property to specify the initial guesses for unknown
parameters, minimum/maximum parameter bounds and fix known parameters.

Set the value of Kp for the second transfer function to 10 and specify it as a fixed
parameter. Initialize the delay values for the two transfer functions to 2 and 5 minutes,
respectively. Specify them as free estimation parameters.
mod.Structure(2).Kp.Value = 10;
mod.Structure(2).Kp.Free = false;

mod.Structure(1).Tp1.Value = 2;
mod.Structure(2).Td.Value = 5;

Estimate Tp1 and Td only.

mod_proc = procest(ze, mod)

mod_proc =
Process model with 2 inputs: y = G11(s)u1 + G12(s)u2
From input "u1" to output "y1":
Kp
G11(s) = ---------- * exp(-Td*s)
1+Tp1*s

Kp = -3.2213
Tp1 = 2.522
Td = 3.8285

From input "u2" to output "y1":

Kp
G12(s) = ---------- * exp(-Td*s)
1+Tp1*s

Kp = 10
Tp1 = 2.4382
Td = 4.111

Parameterization:
'P1D' 'P1D'
Number of free coefficients: 5
Use "getpvec", "getcov" for parameters and their uncertainties.

Status:
Estimated using PROCEST on time domain data "ze".

3-34
 Identifying Process Models

Fit to estimation data: 71.66%

FPE: 24.42, MSE: 24.3

In this case, the value of Kp is fixed, but Tp1 and Td are estimated.

Process Model Structure Specification

This topic describes how to specify the model structure in the estimation procedures
“Estimate Process Models Using the App” on page 3-25 and “Estimate Process Models
at the Command Line” on page 3-29.

In the System Identification app. Specify the model structure by selecting the
number of real or complex poles, and whether to include a zero, delay, and integrator.
The resulting transfer function is displayed in the Process Models dialog box.

At the command line. Specify the model structure using an acronym that includes the
following letters and numbers:

• (Required) P for a process model

• (Required) 0, 1, 2 or 3 for the number of poles
•
(Optional) D to include a time-delay term e-sTd
• (Optional) Z to include a process zero (numerator term)
• (Optional) U to indicate possible complex-valued (underdamped) poles
• (Optional) I to indicate enforced integration

Typically, you specify the model-structure acronym as a string argument in the

estimation command procest:

• procest(data,'P1D') to estimate the following structure:

Kp
G ( s) = e- sTd
1 + sTp1

• procest(data,'P2ZU') to estimate the following structure:

K p (1 + sTz )
G ( s) =
1 + 2 sz Tw + s2Tw2

3-35
3 Linear Model Identification

• procest(data,'P0ID') to estimate the following structure:

K p -sT
G ( s) = e d
s
• procest(data,'P3Z') to estimate the following structure:

K p (1 + sTz )
G ( s) =
(1 + sTp1 ) (1 + sT p2 )(1 + sTp3 )
For more information about estimating models, see “Estimate Process Models at the
Command Line” on page 3-29.

Estimating Multiple-Input, Multi-Output Process Models

If your model contains multiple inputs, multiple outputs, or both, you can specify
whether to estimate the same transfer function for all input-output pairs, or a different
transfer function for each. The information in this section supports the estimation
procedures “Estimate Process Models Using the App” on page 3-25 and “Estimate
Process Models at the Command Line” on page 3-29.

In the System Identification app. To fit a data set with multiple inputs, or multiple
outputs, or both, in the Process Models dialog box, configure the process model settings
for one input-output pair at a time. Use the input and output selection lists to switch to a
different input/output pair.

If you want the same transfer function to apply to all input/output pairs, select the All
same check box. To apply a different structure to each channel, leave this check box
clear, and create a different transfer function for each input.

At the command line. Specify the model structure as a cell array of acronym strings in
the estimation command procest. For example, use this command to specify the first-
order transfer function for the first input, and a second-order model with a zero and an
integrator for the second input:

m = idproc({'P1','P2ZI'})
m = procest(data,m)

To apply the same structure to all inputs, define a single structure in idproc.

3-36
 Identifying Process Models

Disturbance Model Structure for Process Models

This section describes how to specify a noise model in the estimation procedures
“Estimate Process Models Using the App” on page 3-25 and “Estimate Process Models
at the Command Line” on page 3-29.

In addition to the transfer function G, a linear system can include an additive noise term
He, as follows:

y = Gu + He

where e is white noise.

You can estimate only the dynamic model G, or estimate both the dynamic model and the
disturbance model H. For process models, H is a rational transfer function C/D, where
the C and D polynomials for a first- or second-order ARMA model.

In the app. To specify whether to include or exclude a noise model in the Process Models
dialog box, select one of the following options from the Disturbance Model list:

• None — The algorithm does not estimate a noise model (C=D=1). This option also sets
Focus to Simulation.
• Order 1 — Estimates a noise model as a continuous-time, first-order ARMA model.
• Order 2 — Estimates a noise model as a continuous-time, second-order ARMA
model.

At the command line. Specify the disturbance model using the procestOptions
option set. For example, use this command to estimate a first-order transfer function and
a first-order noise model:
opt = procestOptions;
opt.DisturbanceModel = 'arma1';
model = procest(data, 'P1D', opt);

For a complete list of values for the DisturbanceModel model property, see the
procestOptions reference page.

Assigning Estimation Weightings

You can specify how the estimation algorithm weighs the fit at various frequencies. This
information supports the estimation procedures “Estimate Process Models Using the

3-37
3 Linear Model Identification

App” on page 3-25 and “Estimate Process Models at the Command Line” on page
3-29.

In the System Identification app. Set Focus to one of the following options:

• Prediction — Uses the inverse of the noise model H to weigh the relative
importance of how closely to fit the data in various frequency ranges. Corresponds to
minimizing one-step-ahead prediction, which typically favors the fit over a short time
interval. Optimized for output prediction applications.
• Simulation — Uses the input spectrum to weigh the relative importance of the
fit in a specific frequency range. Does not use the noise model to weigh the relative
importance of how closely to fit the data in various frequency ranges. Optimized for
output simulation applications.
• Stability — Behaves the same way as the Prediction option, but also forces the
model to be stable. For more information about model stability, see “Unstable Models”
on page 10-88.
• Filter — Specify a custom filter to open the Estimation Focus dialog box, where
you can enter a filter, as described in “Simple Passband Filter” on page 2-128 or
“Defining a Custom Filter” on page 2-129. This prefiltering applies only for estimating
the dynamics from input to output. The disturbance model is determined from the
estimation data.

At the command line. Specify the focus using the procestOptions option set. For
example, use this command to optimize the fit for simulation and estimate a disturbance
model:

opt = procestOptions('DisturbanceModel','arma2', 'Focus','simulation');

model = procest(data, 'P1D', opt);

Specifying Initial Conditions for Iterative Estimation Algorithms

You can optionally specify how the iterative algorithm treats initial conditions for
estimation of model parameters. This information supports the estimation procedures
“Estimate Process Models Using the App” on page 3-25 and “Estimate Process Models
at the Command Line” on page 3-29.

In the System Identification app. Set Initial condition to one of the following
options:

• Zero — Sets all initial states to zero.

3-38
 Identifying Process Models

• Estimate — Treats the initial states as an unknown vector of parameters and

estimates these states from the data.
• Backcast — Estimates initial states using a backward filtering method (least-
squares fit).
• U-level est — Estimates both the initial conditions and input offset levels. For
multiple inputs, the input level for each input is estimated individually. Use if you
included an integrator in the transfer function.
• Auto — Automatically chooses one of the preceding options based on the estimation
data. If the initial conditions have negligible effect on the prediction errors, they are
taken to be zero to optimize algorithm performance.

At the command line. Specify the initial conditions using the InitialCondition
model estimation option, configured using the procestOptions command. For example,
use this command to estimate a first-order transfer function and set the initial states to
zero:
opt = procestOptions('InitialCondition','zero');
model = procest(data, 'P1D', opt)

3-39
3 Linear Model Identification

Identifying Input-Output Polynomial Models

In this section...
“What Are Polynomial Models?” on page 3-40
“Data Supported by Polynomial Models” on page 3-45
“Preliminary Step – Estimating Model Orders and Input Delays” on page 3-46
“Estimate Polynomial Models in the App” on page 3-54
“Estimate Polynomial Models at the Command Line” on page 3-58
“Polynomial Sizes and Orders of Multi-Output Polynomial Models” on page 3-61
“Assigning Estimation Weightings” on page 3-65
“Specifying Initial States for Iterative Estimation Algorithms” on page 3-65
“Polynomial Model Estimation Algorithms” on page 3-66
“Estimate Models Using armax” on page 3-67

What Are Polynomial Models?

• “Polynomial Model Structure” on page 3-40
• “Understanding the Time-Shift Operator q” on page 3-41
• “Different Configurations of Polynomial Models” on page 3-42
• “Continuous-Time Representation of Polynomial Models” on page 3-44
• “Multi-Output Polynomial Models” on page 3-45

Polynomial Model Structure

A polynomial model uses a generalized notion of transfer functions to express the

relationship between the input, u(t), the output y(t), and the noise e(t) using the equation:
nu
B (q) C( q)
A( q) y( t) = Â Fii(q) ui (t - nki ) + D(q) e(t)
i=1

The variables A, B, C, D, and F are polynomials expressed in the time-shift operator

q^-1. ui is the ith input, nu is the total number of inputs, and nki is the ith input delay
that characterizes the transport delay. The variance of the white noise e(t) is assumed to
be l . For more information about the time-shift operator, see “Understanding the Time-
Shift Operator q” on page 3-41.

3-40
 Identifying Input-Output Polynomial Models

In practice, not all the polynomials are simultaneously active. Often, simpler forms, such
as ARX, ARMAX, Output-Error, and Box-Jenkins are employed. You also have the option
of introducing an integrator in the noise source so that the general model takes the form:

nu
B (q) C( q) 1
A( q) y( t) = Â Fii(q) ui (t - nki ) + D(q) 1 - q-1 e(t)
i=1

For more information, see “Different Configurations of Polynomial Models” on page

3-42.

You can estimate polynomial models using time or frequency domain data.

For estimation, you must specify the model order as a set of integers that represent the
number of coefficients for each polynomial you include in your selected structure—na for
A, nb for B, nc for C, nd for D, and nf for F. You must also specify the number of samples
nk corresponding to the input delay—dead time—given by the number of samples before
the output responds to the input.

The number of coefficients in denominator polynomials is equal to the number of poles,

and the number of coefficients in the numerator polynomials is equal to the number of
zeros plus 1. When the dynamics from u(t) to y(t) contain a delay of nk samples, then the
first nk coefficients of B are zero.

For more information about the family of transfer-function models, see the corresponding
section in System Identification: Theory for the User, Second Edition, by Lennart Ljung,
Prentice Hall PTR, 1999.

Understanding the Time-Shift Operator q

The general polynomial equation is written in terms of the time-shift operator q–1. To
understand this time-shift operator, consider the following discrete-time difference
equation:

y( t) + a1 y(t - T) + a2 y(t - 2T ) =
b1u(t - T ) + b2 u( t - 2 T)

where y(t) is the output, u(t) is the input, and T is the sampling interval. q-1 is
a time-shift operator that compactly represents such difference equations using
q-1u(t) = u( t - T ) :

3-41
3 Linear Model Identification

y( t) + a1 q -1 y(t) + a2q -2 y(t) =

b1 q -1u(t) + b2q -2u(t)
or
A( q) y( t) = B(q) u( t)

In this case, A( q) = 1 + a1 q-1 + a2q -2 and B( q) = b1q -1 + b2 q-2 .

Note: This q description is completely equivalent to the Z-transform form: q corresponds

to z.

Different Configurations of Polynomial Models

These model structures are subsets of the following general polynomial equation:

nu
B (q) C( q)
A( q) y( t) = Â Fii(q) ui (t - nki ) + D(q) e(t)
i=1

The model structures differ by how many of these polynomials are included in the
structure. Thus, different model structures provide varying levels of flexibility for
modeling the dynamics and noise characteristics.

The following table summarizes common linear polynomial model structures supported
by the System Identification Toolbox product. If you have a specific structure in mind for
your application, you can decide whether the dynamics and the noise have common or
different poles. A(q) corresponds to poles that are common for the dynamic model and the
noise model. Using common poles for dynamics and noise is useful when the disturbances
enter the system at the input. F i determines the poles unique to the system dynamics,
and D determines the poles unique to the disturbances.

Model Structure Equation Description

ARX nu The noise model is 1A and the
A( q) y( t) = Â Bi (q)ui ( t - nki ) + e(t) noise is coupled to the dynamics
i=1
model. ARX does not let you
model noise and dynamics
independently. Estimate an ARX

3-42
 Identifying Input-Output Polynomial Models

Model Structure Equation Description

model to obtain a simple model at
good signal-to-noise ratios.
ARIX 1 Extends the ARX structure by
Ay = Bu + e including an integrator in the
1 - q-1 noise source, e(t). This is useful in
cases where the disturbance is not
stationary.
ARMAX nu Extends the ARX structure
A( q) y( t) = Â Bi (q)ui ( t - nki ) + C(q)e(t) by providing more flexibility
i=1 for modeling noise using the C
parameters (a moving average of
white noise). Use ARMAX when
the dominating disturbances enter
at the input. Such disturbances
are called load disturbances.
ARIMAX 1 Extends theARMAX structure
Ay = Bu + C
-1
e by including an integrator in the
1-q noise source, e(t). This is useful in
cases where the disturbance is not
stationary.
Box-Jenkins (BJ) nu Provides completely independent
Bi (q) C( q)
y( t) = Â ui ( t - nki ) + e(t) parameterization for the dynamics
i =1 Fi ( q) D( q) and the noise using rational
polynomial functions.
Use BJ models when the noise
does not enter at the input,
but is primary a measurement
disturbance, This structure
provides additional flexibility for
modeling noise.

3-43
3 Linear Model Identification

Model Structure Equation Description

Output-Error nu Use when you want to
Bi (q)
(OE) y( t) = Â ui ( t - nki ) + e( t) parameterize dynamics, but do not
i =1 Fi ( q) want to estimate a noise model.

Note: In this case, the noise

models is H = 1 in the general
equation and the white noise
source e(t) affects only the output.

The polynomial models can contain one or more outputs and zero or more inputs.

The System Identification app supports direct estimation of ARX, ARMAX, OE and BJ
models. You can add a noise integrator to the ARX, ARMAX and BJ forms. However,
you can use polyest to estimate all five polynomial or any subset of polynomials in the
general equation. For more information about working with pem, see “Using polyest to
Estimate Polynomial Models” on page 3-59.

Continuous-Time Representation of Polynomial Models

In continuous time, the general frequency-domain equation is written in terms of the

Laplace transform variable s, which corresponds to a differentiation operation:

B( s) C( s)
A( s) Y ( s) = U ( s) + E( s)
F(s) D( s)

In the continuous-time case, the underlying time-domain model is a differential equation

and the model order integers represent the number of estimated numerator and
denominator coefficients. For example, na=3 and nb=2 correspond to the following model:

A( s) = s4 + a1 s3 + a2 s2 + a3
B( s) = b1 s + b2

You can only estimate continuous-time polynomial models directly using continuous-time
frequency-domain data. In this case, you must set the Ts data property to 0 to indicate
that you have continuous-time frequency-domain data, and use the oe command to
estimate an Output-Error polynomial model. Continuous-time models of other structures

3-44
 Identifying Input-Output Polynomial Models

such as ARMAX or BJ cannot be estimated. You can obtain those forms only by direct
construction (using idpoly), conversion from other model types, or by converting a
discrete-time model into continuous-time (d2c). Note that the OE form represents a
transfer function expressed as a ratio of numerator (B) and denominator (F) polynomials.
For such forms consider using the transfer function models, represented by idtf models.
You can estimate transfer function models using both time and frequency domain data.
In addition to the numerator and denominator polynomials, you can also estimate
transport delays. See idtf and tfest for more information.

Multi-Output Polynomial Models

You can create multi-output polynomial models by using the idpoly command or
estimate them using ar, arx, bj, oe, armax, and polyest. In the app, you can estimate
such models by choosing a multi-output data set and setting the orders appropriately
in the Polynomial Models dialog box. For more details on the orders of multi-output
models, see “Polynomial Sizes and Orders of Multi-Output Polynomial Models” on page
3-61.

Data Supported by Polynomial Models

• “Types of Supported Data” on page 3-45
• “Designating Data for Estimating Continuous-Time Models” on page 3-46
• “Designating Data for Estimating Discrete-Time Models” on page 3-46

Types of Supported Data

You can estimate linear, black-box polynomial models from data with the following
characteristics:

• Time- or frequency-domain data (iddata or idfrd data objects).

Note: For frequency-domain data, you can only estimate ARX and OE models.

To estimate polynomial models for time-series data, see “Time-Series Model

Identification”.
• Real data or complex data in any domain.
• Single-output and multiple-output.

3-45
3 Linear Model Identification

You must import your data into the MATLAB workspace, as described in “Data
Preparation”.

Designating Data for Estimating Continuous-Time Models

To get a linear, continuous-time model of arbitrary structure for time-domain data, you
can estimate a discrete-time model, and then use d2c to transform it to a continuous-
time model.

For continuous-time frequency-domain data, you can estimate directly only Output-
Error (OE) continuous-time models. Other structures include noise models, which is not
supported for frequency-domain data.

Tip To denote continuous-time frequency-domain data, set the data sampling interval
to 0. You can set the sampling interval when you import data into the app or set the Ts
property of the data object at the command line.

Designating Data for Estimating Discrete-Time Models

You can estimate arbitrary-order, linear state-space models for both time- or frequency-
domain data.

Set the data property Ts to:

• 0, for frequency response data that is measured directly from an experiment.

• Equal to the Ts of the original data, for frequency response data obtained by
transforming time-domain iddata (using spa and etfe).

Tip You can set the sampling interval when you import data into the app or set the Ts
property of the data object at the command line.

Preliminary Step – Estimating Model Orders and Input Delays

• “Why Estimate Model Orders and Delays?” on page 3-47
• “Estimating Orders and Delays in the App” on page 3-47
• “Estimating Model Orders at the Command Line” on page 3-50

3-46
 Identifying Input-Output Polynomial Models

• “Estimating Delays at the Command Line” on page 3-51

• “Selecting Model Orders from the Best ARX Structure” on page 3-52

Why Estimate Model Orders and Delays?

To estimate polynomial models, you must provide input delays and model orders. If you
already have insight into the physics of your system, you can specify the number of poles
and zeros.

In most cases, you do not know the model orders in advance. To get initial model orders
and delays for your system, you can estimate several ARX models with a range of orders
and delays and compare the performance of these models. You choose the model orders
that correspond to the best model performance and use these orders as an initial guess
for further modeling.

Because this estimation procedure uses the ARX model structure, which includes the A
and B polynomials, you only get estimates for the na, nb, and nk parameters. However,
you can use these results as initial guesses for the corresponding polynomial orders and
input delays in other model structures, such as ARMAX, OE, and BJ.

If the estimated nk is too small, the leading nb coefficients are much smaller than their
standard deviations. Conversely, if the estimated nk is too large, there is a significant
correlation between the residuals and the input for lags that correspond to the missing
B terms. For information about residual analysis plots, see “Residual Analysis” on page
10-23.

Estimating Orders and Delays in the App

The following procedure assumes that you have already imported your data into the
app and performed any necessary preprocessing operations. For more information, see
“Represent Data”.

To estimate model orders and input delays in the System Identification app:

1 In the System Identification app, select Estimate > Polynomial Models to open
the Polynomials Models dialog box.

The ARX model is already selected by default in the Structure list.

Note: For time-series models, select the AR model structure.

3-47
3 Linear Model Identification

2 Edit the Orders field to specify a range of poles, zeros, and delays. For example,
enter the following values for na, nb, and nk:

[1:10 1:10 1:10]

Tip As a shortcut for entering 1:10 for each required model order, click Order
Selection.

3 Click Estimate to open the ARX Model Structure Selection window, which displays
the model performance for each combination of model parameters. The following
figure shows an example plot.

3-48
 Identifying Input-Output Polynomial Models

4 Select a rectangle that represents the optimum parameter combination and click
Insert to estimates a model with these parameters. For information about using this
plot, see “Selecting Model Orders from the Best ARX Structure” on page 3-52.

This action adds a new model to the Model Board in the System Identification app.
The default name of the parametric model contains the model type and the number
of poles, zeros, and delays. For example, arx692 is an ARX model with na=6, nb=9,
and a delay of two samples.
5 Click Close to close the ARX Model Structure Selection window.

Note: You cannot estimate model orders when using multi-output data.

After estimating model orders and delays, use these values as initial guesses for
estimating other model structures, as described in “Estimate Polynomial Models in the
App” on page 3-54.

3-49
3 Linear Model Identification

Estimating Model Orders at the Command Line

You can estimate model orders using the struc, arxstruc, and selstruc commands in
combination.

If you are working with a multiple-output system, you must use the struc, arxstruc,
and selstruc commands one output at a time. You must subreference the correct
output channel in your estimation and validation data sets.

For each estimation, you use two independent data sets—an estimation data set and a
validation data set. These independent data set can be from different experiments, or
data subsets from a single experiment. For more information about subreferencing data,
see “Select Data Channels, I/O Data and Experiments in iddata Objects” on page 2-55
and “Select I/O Channels and Data in idfrd Objects” on page 2-85.

For an example of estimating model orders for a multiple-input system, see “Estimating
Delays in the Multiple-Input System” in System Identification Toolbox Getting Started
Guide.
struc

The struc command creates a matrix of possible model-order combinations for a

specified range of na, nb, and nk values.

For example, the following command defines the range of model orders and delays
na=2:5, nb=1:5, and nk=1:5:

NN = struc(2:5,1:5,1:5))

arxstruc

The arxstruc command takes the output from struc, estimates an ARX model for each
model order, and compares the model output to the measured output. arxstruc returns
the loss for each model, which is the normalized sum of squared prediction errors.

For example, the following command uses the range of specified orders NN to compute the
loss function for single-input/single-output estimation data data_e and validation data
data_v:

V = arxstruc(data_e,data_v,NN)

Each row in NN corresponds to one set of orders:

3-50
 Identifying Input-Output Polynomial Models

[na nb nk]

selstruc

The selstruc command takes the output from arxstruc and opens the ARX Model
Structure Selection window to guide your choice of the model order with the best
performance.

For example, to open the ARX Model Structure Selection window and interactively
choose the optimum parameter combination, use the following command:

selstruc(V)

For more information about working with the ARX Model Structure Selection window,
see “Selecting Model Orders from the Best ARX Structure” on page 3-52.

To find the structure that minimizes Akaike's Information Criterion, use the following
command:

nn = selstruc(V,'AIC')

where nn contains the corresponding na, nb, and nk orders.

Similarly, to find the structure that minimizes the Rissanen's Minimum Description
Length (MDL), use the following command:

nn = selstruc(V,'MDL')

To select the structure with the smallest loss function, use the following command:

nn = selstruc(V,0)

After estimating model orders and delays, use these values as initial guesses for
estimating other model structures, as described in “Using polyest to Estimate Polynomial
Models” on page 3-59.

Estimating Delays at the Command Line

The delayest command estimates the time delay in a dynamic system by estimating a
low-order, discrete-time ARX model and treating the delay as an unknown parameter.

By default, delayest assumes that na=nb=2 and that there is a good signal-to-noise
ratio, and uses this information to estimate nk.

3-51
3 Linear Model Identification

To estimate the delay for a data set data, type the following at the prompt:

delayest(data)

If your data has a single input, MATLAB computes a scalar value for the input delay—
equal to the number of data samples. If your data has multiple inputs, MATLAB returns
a vector, where each value is the delay for the corresponding input signal.

To compute the actual delay time, you must multiply the input delay by the sampling
interval of the data.

You can also use the ARX Model Structure Selection window to estimate input delays
and model order together, as described in “Estimating Model Orders at the Command
Line” on page 3-50.

Selecting Model Orders from the Best ARX Structure

You generate the ARX Model Structure Selection window for your data to select the best-
fit model.

For a procedure on generating this plot in the System Identification app, see “Estimating
Orders and Delays in the App” on page 3-47. To open this plot at the command line,
see “Estimating Model Orders at the Command Line” on page 3-50.

The following figure shows a sample plot in the ARX Model Structure Selection window.

You use this plot to select the best-fit model.

3-52
 Identifying Input-Output Polynomial Models

• The horizontal axis is the total number of parameters — na + nb.

• The vertical axis, called Unexplained output variance (in %), is the portion of the
output not explained by the model—the ARX model prediction error for the number of
parameters shown on the horizontal axis.

The prediction error is the sum of the squares of the differences between the
validation data output and the model one-step-ahead predicted output.
• nk is the delay.

Three rectangles are highlighted on the plot in green, blue, and red. Each color indicates
a type of best-fit criterion, as follows:

• Red — Best fit minimizes the sum of the squares of the difference between the
validation data output and the model output. This rectangle indicates the overall best
fit.
• Green — Best fit minimizes Rissanen MDL criterion.
• Blue — Best fit minimizes Akaike AIC criterion.

3-53
3 Linear Model Identification

In the ARX Model Structure Selection window, click any bar to view the orders that give
the best fit. The area on the right is dynamically updated to show the orders and delays
that give the best fit.

For more information about the AIC criterion, see “Akaike's Criteria for Model
Validation” on page 10-80.

Estimate Polynomial Models in the App

Prerequisites
Before you can perform this task, you must have:

• Imported data into the System Identification app. See “Import Time-Domain Data
into the App” on page 2-16. For supported data formats, see “Data Supported by
Polynomial Models”.
• Performed any required data preprocessing operations. To improve the accuracy of
your model, you should detrend your data. Removing offsets and trends is especially
important for Output-Error (OE) models and has less impact on the accuracy of
models that include a flexible noise model structure, such as ARMAX and Box-
Jenkins. See “Ways to Prepare Data for System Identification” on page 2-6.
• Select a model structure, model orders, and delays. For a list of available structures,
see “What Are Polynomial Models?” on page 3-40 For more information about how
to estimate model orders and delays, see “Estimating Orders and Delays in the App”
on page 3-47. For multiple-output models, you must specify order matrices in the
MATLAB workspace, as described in “Polynomial Sizes and Orders of Multi-Output
Polynomial Models” on page 3-61.

1 In the System Identification app, select Estimate > Polynomial Models to open
the Polynomial Models dialog box.

3-54
 Prerequisites

For more information on the options in the dialog box, click Help.
2 In the Structure list, select the polynomial model structure you want to estimate
from the following options:

• ARX:[na nb nk]
• ARMAX:[na nb nc nk]
• OE:[nb nf nk]
• BJ:[nb nc nd nf nk]

This action updates the options in the Polynomial Models dialog box to correspond
with this model structure. For information about each model structure, see “What
Are Polynomial Models?” on page 3-40.

3-55
3 Linear Model Identification

Note: For time-series data, only AR and ARMA models are available. For more
information about estimating time-series models, see “Time-Series Model
Identification”.
3 In the Orders field, specify the model orders and delays, as follows:

• For single-output polynomial models. Enter the model orders and delays
according to the sequence displayed in the Structure field. For multiple-input
models, specify nb and nk as row vectors with as many elements as there are
inputs. If you are estimating BJ and OE models, you must also specify nf as a
vector.

For example, for a three-input system, nb can be [1 2 4], where each element
corresponds to an input.
• For multiple-output models. Enter the model orders, as described in
“Polynomial Sizes and Orders of Multi-Output Polynomial Models” on page
3-61.

Tip To enter model orders and delays using the Order Editor dialog box, click Order
Editor.
4 (ARX models only) Select the estimation Method as ARX or IV (instrumental
variable method). For information about the algorithms, see “Polynomial Model
Estimation Algorithms” on page 3-66.
5 (ARX, ARMAX, and BJ models only) Check the Add noise integration check box to
add an integrator to the noise source, e.
6 Specify the delay using the Input delay edit box. The value must be a vector
of length equal to the number of input channels in the data. For discrete-time
estimations (any estimation using data with nonzero sample-time), the delay must
be expressed in the number of lags. These delays are separate from the “in-model”
delays specified by the nk order in the Orders edit box.
7 In the Name field, edit the name of the model or keep the default.
8 In the Focus list, select how to weigh the relative importance of the fit at different
frequencies. For more information about each option, see “Assigning Estimation
Weightings” on page 3-65.

3-56
 Prerequisites

9 In the Initial state list, specify how you want the algorithm to treat initial
conditions. For more information about the available options, see “Specifying Initial
Conditions for Iterative Estimation Algorithms” on page 3-38.

Tip If you get an inaccurate fit, try setting a specific method for handling initial
states rather than choosing it automatically.
10 In the Covariance list, select Estimate if you want the algorithm to compute
parameter uncertainties. Effects of such uncertainties are displayed on plots as
model confidence regions.

To omit estimating uncertainty, select None. Skipping uncertainty computation for

large, multiple-output models might reduce computation time.
11 Click Regularization to obtain regularized estimates of model parameters. Specify
the regularization constants in the Regularization Options dialog box. To learn more,
see “Regularized Estimates of Model Parameters”.
12 (ARMAX, OE, and BJ models only) To view the estimation progress in the MATLAB
Command Window, select the Display progress check box. This launches a
progress viewer window in which estimation progress is reported.
13 Click Estimate to add this model to the Model Board in the System Identification
app.
14 (Prediction-error method only) To stop the search and save the results after the
current iteration has been completed, click Stop Iterations. To continue iterations
from the current model, click the Continue iter button to assign current parameter
values as initial guesses for the next search.

Next Steps

• Validate the model by selecting the appropriate check box in the Model Views area
of the System Identification app. For more information about validating models, see
“Validating Models After Estimation” on page 10-2.
• Export the model to the MATLAB workspace for further analysis by dragging it to the
To Workspace rectangle in the System Identification app.

Tip For ARX and OE models, you can use the exported model for initializing a
nonlinear estimation at the command line. This initialization may improve the fit of
the model. See “Using Linear Model for Nonlinear ARX Estimation” on page 5-28,
and “Using Linear Model for Hammerstein-Wiener Estimation” on page 5-64.

3-57
3 Linear Model Identification

Estimate Polynomial Models at the Command Line

• “Using arx and iv4 to Estimate ARX Models” on page 3-58
• “Using polyest to Estimate Polynomial Models” on page 3-59

Prerequisites
Before you can perform this task, you must have

• Input-output data as an iddata object or frequency response data as an frd

or idfrd object. See “Representing Time- and Frequency-Domain Data Using
iddata Objects” on page 2-50. For supported data formats, see “Data Supported by
Polynomial Models”.
• Performed any required data preprocessing operations. To improve the accuracy of
results when using time domain data, you can detrend the data or specify the input/
output offset levels as estimation options. Removing offsets and trends is especially
important for Output-Error (OE) models and has less impact on the accuracy of
models that include a flexible noise model structure, such as ARMAX and Box-
Jenkins. See “Ways to Prepare Data for System Identification” on page 2-6.
• Select a model structure, model orders, and delays. For a list of available structures,
see “What Are Polynomial Models?” on page 3-40 For more information about
how to estimate model orders and delays, see “Estimating Model Orders at the
Command Line” on page 3-50 and “Estimating Delays at the Command Line” on
page 3-51. For multiple-output models, you must specify order matrices in the
MATLAB workspace, as described in “Polynomial Sizes and Orders of Multi-Output
Polynomial Models” on page 3-61.

Using arx and iv4 to Estimate ARX Models

You can estimate single-output and multiple-output ARX models using the arx and iv4
commands. For information about the algorithms, see “Polynomial Model Estimation
Algorithms” on page 3-66.

You can use the following general syntax to both configure and estimate ARX models:

% Using ARX method

m = arx(data,[na nb nk],opt)
% Using IV method
m = iv4(data,[na nb nk],opt)

3-58
 Prerequisites

data is the estimation data and [na nb nk] specifies the model orders, as discussed in
“What Are Polynomial Models?” on page 3-40.

The third input argument opt contains the options for configuring the estimation of the
ARX model, such as handling of initial conditions and input offsets. You can create and
configure the option set opt using the arxOptions and iv4Options commands. The
three input arguments can also be followed by name and value pairs to specify optional
model structure attributes such as InputDelay, ioDelay, and IntegrateNoise.

To get discrete-time models, use the time-domain data (iddata object).

Note: Continuous-time polynomials of ARX structure are not supported.

For more information about validating you model, see “Validating Models After
Estimation” on page 10-2.

You can use pem or polyest to refine parameter estimates of an existing polynomial
model, as described in “Refining Linear Parametric Models” on page 3-86.

For detailed information about these commands, see the corresponding reference page.

Tip You can use the estimated ARX model for initializing a nonlinear estimation at
the command line, which improves the fit of the model. See “Using Linear Model for
Nonlinear ARX Estimation” on page 5-28.

Using polyest to Estimate Polynomial Models

You can estimate any polynomial model using the iterative prediction-error estimation
method polyest. For Gaussian disturbances of unknown variance, this method gives the
maximum likelihood estimate. The resulting models are stored as idpoly model objects.

Use the following general syntax to both configure and estimate polynomial models:

m = polyest(data, [na nb nc nd nf nk], opt,Name,Value)

where data is the estimation data. na, nb, nc, nd, nf are integers that specify the model
orders, and nk specifies the input delays for each input.For more information about
model orders, see “What Are Polynomial Models?” on page 3-40.

3-59
3 Linear Model Identification

Tip You do not need to construct the model object using idpoly before estimation.

If you want to estimate the coefficients of all five polynomials, A, B, C, D, and F, you
must specify an integer order for each polynomial. However, if you want to specify an
ARMAX model for example, which includes only the A, B, and C polynomials, you must
set nd and nf to zero matrices of the appropriate size. For some simpler configurations,
there are dedicated estimation commands such as arx, armax, bj, and oe, which deliver
the required model by using just the required orders. For example, oe(data, [nb nf
nk],opt) estimates an output-error structure polynomial model.

Note: To get faster estimation of ARX models, use arx or iv4 instead of polyest.

In addition to the polynomial models listed in “What Are Polynomial Models?” on page
3-40, you can use polyest to model the ARARX structure—called the generalized
least-squares model—by setting nc=nf=0. You can also model the ARARMAX structure—
called the extended matrix model—by setting nf=0.

The third input argument, opt, contains the options for configuring the estimation of
the polynomial model, such as handling of initial conditions, input offsets and search
algorithm. You can create and configure the option set opt using the polyestOptions
command. The three input arguments can also be followed by name and value pairs
to specify optional model structure attributes such as InputDelay, ioDelay, and
IntegrateNoise.

For ARMAX, Box-Jenkins, and Output-Error models—which can only be estimated using
the iterative prediction-error method—use the armax, bj, and oe estimation commands,
respectively. These commands are versions of polyest with simplified syntax for these
specific model structures, as follows:

m = armax(Data,[na nb nc nk])
m = oe(Data,[nb nf nk])
m = bj(Data,[nb nc nd nf nk])

Similar to polyest, you can specify as input arguments the option set configured using
commands armaxOptions, oeOptions, and bjOptions for the estimators armax, oe,
and bj respectively. You can also use name and value pairs to configure additional model
structure attributes.

3-60
 Prerequisites

Tip If your data is sampled fast, it might help to apply a lowpass filter to the data
before estimating the model, or specify a frequency range for the Focus property during
estimation. For example, to model only data in the frequency range 0-10 rad/s, use the
Focus property, as follows:

opt = oeOptions('Focus',[0 10])

m = oe(Data, [nb nf nk], opt)

For more information about validating your model, see “Validating Models After
Estimation” on page 10-2.

You can use pem or polyest to refine parameter estimates of an existing polynomial
model (of any configuration), as described in “Refining Linear Parametric Models” on
page 3-86.

For more information, see polyest, pem and idpoly.

Polynomial Sizes and Orders of Multi-Output Polynomial Models

For a model with Ny (Ny > 1) outputs and Nu inputs, the polynomials A, B, C, D, and
F are specified as cell arrays of row vectors. Each entry in the cell array contains the
coefficients of a particular polynomial that relates input, output, and noise values. Orders
are matrices of integers used as input arguments to the estimation commands.

Polynomial Dimension Relation Described Orders

A Ny-by-Ny array of row vectors A{i,j} contains coefficients of na: Ny-
relation between output yi and by-Ny
output yj matrix
such that
each entry
contains
the degree
of the
corresponding
A
polynomial.
B Ny-by-Nu array of row vectors B{i,j} contain coefficients of nk: Ny-
relations between output yi and by-Nu
input uj matrix

3-61
3 Linear Model Identification

Polynomial Dimension Relation Described Orders

such that
each entry
contains
the
number
of leading
fixed zeros
of the
corresponding
B
polynomial
(input
delay).

nb: Ny-
by-Nu
matrix
such
nb(i,j)
=
length(B{i,j})-
nk(i,j).
C,D Ny-by-1 array of row vectors C{i} and D{i} contain coefficients nc and
of relations between output yi and nd are
noise ei Ny-by-1
matrices
such that
each entry
contains
the degree
of the
corresponding
C and D
polynomial,
respectively.

3-62
 Prerequisites

Polynomial Dimension Relation Described Orders

F Ny-by-Nu array of row vectors F{i,j} contains coefficients of nf: Ny-
relations between output yi and by-Nu
input uj matrix
such that
each entry
contains
the degree
of the
corresponding
F
polynomial.

For more information, see idpoly.

For example, consider the ARMAX set of equations for a 2 output, 1 input model:

y1 (t) + 0.5 y1 (t-1) + 0.9 y 2 (t-1) + 0.1 y 2 (t-2) = u(t) + 5 u(t-1) + 2 u(t-2) + e1 (t) + 0.01 e1 (t-1)
y2 (t) + 0.05 y 2 (t-1) + 0.3 y 2 (t-2) = 10 u(t-2) + e2 (t) + 0.1 e2 (t-1) + 0.02 e2 (t-2)

y1 andy2 represent the two outputs and u represents the input variable. e1 and e2
represent the white noise disturbances on the outputs, y1 and y2, respectively. To
represent these equations as an ARMAX form polynomial using idpoly, configure the A,
B, and C polynomials as follows:

A = cell(2,2);
A{1,1} = [1 0.5];
A{1,2} = [0 0.9 0.1];
A{2,1} = [0];
A{2,2} = [1 0.05 0.3];

B = cell(2,1);
B{1,1} = [1 5 2];
B{2,1} = [0 0 10];

C = cell(2,1);
C{1} = [1 0.01];
C{2} = [1 0.1 0.02];

model = idpoly(A,B,C)

3-63
3 Linear Model Identification

model =
Discrete-time ARMAX model:
Model for output number 1: A(z)y_1(t) = - A_i(z)y_i(t) + B(z)u(t) + C(z)e_1(t)
A(z) = 1 + 0.5 z^-1

A_2(z) = 0.9 z^-1 + 0.1 z^-2

B(z) = 1 + 5 z^-1 + 2 z^-2

C(z) = 1 + 0.01 z^-1

Model for output number 2: A(z)y_2(t) = B(z)u(t) + C(z)e_2(t)

A(z) = 1 + 0.05 z^-1 + 0.3 z^-2

B(z) = 10 z^-2

C(z) = 1 + 0.1 z^-1 + 0.02 z^-2

Sample time: unspecified

Parameterization:
Polynomial orders: na=[1 2;0 2] nb=[3;1] nc=[1;2]
nk=[0;2]
Number of free coefficients: 12
Use "polydata", "getpvec", "getcov" for parameters and their uncertainties.

Status:
Created by direct construction or transformation. Not estimated.

model is a discrete-time ARMAX model with unspecified sample-time. When estimating

such models, you need to specify the orders of these polynomials as input arguments.

In the System Identification app. You can enter the matrices directly in the Orders
field.

At the command line. Define variables that store the model order matrices and specify
these variables in the model-estimation command.

Tip To simplify entering large matrices orders in the System Identification app, define
the variable NN=[NA NB NK] at the command line. You can specify this variable in the
Orders field.

3-64
 Prerequisites

Assigning Estimation Weightings

You can specify how the estimation algorithm weighs the fit at various frequencies.
This information supports the estimation procedures “Estimate Polynomial Models in
the App” on page 3-54 and “Using polyest to Estimate Polynomial Models” on page
3-59.

In the System Identification app. Set Focus to one of the following options:

• Prediction — Uses the inverse of the noise model H to weigh the relative
importance of how closely to fit the data in various frequency ranges. Corresponds to
minimizing one-step-ahead prediction, which typically favors the fit over a short time
interval. Optimized for output prediction applications.
• Simulation — Uses the input spectrum to weigh the relative importance of the
fit in a specific frequency range. Does not use the noise model to weigh the relative
importance of how closely to fit the data in various frequency ranges. Optimized for
output simulation applications.
• Stability — Estimates the best stable model. For more information about model
stability, see “Unstable Models” on page 10-88.
• Filter — Specify a custom filter to open the Estimation Focus dialog box, where
you can enter a filter, as described in “Simple Passband Filter” on page 2-128 or
“Defining a Custom Filter” on page 2-129. This prefiltering applies only for estimating
the dynamics from input to output. The disturbance model is determined from the
unfiltered estimation data.

At the command line. Specify the focus as an estimation option (created using
polyestOptions, oeOptions etc.) using the same options as in the app. For example,
use this command to estimate an ARX model and emphasize the frequency content
related to the input spectrum only:
opt = arxOptions('Focus', 'simulation');
m = arx(data,[2 2 3],opt)

This Focus setting might produce more accurate simulation results, provided the orders
picked are optimal for the given data..

Specifying Initial States for Iterative Estimation Algorithms

When you use the pem or polyest to estimate ARMAX, Box-Jenkins (BJ), Output-Error
(OE), you must specify how the algorithm treats initial conditions.

3-65
3 Linear Model Identification

This information supports the estimation procedures “Estimate Polynomial Models in

the App” on page 3-54 and “Using polyest to Estimate Polynomial Models” on page
3-59.

In the System Identification app. For ARMAX, OE, and BJ models, set Initial state
to one of the following options:

• Auto — Automatically chooses Zero, Estimate, or Backcast based on the

estimation data. If initial states have negligible effect on the prediction errors, the
initial states are set to zero to optimize algorithm performance.
• Zero — Sets all initial states to zero.
• Estimate — Treats the initial states as an unknown vector of parameters and
estimates these states from the data.
• Backcast — Estimates initial states using a smoothing filter.

At the command line. Specify the initial conditions as an estimation option. Use
polyestOptions to configure options for the polyest command, armaxOptions for
the armax command etc. Set the InitialCondition option to the desired value in the
option set. For example, use this command to estimate an ARMAX model and set the
initial states to zero:
opt = armaxOptions('InitialCondition','zero')
m = armax(data,[2 2 2 3],opt)

For a complete list of values for the InitialCondition estimation option, see the
armaxOptions reference page.

Polynomial Model Estimation Algorithms

For linear ARX and AR models, you can choose between the ARX and IV algorithms.
ARX implements the least-squares estimation method that uses QR-factorization
for overdetermined linear equations. IV is the instrument variable method. For more
information about IV, see the section on variance-optimal instruments in System
Identification: Theory for the User, Second Edition, by Lennart Ljung, Prentice Hall PTR,
1999.

The ARX and IV algorithms treat noise differently. ARX assumes white noise. However,
the instrumental variable algorithm, IV, is not sensitive to noise color. Thus, use IV
when the noise in your system is not completely white and it is incorrect to assume white
noise. If the models you obtained using ARX are inaccurate, try using IV.

3-66
 Prerequisites

Note: AR models apply to time-series data, which has no input. For more information,
see “Time-Series Model Identification”. For more information about working with AR and
ARX models, see “Identifying Input-Output Polynomial Models” on page 3-40.

Estimate Models Using armax

This example shows how to estimate a linear, polynomial model with an ARMAX
structure for a three-input and single-output (MISO) system using the iterative
estimation method armax. For a summary of all available estimation commands in the
toolbox, see “Model Estimation Commands” on page 1-37.

Load a sample data set z8 with three inputs and one output, measured at 1 -second
intervals and containing 500 data samples.

load iddata8

Use armax to both construct the idpoly model object, and estimate the parameters:

Typically, you try different model orders and compare results, ultimately choosing
the simplest model that best describes the system dynamics. The following command
specifies the estimation data set, z8 , and the orders of the A , B , and C polynomials as
na , nb , and nc, respectively. nk of [0 0 0] specifies that there is no input delay for all
three input channels.

opt = armaxOptions;
opt.Focus = 'simulation';
opt.SearchOption.MaxIter = 50;
opt.SearchOption.Tolerance = 1e-5;
na = 4;
nb = [3 2 3];
nc = 4;
nk = [0 0 0];
m_armax = armax(z8, [na nb nc nk], opt);

Focus, Tolerance, and MaxIter are estimation options that configure the estimation
objective function and the attributes of the search algorithm. The Focus option specifies
whether the model is optimized for simulation or prediction applications. The Tolerance

3-67
3 Linear Model Identification

and MaxIter search options specify when to stop estimation. For more information about
these properties, see the armaxOptions reference page.

armax is a version of polyest with simplified syntax for the ARMAX model structure.
The armax method both constructs the idpoly model object and estimates its
parameters.

View information about the resulting model object.

m_armax

m_armax =
Discrete-time ARMAX model: A(z)y(t) = B(z)u(t) + C(z)e(t)
A(z) = 1 - 1.284 z^-1 + 0.3048 z^-2 + 0.2648 z^-3 - 0.05708 z^-4

B1(z) = -0.07547 + 1.087 z^-1 + 0.7166 z^-2

B2(z) = 1.019 + 0.1142 z^-1

B3(z) = -0.06739 + 0.06828 z^-1 + 0.5509 z^-2

C(z) = 1 - 0.06096 z^-1 - 0.1296 z^-2 + 0.02489 z^-3 - 0.04699 z^-4

Sample time: 1 seconds

Parameterization:
Polynomial orders: na=4 nb=[3 2 3] nc=4 nk=[0 0 0]
Number of free coefficients: 16
Use "polydata", "getpvec", "getcov" for parameters and their uncertainties.

Status:
Estimated using ARMAX on time domain data "z8".
Fit to estimation data: 80.86% (simulation focus)
FPE: 1.056, MSE: 0.9868

m_armax is an idpoly model object. The coefficients represent estimated parameters of

this polynomial model. You can use present(m_armax) to show additional information
about the model, including parameter uncertainties.

View all property values for this model.

get(m_armax)

a: [1 -1.2836 0.3048 0.2648 -0.0571]

3-68
 Prerequisites

b: {[-0.0755 1.0870 0.7166] [1.0188 0.1142] [1x3 double]}

c: [1 -0.0610 -0.1296 0.0249 -0.0470]
d: 1
f: {[1] [1] [1]}
IntegrateNoise: 0
Variable: 'z^-1'
ioDelay: [0 0 0]
Structure: [1x1 pmodel.polynomial]
NoiseVariance: 0.9899
Report: [1x1 idresults.polyest]
InputDelay: [3x1 double]
OutputDelay: 0
Ts: 1
TimeUnit: 'seconds'
InputName: {3x1 cell}
InputUnit: {3x1 cell}
InputGroup: [1x1 struct]
OutputName: {'y1'}
OutputUnit: {''}
OutputGroup: [1x1 struct]
Name: ''
Notes: {}
UserData: []
SamplingGrid: [1x1 struct]

The Report model property contains detailed information on the estimation results. To
view the properties and values inside Report, use dot notation. For example:
m_armax.Report

Status: 'Estimated using POLYEST with Focus = "simulation"'

Method: 'ARMAX'
InitialCondition: 'zero'
Fit: [1x1 struct]
Parameters: [1x1 struct]
OptionsUsed: [1x1 idoptions.polyest]
RandState: [1x1 struct]
DataUsed: [1x1 struct]
Termination: [1x1 struct]

This action displays the contents of estimation report such as model quality measures
(Fit), search termination criterion (Termination), and a record of estimation data
(DataUsed) and options (OptionsUsed).

3-69
3 Linear Model Identification

Identifying Transfer Function Models

In this section...
“What are Transfer Function Models?” on page 3-70
“Data Supported by Transfer Function Models” on page 3-71
“How to Estimate Transfer Function Models in the System Identification App” on page
3-72
“How to Estimate Transfer Function Models at the Command Line” on page 3-78
“Transfer Function Structure Specification” on page 3-79
“Estimate Transfer Function Models by Specifying Number of Poles” on page 3-80
“Estimate Transfer Function Models with Transport Delay to Fit Given Frequency
Response Data” on page 3-80
“Estimate Transfer Function Models With Prior Knowledge of Model Structure and
Constraints” on page 3-81
“Estimate Transfer Function Models with Unknown Transport Delays” on page 3-82
“Estimate Transfer Functions with Delays” on page 3-83
“Specifying Initial Conditions for Iterative Estimation Algorithms” on page 3-84

What are Transfer Function Models?

Definition of Transfer Function Models

Transfer function models describe the relationship between the inputs and outputs
of a system using a ratio of polynomials. The model order is equal to the order of the
denominator polynomial. The roots of the denominator polynomial are referred to as the
model poles. The roots of the numerator polynomial are referred to as the model zeros.

The parameters of a transfer function model are its poles, zeros and transport delays.

Continuous-Time Representation

In continuous-time, a transfer function model has the form:

num( s)
Y ( s) = U ( s) + E( s)
den(s)

3-70
 Identifying Transfer Function Models

Where, Y(s), U(s) and E(s) represent the Laplace transforms of the output, input and
noise, respectively. num(s) and den(s) represent the numerator and denominator
polynomials that define the relationship between the input and the output.

Discrete-Time Representation

In discrete-time, a transfer function model has the form:

num( q -1 )
y( t) = u(t) + e(t)
den(q -1 )
num( q -1) = b0 + b1 q -1 + b2 q-2 + º
den(q -1 ) = 1 + a1q -1 + a2 q -2 + º

The roots of num(q^-1) and den(q^-1) are expressed in terms of the lag variable q^-1.

Delays

In continuous-time, input and transport delays are of the form:

num( s) - st
Y ( s) = e U ( s) + E( s)
den(s)

Where τ represents the delay.

In discrete-time:

num
y( t) = u(t - t ) + e( t)
den

where num and den are polynomials in the lag operator q^(-1).

Data Supported by Transfer Function Models

You can estimate transfer function models from data with the following characteristics:

• Real data or complex data

• Single-output and multiple-output
• Time- or frequency-domain data

Note that you cannot use time-series data for transfer function model identification.

3-71
3 Linear Model Identification

You must first import your data into the MATLAB workspace, as described in “Data
Preparation”.

How to Estimate Transfer Function Models in the System Identification

App

Prerequisites
• Import data into the System Identification app. See “Represent Data”. For supported
data formats, see “Data Supported by Transfer Function Models” on page 3-71.
• Perform any required data preprocessing operations. If input and/or output signals
contain nonzero offsets, consider detrending your data. See “Ways to Prepare Data for
System Identification” on page 2-6.

1 In the System Identification app, select Estimate > Transfer Function Models

The Transfer Functions dialog box opens.

3-72
 Multi-Input, Multi-Output Models

Tip For more information on the options in the dialog box, click Help.
2 In the Number of poles and Number of zeros fields, specify the number of poles
and zeros of the transfer function as nonnegative integers.

Multi-Input, Multi-Output Models

For systems that are multiple input, multiple output, or both:

• To use the same number of poles or zeros for all the input/output pairs, specify a
scalar.
• To use a different number of poles and zeros for the input/output pairs, specify an
ny-by-nu matrix. ny is the number of outputs and nu is the number of inputs.

Alternatively, click .

The Model Orders dialog box opens where you specify the number of poles and
zeros for each input/output pair. Use the Output list to select an output.

3-73
3 Linear Model Identification

3 Select Continuous-time or Discrete-time to specify whether the model is a

continuous- or discrete-time transfer function.

For discrete-time models, the number of poles and zeros refers to the roots of the
numerator and denominator polynomials expressed in terms of the lag variable
q^-1.
4 (For discrete-time models only) Specify whether to estimate the model feedthrough.
Select the Feedthrough check box.

A discrete-time model with 2 poles and 3 zeros takes the following form:

b0 + b1 z-1 + b2 z-2 + b 3 z-3

Hz-1 =
1 + a1 z -1 + a2 z-2

When the model has direct feedthrough, b0 is a free parameter whose value is
estimated along with the rest of the model parameters b1, b2, b3, a1, a2. When the
model has no feedthrough, b0 is fixed to zero.

Multi-Input, Multi-Output Models

For models that are multi input, multi output or both, click Feedthrough.

3-74
 Multi-Input, Multi-Output Models

The Model Orders dialog box opens, where you specify to estimate the feedthrough
for each input/output pair separately. Use the Output list to select an output.

5 Expand the I/O Delay section to specify nominal values and constraints for
transport delays for different input/output pairs.

Use the Output list to select an output. Select the Fixed check box to specify a
transport delay as a fixed value. Specify its nominal value in the Delay field.
6 Expand the Estimation Options section to specify estimation options.

3-75
3 Linear Model Identification

• Set the range slider to the desired passband to specify the frequency range
over which the transfer function model must fit the data. By default the entire
frequency range (0 to Nyquist frequency) is covered.
• Select Display progress to view the progress of the optimization.
• Select Estimate covariance to estimate the covariance of the transfer function
parameters.
• (For frequency-domain data only) Specify whether to allow the estimation process
to use parameter values that may lead to unstable models. Select the Allow
unstable models option.

Setting this option is same as setting the estimation option Focus to

'prediction' at the command line. An unstable model is delivered only if it
produces a better fit to the data than other stable models computed during the
estimation process.
• Specify how to treat the initial conditions in the Initial condition list. For
more information, see “Specifying Initial Conditions for Iterative Estimation
Algorithms” on page 3-84.
• Specify the algorithm used to initialize the values of the numerator and
denominator coefficients in the Initialization method list.

• IV — Instrument Variable approach.

• SVF — State Variable Filters approach.

3-76
 Iteration Options

• N4SID — Generalized Poisson Moment Functions approach.

• GPMF — Subspace state-space estimation approach.
• All — Combination of all of the above approaches. The software tries all
the above methods and selects the method that yields the smallest value of
prediction error norm.
7 Click Regularization to obtain regularized estimates of model parameters. Specify
the regularization constants in the Regularization Options dialog box. To learn more,
see “Regularized Estimates of Model Parameters”.
8 Click Iterations Options to specify options for controlling the iterations. The
Options for Iterative Minimization dialog box opens.

Iteration Options
In the Options for Iterative Minimization dialog box, you can specify the following
iteration options:

3-77
3 Linear Model Identification

• Search Method — Method used by the iterative search algorithm. Search

method is auto by default. The descent direction is calculated using gn (Gauss-
Newton), gna (Adaptive Gauss-Newton), lm (Levenberg-Marquardt), lsqnonlin
(Trust-Region Reflective Newton), and grad (Gradient Search) successively at
each iteration until a sufficient reduction in error is achieved.
• Output weighting — Weighting applied to the loss function to be minimized.
Use this option for multi-output estimations only. Specify as 'noise' or a
positive semidefinite matrix of size equal the number of outputs.
• Maximum number of iterations — Maximum number of iterations to use
during search.
• Termination tolerance — Tolerance value when the iterations should
terminate.
• Error threshold for outlier penalty — Robustification of the quadratic
criterion of fit.
9 Click Estimate to estimate the model. A new model gets added to the System
Identification app.

Next Steps
• Validate the model by selecting the appropriate check box in the Model Views area
of the System Identification app. For more information about validating models, see
“Validating Models After Estimation” on page 10-2.
• Export the model to the MATLAB workspace for further analysis. Drag the model to
the To Workspace rectangle in the System Identification app.

How to Estimate Transfer Function Models at the Command Line

Before you estimate a transfer function model, you must have:

• Input/Output data. See “Representing Time- and Frequency-Domain Data Using

iddata Objects” on page 2-50. For supported data formats, see “Data Supported by
Transfer Function Models” on page 3-71.
• Performed any required data preprocessing operations. You can detrend your data
before estimation. For more information, see “Ways to Prepare Data for System
Identification” on page 2-6.

3-78
 Next Steps

Alternatively, you can specify the input/output offset for the data using an estimation
option set. Use tfestOptions to create the estimation option set. Use the
InputOffset and OutputOffset name and value pairs to specify the input/output
offset.

Estimate continuous-time and discrete-time transfer function models using tfest. The
output of tfest is an idtf object, which represents the identified transfer function.

The general workflow in estimating a transfer function model is:

1 Create a data object (iddata or idfrd) that captures the experimental data.
2 (Optional) Specify estimation options using tfestOptions.
3 (Optional) Create a transfer function model that specifies the expected model
structure and any constraints on the estimation parameters.
4 Use tfest to identify the transfer function model, based on the data.
5 Validate the model. See “Model Validation”.

Transfer Function Structure Specification

You can use a priori knowledge of the expected transfer function model structure to
initialize the estimation. The Structure property of an idtf model contains parameters
that allow you to specify the values and constraints for the numerator, denominator and
transport delays.

For example, specify a third-order transfer function model that contains an integrator
and has a transport delay of at most 1.5 seconds:

init_sys = idtf([nan nan],[1 2 1 0]);

init_sys.Structure.ioDelay.Maximum = 1.5;
init_sys.Structure.den.Free(end) = false;

int_sys is an idtf model with three poles and one zero. The denominator coefficient for
the s^0 term is zero and implies that one of the poles is an integrator.

init_sys.Structure.ioDelay.Maximum = 1.5 constrains the transport delay to

a maximum of 1.5 seconds. The last element of the denominator coefficients (associated
with the s^0 term) is not a free estimation variable. This constraint forces one of the
estimated poles to be at s = 0.

3-79
3 Linear Model Identification

For more information regarding configuring the initial parameterization of an estimated

transfer function, see Structure in idtf.

Estimate Transfer Function Models by Specifying Number of Poles

This example shows how to identify a transfer function containing a specified number of
poles for given data.

Load time-domain system response data and use it to estimate a transfer function for the
system.

load iddata1 z1;

np = 2;
sys = tfest(z1,np);

z1 is an iddata object that contains time-domain, input-output data.

np specifies the number of poles in the estimated transfer function.

sys is an idtf model containing the estimated transfer function.

To see the numerator and denominator coefficients of the resulting estimated model sys,
enter:

sys.num
sys.den

To view the uncertainty in the estimates of the numerator and denominator and other
information, use tfdata.

Estimate Transfer Function Models with Transport Delay to Fit Given

Frequency Response Data
This example shows how to identify a transfer function to fit a given frequency response
data (FRD) containing additional phase roll off induced by input delay.

Obtain frequency response data.

For this example, use bode to obtain the magnitude and phase response data for the
following system:

3-80
 Next Steps

s + 0 .2
H ( s) = e-.5s
s + 2s2 + s + 1
3

Use 100 frequency points, ranging from 0.1 rad/s to 10 rad/s, to obtain the frequency
response data. Use frd to create a frequency response data object.
freq = logspace(-1,1,100);
[mag, phase] = bode(tf([1 .2],[1 2 1 1],'InputDelay',.5),freq);
data = frd(mag.*exp(1j*phase*pi/180),freq);

data is an iddata object that contains frequency response data for the described system.

Estimate a transfer function using data. Specify an unknown transport delay for the
identified transfer function.
np = 3;
nz = 1;
iodelay = NaN;
sys = tfest(data,np,nz,iodelay)

np and nz specify the number of poles and zeros in the identified transfer function,
respectively.

iodelay specifies an unknown transport delay for the identified transfer function.

sys is an idtf model containing the identified transfer function.

Estimate Transfer Function Models With Prior Knowledge of Model

Structure and Constraints
This example shows how to estimate a transfer function for given data when the
structure of the expected model is known. This example also shows how to apply
constraints to the numerator and denominator coefficients.

Load time domain data.

load iddata1 z1;
z1.y = cumsum(z1.y);

cumsum integrates the output data of z1. The estimated transfer function should
therefore contain an integrator.

Create a transfer function model with the expected structure.

3-81
3 Linear Model Identification

init_sys = idtf([100 1500],[1 10 10 0]);

int_sys is an idtf model with three poles and one zero. The denominator coefficient for
the s^0 term is zero. Therefore, int_sys contains an integrator.

Specify constraints on the numerator and denominator coefficients of the transfer

function model. To do so, configure fields in the Structure property:
init_sys.Structure.num.Minimum = eps;
init_sys.Structure.den.Minimum = eps;
init_sys.Structure.den.Free(end) = false;

The constraints specify that the numerator and denominator coefficients are
nonnegative. Additionally, the last element of the denominator coefficients (associated
with the s^0 term) is not an estimable parameter. This constraint forces one of the
estimated poles to be at s = 0.

Create an estimation option set that specifies using the Levenberg–Marquardt search
method.
opt = tfestOptions('SearchMethod', 'lm');

Estimate a transfer function for z1 using init_sys and the estimation option set.
sys = tfest(z1,init_sys,opt);

tfest uses the coefficients of init_sys to initialize the estimation of sys. Additionally,
the estimation is constrained by the constraints you specify in the Structure property
of init_sys. The resulting idtf model sys contains the parameter values that result
from the estimation.

Estimate Transfer Function Models with Unknown Transport Delays

This example shows how to estimate a transfer function for given data with unknown
transport delays. This example also shows how to apply an upper bound on the unknown
transport delays.

Create a transfer function model with the expected numerator and denominator
structure and delay constraints.

For this example, the experiment data consists of two inputs and one output. Both
transport delays are unknown and have an identical upper bound. Additionally, the
transfer functions from both inputs to the output are identical in structure.

3-82
 Next Steps

init_sys = idtf(NaN(1,2),[1, NaN(1,3)],'ioDelay',NaN);

init_sys.Structure(1).ioDelay.Free = true;
init_sys.Structure(1).ioDelay.Maximum = 7;

init_sys is an idtf model describing the structure of the transfer function from
one input to the output. The transfer function consists of one zero, three poles and a
transport delay. The use of NaN indicates unknown coefficients.

init_sys.Structure(1).ioDelay.Free = true indicates that the transport delay

is not fixed.

init_sys.Structure(1).ioDelay.Maximum = 7 sets the upper bound for the

transport delay to 7 seconds.

Specify the transfer function from both inputs to the output.

init_sys = [init_sys, init_sys];

Load time domain system response data and detrend .

load co2data;
Ts = 0.5;
data = iddata(Output_exp1,Input_exp1,Ts);
T = getTrend(data);
T.InputOffset = [170, 50];
T.OutputOffset = mean(data.y(1:75));
data = detrend(data, T);

data is an iddata object and has a sample rate of 0.5 seconds.

Identify a transfer function for the measured data using the specified delay constraints.

sys = tfest(data,init_sys);

sys is an idtf model containing the identified transfer function.

Estimate Transfer Functions with Delays

This example shows how to estimate transfer function models with I/O delays.

The tfest command supports estimation of IO delays. In the simplest case, if you
specify NaN as the value for the ioDelay input argument, tfest estimates the
corresponding delay value.

3-83
3 Linear Model Identification

load iddata1 z1
sys = tfest(z1, 2, 2, NaN); % 2 poles, 2 zeros, unknown transport delay

If you want to assign an initial guess to the value of delay or prescribe bounds for its
value, you must first create a template idtf model and configure ioDelay using the
model's Structure property:
sys0 = idtf([nan nan nan],[1 nan nan]);
sys0.Structure.ioDelay.Value = 0.1; % initial guess
sys0.Structure.ioDelay.Maximum = 1; % maximum allowable value for delay
sys0.Structure.ioDelay.Free = true; % treat delay as estimatable quantity
sys = tfest(z1, sys0);

If estimation data is in the time-domain, the delays are not estimated iteratively. If a
finite initial value is specified, that value is retained as is with no iterative updates. The
same is true of discrete-time frequency domain data. Thus in the example above, if data
has a nonzero sample time, the estimated value of delay in the returned model sys is 0.1
(same as the initial guess specified for sys0 ). The delays are updated iteratively only
for continuos-time frequency domain data. If, on the other hand, a finite initial value for
delay is not specified (e.g., sys0.Structure.ioDelay.Value = NaN ), then a value for
delay is determined using the delayest function, regardless of the nature of the data.

Determination of delay as a quantity independent of the model's poles and zeros is

a difficult task. Estimation of delays becomes especially difficult for multi-input or
multi-output data. It is strongly recommended that you perform some investigation to
determine delays before estimation. You can use functions such as delayest, arxstruc,
selstruc and impulse response analysis to determine delays. Often, physical knowledge
of the system or dedicated transient tests (how long does it take for a step change in
input to show up in a measured output?) will reveal the value of transport delays. Use
the results of such analysis to assign initial guesses as well as minimum and maximum
bounds on the estimated values of delays.

Specifying Initial Conditions for Iterative Estimation Algorithms

If you estimate transfer function models using tfest, you can specify how the algorithm
treats initial conditions.

In the System Identification app, set Initial condition to one of the following options:

• auto — Automatically chooses Zero, Estimate, or Backcast based on the

estimation data. If initial conditions have negligible effect on the prediction errors, the
initial conditions are set to zero to optimize algorithm performance.

3-84
 Next Steps

• Zero — Sets all initial conditions to zero.

• Estimate — Treats the initial conditions as an estimation parameters.
• Backcast — Estimates initial conditions using a backward filtering method (least-
squares fit).

At the command line. Specify the initial conditions by using an estimation option set.
Use tfestOptions to create the estimation option set. For example, create an options
set that sets the initial conditions to zero:
opt = tfestOptions('InitialCondition','zero);

For more information, see tfestOptions.

3-85
3 Linear Model Identification

Refining Linear Parametric Models

In this section...
“When to Refine Models” on page 3-86
“What You Specify to Refine a Model” on page 3-86
“Refine Linear Parametric Models Using System Identification App” on page 3-86
“Refine Linear Parametric Models at the Command Line” on page 3-88

When to Refine Models

There are two situations where you can refine estimates of linear parametric models.

In the first situation, you have already estimated a parametric model and wish to update
the values of its free parameters to improve the fit to the estimation data. This is useful
if your previous estimation terminated because of search algorithm constraints such as
maximum number of iterations or function evaluations allowed reached. However, if your
model captures the essential dynamics, it is usually not necessary to continue improving
the fit—especially when the improvement is a fraction of a percent.

In the second situation, you might have constructed a model using one of the model
constructors described in “Commands for Constructing Model Structures” on page 1-22.
In this case, you built initial parameter guesses into the model structure and wish to
refine these parameter values.

What You Specify to Refine a Model

When you refine a model, you must provide two inputs:

• Parametric model
• Data — You can either use the same data set for refining the model as the one you
originally used to estimate the model, or you can use a different data set.

Refine Linear Parametric Models Using System Identification App

The following procedure assumes that the model you want to refine is already in the
System Identification app. You might have estimated this model in the current session

3-86
 Refining Linear Parametric Models

or imported the model from the MATLAB workspace. For information about importing
models into the app, see “Importing Models into the App” on page 14-7.

To refine your model:

1 In the System Identification app, verify that you have the correct data set in the
Working Data area for refining your model.

If you are using a different data set than the one you used to estimate the model,
drag the correct data set into the Working Data area. For more information about
specifying estimation data, see “Specify Estimation and Validation Data in the App”
on page 2-30.
2 Select Estimate > Refine Existing Models to open the Linear Model Refinement
dialog box.

For more information on the options in the dialog box, click Help.
3 Select the model you want to refine in the Initial Model drop-down list or type
the model name.

3-87
3 Linear Model Identification

The model name must be in the Model Board of the System Identification app or a
variable in the MATLAB workspace. The model can be a state-space, polynomial,
process, transfer function or linear grey-box model. The input-output dimensions of
the model must match that of the working data.
4 (Optional) Modify the Estimation Options.

When you enter the model name, the estimation options in the Linear Model
Refinement dialog box override the initial model settings.
5 Click Regularization to obtain regularized estimates of model parameters. Specify
the regularization constants in the Regularization Options dialog box. To learn more,
see “Regularized Estimates of Model Parameters”.
6 Click Estimate to refine the model.
7 Validate the new model. See “Ways to Validate Models” on page 10-2.

Refine Linear Parametric Models at the Command Line

If you are working at the command line, you can use pem to refine parametric model
estimates. You can also use the various model-structure specific estimators — ssest for
idss models, polyest for idpoly models, tfest for idtf models, and greyest for
idgrey models.

The general syntax for refining initial models is as follows:

m = pem(data,init_model)

pem uses the properties of the initial model.

You can also specify the estimation options configuring the objective function and search
algorithm settings. For more information, see the reference page of the estimating
function.

3-88
 Refine ARMAX Model with Initial Parameter Guesses at Command Line

Refine ARMAX Model with Initial Parameter Guesses at Command

Line
This example shows how to refine models for which you have initial parameter guesses.

Estimate an ARMAX model for the data by initializing the A , B , and C polynomials.
You must first create a model object and set the initial parameter values in the model
properties. Next, you provide this initial model as input to armax , polyest , or pem ,
which refine the initial parameter guesses using the data.

Load estimation data.

load iddata8

Define model parameters.

Leading zeros in B indicate input delay (nk), which is 1 for each input channel.

A = [1 -1.2 0.7];
B{1} = [0 1 0.5 0.1]; % first input
B{2} = [0 1.5 -0.5]; % second input
B{3} = [0 -0.1 0.5 -0.1]; % third input
C = [1 0 0 0 0];
Ts = 1;

Create model object.

init_model = idpoly(A,B,C,'Ts',1);

Use polyest to update the parameters of the initial model.

model = polyest(z8,init_model);

Compare the two models.

compare(z8,init_model,model)

3-89
3 Linear Model Identification

Related Examples
• “Identifying Input-Output Polynomial Models” on page 3-40

3-90
 Refine Initial ARMAX Model at Command Line

Refine Initial ARMAX Model at Command Line

This example shows how to estimate an initial model and refine it using pem.

Load measured data.

load iddata8;

Split the data into an initial estimation data set and a refinement data set.

init_data = z8(1:100);
refine_data = z8(101:end);

init_data is an iddata object containing the first 100 samples from z8 and
refine_data is an iddata object representing the remaining data in z8.

Estimate an ARMAX model.

na=4;
nb=[3 2 3];
nc=2;
nk=[0 0 0];

sys = armax(init_data,[na nb nc nk]);

armax uses the default algorithm properties to estimate sys.

Refine the estimated model by specifying the estimation algorithm options. Specify
stricter tolerance and increase the maximum iterations.

opt = armaxOptions;
opt.SearchOption.Tolerance = 1e-5;
opt.SearchOption.MaxIter = 50;

refine_sys = pem(refine_data,sys,opt);

Compare the fit of the initial and refined models.

compare(refine_data,sys,refine_sys);

3-91
3 Linear Model Identification

refine_sys provides a closer fit to the data than sys.

You can similarly use polyest or armax to refine the estimated model.

See Also
Functions
armax | pem | polyest

3-92
 Extracting Numerical Model Data

Extracting Numerical Model Data

You can extract the following numerical data from linear model objects:

• Coefficients and uncertainty

For example, extract state-space matrices (A, B, C, D and K) for state-space models, or
polynomials (a, b, c, d and f) for polynomial models.

If you estimated model uncertainty data, this information is stored in the model in the
form of the parameter covariance matrix. You can fetch the covariance matrix (in its
raw or factored form) using the getcov command. The covariance matrix represents
uncertainties in parameter estimates and is used to compute:

• Confidence bounds on model output plots, Bode plots, residual plots, and pole-zero
plots
• Standard deviation in individual parameter values. For example, one standard
deviation in the estimated value of the A polynomial in an ARX model, returned by
the polydata command and displayed by the present command.

The following table summarizes the commands for extracting model coefficients and
uncertainty.

Commands for Extracting Model Coefficients and Uncertainty Data

Command Description Syntax

freqresp Extracts frequency- [H,w,CovH] = freqresp(m)
response data (H) and
corresponding covariance
(CovH) from any linear
identified model.
polydata Extracts polynomials [A,B,C,D,F,dA,dB,dC,dD,dF] = ...
(such as A) from any polydata(m)
linear identified
model. The polynomial
uncertainties (such as
dA) are returned only for
idpoly models.

3-93
3 Linear Model Identification

Command Description Syntax

idssdata Extracts state-space [A,B,C,D,K,X0,...
matrices (such as A) from dA,dB,dC,dD,dK,dX0] = ...
any linear identified idssdata(m)
model. The matrix
uncertainties (such as
dA) are returned only for
idss models.
tfdata Extracts numerator and [Num,Den,Ts,dNum,dDen] = ...
denominator polynomials tfdata(m)
(num, den) and their
uncertainties (dnum,
dden) from any linear
identified model.
zpkdata Extracts zeros, poles, and [Z,P,K,Ts,covZ,covP,covK] = ...
gains (Z, P, K) and their zpkdata(m)
covariances (covZ, covP,
covK) from any linear
identified model.
getpvec Obtain a list of model pvec = getpvec(m)
parameters and their
uncertainties.
To access parameter
attributes such as values,
free status, bounds or
labels, use getpar.
getcov Obtain parameter cov_data = getcov(m)
covariance information

You can also extract numerical model data by using dot notation to access model
properties. For example, m.A displays the A polynomial coefficients from model m.
Alternatively, you can use the get command, as follows: get(m,'A').

Tip To view a list of model properties, type get(model).

• Dynamic and noise models

3-94
 Extracting Numerical Model Data

For linear models, the general symbolic model description is given by:

y = Gu + He

G is an operator that takes the measured inputs u to the outputs and captures the
system dynamics, also called the measured model. H is an operator that describes
the properties of the additive output disturbance and takes the hypothetical
(unmeasured) noise source inputs e to the outputs, also called the noise model. When
you estimate a noise model, the toolbox includes one noise channel e for each output in
your system.

You can operate on extracted model data as you would on any other MATLAB vectors,
matrices and cell arrays. You can also pass these numerical values to Control System
Toolbox commands, for example, or Simulink blocks.

3-95
3 Linear Model Identification

Transforming Between Discrete-Time and Continuous-Time

Representations
In this section...
“Why Transform Between Continuous and Discrete Time?” on page 3-96
“Using the c2d, d2c, and d2d Commands” on page 3-96
“Specifying Intersample Behavior” on page 3-98
“Effects on the Noise Model” on page 3-98

Why Transform Between Continuous and Discrete Time?

Transforming between continuous-time and discrete-time representations is useful, for
example, if you have estimated a discrete-time linear model and require a continuous-
time model instead for your application.

You can use c2d and d2c to transform any linear identified model between continuous-
time and discrete-time representations. d2d is useful is you want to change the sampling
interval of a discrete-time model. All of these operations change the sampling interval,
which is called resampling the model.

These commands do not transform the estimated model uncertainty. If you want to
translate the estimated parameter covariance during the conversion, use translatecov.

Note: c2d and d2d correctly approximate the transformation of the noise model only
when the sampling interval T is small compared to the bandwidth of the noise.

Using the c2d, d2c, and d2d Commands

The following table summarizes the commands for transforming between continuous-
time and discrete-time model representations.

Command Description Usage Example

c2d Converts continuous-time To transform a continuous-time model
models to discrete-time mod_c to a discrete-time form, use the
models. following command:

3-96
 Transforming Between Discrete-Time and Continuous-Time Representations

Command Description Usage Example

You cannot use c2d for mod_d = c2d(mod_c,T)
idproc models and for
idgrey models whose where T is the sampling interval of
FcnType is not 'cd'. the discrete-time model.
Convert these models into
idpoly, idtf, or idss
models before calling c2d.

d2c Converts parametric To transform a discrete-time model

discrete-time models to mod_d to a continuous-time form, use
continuous-time models. the following command:

You cannot use d2c for mod_c = d2c(mod_d)

idgrey models whose
FcnType is not 'cd'.
Convert these models into
idpoly, idtf, or idss
models before calling d2c.
d2d Resample a linear discrete- To resample a discrete-time model
time model and produce an mod_d1 to a discrete-time form with
equivalent discrete-time a new sampling interval Ts, use the
model with a new sampling following command:
interval.
You can use the resampled mod_d2 = d2d(mod_d1,Ts)
model to simulate or predict
output with a specified time
interval.

The following commands compare estimated model m and its continuous-time counterpart
mc on a Bode plot:

% Estimate discrete-time ARMAX model

% from the data
m = armax(data,[2 3 1 2]);
% Convert to continuous-time form
mc = d2c(m);
% Plot bode plot for both models
bode(m,mc)

3-97
3 Linear Model Identification

Specifying Intersample Behavior

A sampled signal is characterized only by its values at the sampling instants. However,
when you apply a continuous-time input to a continuous-time system, the output values
at the sampling instants depend on the inputs at the sampling instants and on the
inputs between these points. Thus, the InterSample data property describes how the
algorithms should handle the input between samples. For example, you can specify the
behavior between the samples to be piece-wise constant (zero-order hold, zoh) or linearly
interpolated between the samples (first order hold, foh). The transformation formulas for
c2d and d2c are affected by the intersample behavior of the input.

By default, c2d and d2c use the intersample behavior you assigned to the estimation
data. To override this setting during transformation, add an extra argument in the
syntax. For example:

% Set first-order hold intersample behavior

mod_d = c2d(mod_c,T,'foh')

Effects on the Noise Model

c2d, d2c, and d2d change the sampling interval of both the dynamic model and the noise
model. Resampling a model affects the variance of its noise model.

A parametric noise model is a time-series model with the following mathematical

description:

y( t) = H ( q) e(t)
Ee2 = l

The noise spectrum is computed by the following discrete-time equation:

2
F v (w) = lT H eiw T ( )
where l is the variance of the white noise e(t), and lT represents the spectral density of
e(t). Resampling the noise model preserves the spectral density l T . The spectral density
l T is invariant up to the Nyquist frequency. For more information about spectrum
normalization, see “Spectrum Normalization” on page 3-13.

3-98
 Transforming Between Discrete-Time and Continuous-Time Representations

d2d resampling of the noise model affects simulations with noise using sim. If you
resample a model to a faster sampling rate, simulating this model results in higher noise
level. This higher noise level results from the underlying continuous-time model being
subject to continuous-time white noise disturbances, which have infinite, instantaneous
variance. In this case, the underlying continuous-time model is the unique representation
for discrete-time models. To maintain the same level of noise after interpolating the noise

signal, scale the noise spectrum by TNew

TOld , where Tnew is the new sampling interval
and Told is the original sampling interval. before applying sim.

3-99
3 Linear Model Identification

Continuous-Discrete Conversion Methods

In this section...
“Choosing a Conversion Method” on page 3-100
“Zero-Order Hold” on page 3-101
“First-Order Hold” on page 3-102
“Impulse-Invariant Mapping” on page 3-103
“Tustin Approximation” on page 3-104
“Zero-Pole Matching Equivalents” on page 3-108

Choosing a Conversion Method

The c2d command discretizes continuous-time models. Conversely, d2c converts
discrete-time models to continuous time. Both commands support several discretization
and interpolation methods, as shown in the following table.

Discretization Method Use when:

“Zero-Order Hold” on page 3-101 You want an exact discretization in the
time domain for staircase inputs.
“First-Order Hold” on page 3-102 You want an exact discretization in the
time domain for piecewise linear inputs.
“Impulse-Invariant Mapping” on page You want an exact discretization in the
3-103 (c2d only) time domain for impulse train inputs.
“Tustin Approximation” on page 3-104 • You want good matching in the
frequency domain between the
continuous- and discrete-time models.
• Your model has important dynamics at
some particular frequency.
“Zero-Pole Matching Equivalents” on page You have a SISO model, and you want good
3-108 matching in the frequency domain between
the continuous- and discrete-time models.

3-100
 Continuous-Discrete Conversion Methods

Zero-Order Hold
The Zero-Order Hold (ZOH) method provides an exact match between the continuous-
and discrete-time systems in the time domain for staircase inputs.

The following block diagram illustrates the zero-order-hold discretization Hd(z) of a

continuous-time linear model H(s)

The ZOH block generates the continuous-time input signal u(t) by holding each sample
value u(k) constant over one sample period:

u ( t ) = u [ k] , kTs £ t £ ( k + 1 ) Ts

The signal u(t) is the input to the continuous system H(s). The output y[k] results from
sampling y(t) every Ts seconds.

Conversely, given a discrete system Hd(z), d2c produces a continuous system H(s). The
ZOH discretization of H(s) coincides with Hd(z).

The ZOH discrete-to-continuous conversion has the following limitations:

• d2c cannot convert LTI models with poles at z = 0.

• For discrete-time LTI models having negative real poles, ZOH d2c conversion
produces a continuous system with higher order. The model order increases because
a negative real pole in the z domain maps to a pure imaginary value in the s domain.
Such mapping results in a continuous-time model with complex data. To avoid this,
the software instead introduces a conjugate pair of complex poles in the s domain.

ZOH Method for Systems with Time Delays

You can use the ZOH method to discretize SISO or MIMO continuous-time models with
time delays. The ZOH method yields an exact discretization for systems with input
delays, output delays, or transfer delays.

3-101
3 Linear Model Identification

For systems with internal delays (delays in feedback loops), the ZOH method results in
approximate discretizations. The following figure illustrates a system with an internal
delay.

H(s)

e-ts

For such systems, c2d performs the following actions to compute an approximate ZOH
discretization:

1 Decomposes the delay τ as t = kTs + r with 0 £ r < Ts .

2 Absorbs the fractional delay r into H(s).

3 Discretizes H(s) to H(z).

4 Represents the integer portion of the delay kTs as an internal discrete-time delay z–k.
The final discretized model appears in the following figure:

H(z)

H(s) e-sr

z-k

First-Order Hold
The First-Order Hold (FOH) method provides an exact match between the continuous-
and discrete-time systems in the time domain for piecewise linear inputs.

FOH differs from ZOH by the underlying hold mechanism. To turn the input samples
u[k] into a continuous input u(t), FOH uses linear interpolation between samples:

3-102
 Continuous-Discrete Conversion Methods

t - kTs
u ( t ) = u [ k] +
Ts
(u [k + 1] - u [k]) , kTs £ t £ (k + 1) Ts

This method is generally more accurate than ZOH for systems driven by smooth inputs.

This FOH method differs from standard causal FOH and is more appropriately called
triangle approximation (see [2], p. 228). The method is also known as ramp-invariant
approximation.

FOH Method for Systems with Time Delays

You can use the FOH method to discretize SISO or MIMO continuous-time models with
time delays. The FOH method handles time delays in the same way as the ZOH method.
See “ZOH Method for Systems with Time Delays” on page 3-101.

Impulse-Invariant Mapping
The impulse-invariant mapping produces a discrete-time model with the same impulse
response as the continuous time system. For example, compare the impulse response of a
first-order continuous system with the impulse-invariant discretization:

G = tf(1,[1,1]);
Gd1 = c2d(G,0.01,'impulse');
impulse(G,Gd1)

3-103
3 Linear Model Identification

The impulse response plot shows that the impulse responses of the continuous and
discretized systems match.

Impulse-Invariant Mapping for Systems with Time Delays

You can use impulse-invariant mapping to discretize SISO or MIMO continuous-time

models with time delay, except that the method does not support ss models with internal
delays. For supported models, impulse-invariant mapping yields an exact discretization
of the time delay.

Tustin Approximation
The Tustin or bilinear approximation yields the best frequency-domain match between
the continuous-time and discretized systems. This method relates the s-domain and z-
domain transfer functions using the approximation:

3-104
 Continuous-Discrete Conversion Methods

sTs 1 + sTs / 2
z=e ª .
1 - sTs / 2

In c2d conversions, the discretization Hd(z) of a continuous transfer function H(s) is:

2 z-1
H d ( z ) = H ( s¢ ) , s¢ =
Ts z + 1

Similarly, the d2c conversion relies on the inverse correspondence

1 + sTs / 2
H ( s ) = Hd ( z¢ ) , z¢ =
1 - sTs / 2

When you convert a state-space model using the Tustin method, the states are not
preserved. The state transformation depends upon the state-space matrices and whether
the system has time delays. For example, for an explicit (E = I) continuous-time model
with no time delays, the state vector w[k] of the discretized model is related to the
continuous-time state vector x(t) by:

Ê T ˆ T T
w [ kTs ] = Á I - A s ˜ x ( kTs ) - s Bu ( kTs ) = x ( kTs ) - s ( Ax ( kTs ) + Bu ( kTs ) ) .
Ë 2 ¯ 2 2

Ts is the sampling time of the discrete-time model. A and B are state-space matrices of
the continuous-time model.

Tustin Approximation with Frequency Prewarping

If your system has important dynamics at a particular frequency that you want the
transformation to preserve, you can use the Tustin method with frequency prewarping.
This method ensures a match between the continuous- and discrete-time responses at the
prewarp frequency.

The Tustin approximation with frequency prewarping uses the following transformation
of variables:

w z-1
H d ( z ) = H ( s¢ ) , s¢ =
tan (w Ts / 2 ) z + 1

3-105
3 Linear Model Identification

This change of variable ensures the matching of the continuous- and discrete-
time frequency responses at the prewarp frequency ω, because of the following
correspondence:

(
H ( jw ) = H d e jw Ts )
Tustin Approximation for Systems with Time Delays

You can use the Tustin approximation to discretize SISO or MIMO continuous-time
models with time delays.

By default, the Tustin method rounds any time delay to the nearest multiple of the
sampling time. Therefore, for any time delay tau, the integer portion of the delay, k*Ts,
maps to a delay of k sampling periods in the discretized model. This approach ignores the
residual fractional delay, tau - k*Ts.

You can to approximate the fractional portion of the delay by a discrete all-pass filter
(Thiran filter) of specified order. To do so, use the FractDelayApproxOrder option of
c2dOptions.

To understand how the Tustin method handles systems with time delays, consider the
following SISO state-space model G(s). The model has input delay τi, output delay τo, and
internal delay τ.

G(s)

e-tis e-tos
H(s)

e-ts

The following figure shows the general result of discretizing G(s) using the Tustin
method.

3-106
 Continuous-Discrete Conversion Methods

Gd(z)

z-mi Fi(z) Fo(z) z-mo

Hd(z)

z-m F(z)

By default, c2d converts the time delays to pure integer time delays. The c2d command
computes the integer delays by rounding each time delay to the nearest multiple of
the sample time Ts. Thus, in the default case, mi = round(τi /Ts), mo = round(τo/Ts), and
m = round(τ/Ts).. Also in this case, Fi(z) = Fo(z) = F(z) = 1.

If you set FractDelayApproxOrder to a non-zero value, c2d approximates the

fractional portion of the time delays by Thiran filters Fi(z), Fo(z), and F(z).

The Thiran filters add additional states to the model. The maximum number of
additional states for each delay is FractDelayApproxOrder.

For example, for the input delay τi, the order of the Thiran filter Fi(z) is:
order(Fi(z)) = max(ceil(τi /Ts), FractDelayApproxOrder).

If ceil(τi /Ts) < FractDelayApproxOrder, the Thiran filter Fi(z) approximates the
entire input delay τi. If ceil(τi/Ts) > FractDelayApproxOrder, the Thiran filter only
approximates a portion of the input delay. In that case, c2d represents the remainder of
the input delay as a chain of unit delays z–mi, where
mi = ceil(τi /Ts) – FractDelayApproxOrder.

c2d uses Thiran filters and FractDelayApproxOrder in a similar way to approximate

the output delay τo and the internal delay τ.

When you discretizetf and zpk models using the Tustin method, c2d first aggregates all
input, output, and transfer delays into a single transfer delay τTOT for each channel. c2d
then approximates τTOT as a Thiran filter and a chain of unit delays in the same way as
described for each of the time delays in ss models.

3-107
3 Linear Model Identification

For more information about Thiran filters, see the thiran reference page and [4].

Zero-Pole Matching Equivalents

The method of conversion by computing zero-pole matching equivalents applies only to
SISO systems. The continuous and discretized systems have matching DC gains. Their
poles and zeros are related by the transformation:

zi = esi Ts

where:

• zi is the ith pole or zero of the discrete-time system.

• si is the ith pole or zero of the continuous-time system.
• Ts is the sampling time.

See [2] for more information.

Zero-Pole Matching for Systems with Time Delays

You can use zero-pole matching to discretize SISO continuous-time models with time
delay, except that the method does not support ss models with internal delays. The zero-
pole matching method handles time delays in the same way as the Tustin approximation.
See “Tustin Approximation for Systems with Time Delays” on page 3-106.

References
[1] Åström, K.J. and B. Wittenmark, Computer-Controlled Systems: Theory and Design,
Prentice-Hall, 1990, pp. 48-52.

[2] Franklin, G.F., Powell, D.J., and Workman, M.L., Digital Control of Dynamic Systems
(3rd Edition), Prentice Hall, 1997.

[3] Smith, J.O. III, “Impulse Invariant Method”, Physical Audio Signal
Processing, August 2007. http://www.dsprelated.com/dspbooks/pasp/
Impulse_Invariant_Method.html.

[4] T. Laakso, V. Valimaki, “Splitting the Unit Delay”, IEEE Signal Processing Magazine,
Vol. 13, No. 1, p.30-60, 1996.

3-108
 Continuous-Discrete Conversion Methods

See Also
c2d | c2dOptions | d2c | d2cOptions | d2d | thiran

3-109
3 Linear Model Identification

Effect of Input Intersample Behavior on Continuous-Time Models

The intersample behavior of the input signals influences the estimation, simulation
and prediction of continuous-time models. A sampled signal is characterized only by its
values at the sampling instants. However, when you apply a continuous-time input to a
continuous-time system, the output values at the sampling instants depend on the inputs
at the sampling instants and on the inputs between these points.

The iddata and idfrd objects have an InterSample property which stores how the
input behaves between the sampling instants. You can specify the behavior between the
samples to be piecewise constant (zero-order hold), linearly interpolated between the
samples (first-order hold) or band-limited. A band-limited intersample behavior of the
input signal means:

• A filtered input signal (an input of finite bandwidth) was used to excite the system
dynamics
• The input was measured using a sampling device (A/D converter with antialiasing)
that reported it to be band-limited even though the true input entering the system
was piecewise constant or linear. In this case, the sampling devices can be assumed to
be a part of the system being modeled.

When input signal is band-limited, the estimation is performed as follows:

• Time-domain data is converted into frequency domain data using fft and the sample
time of the data is set to zero.
• Discrete-time frequency domain data (iddata with domain = 'frequency' or
idfrd with sample time Ts≠0) is treated as continuous-time data by setting the
sample time Ts to zero.

The resulting continuous-time frequency domain data is used for model estimation. For
more information, see Pintelon, R. and J. Schoukens, System Identification. A Frequency
Domain Approach, section 10.2, pp-352-356,Wiley-IEEE Press, New York, 2001.

Similarly, the intersample behavior of the input data affects the results of simulation
and prediction of continuous-time models. sim and predict commands use the
InterSample property to choose the right algorithm for computing model response.

The following example simulates a system using first-order hold ( foh ) intersample
behavior for input signal.
sys = idtf([-1 -2],[1 2 1 0.5]);

3-110
 Effect of Input Intersample Behavior on Continuous-Time Models

rng('default')
u = idinput([100 1 5],'sine',[],[],[5 10 1]);
Ts = 2;
y = lsim(sys, u, (0:Ts:999)', 'foh');

Create an iddata object for the simulated input-output data.

data = iddata(y,u,Ts);

The default intersample behavior is zero-order hold ( zoh ).

data.InterSample

ans =

zoh

Estimate a transfer function using this data.

np = 3; % number of poles
nz = 1; % number of zeros
opt = tfestOptions('InitMethod','all','Display','on');
opt.SearchOption.MaxIter = 100;
modelZOH = tfest(data,np,nz,opt)

modelZOH =

From input "u1" to output "y1":

-217.2 s - 391.6
---------------------------------
s^3 + 354.4 s^2 + 140.2 s + 112.4

Continuous-time identified transfer function.

Parameterization:
Number of poles: 3 Number of zeros: 1
Number of free coefficients: 5
Use "tfdata", "getpvec", "getcov" for parameters and their uncertainties.

Status:
Estimated using TFEST on time domain data "data".

3-111
3 Linear Model Identification

Fit to estimation data: 81.38% (simulation focus)

FPE: 0.1261, MSE: 0.111

The model gives about 80% fit to data. The sample time of the data is large enough that
intersample inaccuracy (using zoh rather than foh ) leads to significant modeling errors.

Re-estimate the model using foh intersample behavior.

data.InterSample = 'foh';
modelFOH = tfest(data, np, nz,opt)

modelFOH =

From input "u1" to output "y1":

-1.197 s - 0.06843
-------------------------------------
s^3 + 0.4824 s^2 + 0.3258 s + 0.01723

Continuous-time identified transfer function.

Parameterization:
Number of poles: 3 Number of zeros: 1
Number of free coefficients: 5
Use "tfdata", "getpvec", "getcov" for parameters and their uncertainties.

Status:
Estimated using TFEST on time domain data "data".
Fit to estimation data: 97.7% (simulation focus)
FPE: 0.001747, MSE: 0.001693

modelFOH is able to retrieve the original system correctly.

Compare the model outputs with data.

compare(data, modelZOH, modelFOH)

3-112
 Effect of Input Intersample Behavior on Continuous-Time Models

modelZOH is compared to data whose intersample behavior is foh. Therefore, its fit
decreases to around 70%.

3-113
3 Linear Model Identification

Transforming Between Linear Model Representations

You can transform linear models between state-space and polynomial forms. You can also
transform between frequency-response, state-space, and polynomial forms.

If you used the System Identification app to estimate models, you must export the models
to the MATLAB workspace before converting models.

For detailed information about each command in the following table, see the
corresponding reference page.

Commands for Transforming Model Representations

Command Model Type to Convert Usage Example
idfrd Converts any linear model to To get frequency response of m at default
an idfrd model. frequencies, use the following command:
If you have the Control System
Toolbox product, this command m_f = idfrd(m)
converts any numeric LTI To get frequency response at specific
model too. frequencies, use the following command:
m_f = idfrd(m,f)
To get frequency response for a submodel
from input 2 to output 3, use the following
command:
m_f = idfrd(m(2,3))
idpoly Converts any linear identified To get an ARMAX model from state-space
model, except idfrd, to model m_ss, use the following command:
ARMAX representation if the
original model has a nontrivial m_p = idpoly(m_ss)
noise component, or OE if the
noise model is trivial (H = 1).
If you have the Control System
Toolbox product, this command
converts any numeric LTI
model, except frd.
idss Converts any linear identified To get a state-space model from an ARX
model, except idfrd, to state- model m_arx, use the following command:
space representation.
m_ss = idss(m_arx)

3-114
 Transforming Between Linear Model Representations

Command Model Type to Convert Usage Example

If you have the Control System
Toolbox product, this command
converts any numeric LTI
model, except frd.
idtf Converts any linear To get a transfer function from a state-
identified model, except space model m_ss, use the following
idfrd, to transfer function command:
representation. The noise
component of the original m_tf = idtf(m_ss)
model is lost since an idtf
object has no elements to model
noise dynamics.
If you have the Control System
Toolbox product, this command
converts any numeric LTI
model, except frd.

Note: Most transformations among identified models (among idss, idtf, idpoly)
causes the parameter covariance information to be lost, with few exceptions:

• Conversion of an idtf model to an idpoly model.

• Conversion of an idgrey model to an idss model.

If you want to translate the estimated parameter covariance during conversion, use
translatecov.

3-115
3 Linear Model Identification

Subreferencing Models
In this section...
“What Is Subreferencing?” on page 3-116
“Limitation on Supported Models” on page 3-116
“Subreferencing Specific Measured Channels” on page 3-116
“Separation of Measured and Noise Components of Models” on page 3-117
“Treating Noise Channels as Measured Inputs” on page 3-118

What Is Subreferencing?
You can use subreferencing to create models with subsets of inputs and outputs from
existing multivariable models. Subreferencing is also useful when you want to generate
model plots for only certain channels, such as when you are exploring multiple-output
models for input channels that have minimal effect on the output.

The toolbox supports subreferencing operations for idtf, idpoly, idproc, idss, and
idfrd model objects.

Subreferencing is not supported for idgrey models. If you want to analyze the sub-
model, convert it into an idss model first, and then subreference the I/Os of the idss
model. If you want a grey-box representation of a subset of I/Os, create a new idgrey
model that uses an ODE function returning the desired I/O dynamics.

In addition to subreferencing the model for specific combinations of measured inputs and
output, you can subreference dynamic and noise models individually.

Limitation on Supported Models

Subreferencing nonlinear models is not supported.

Subreferencing Specific Measured Channels

Use the following general syntax to subreference specific input and output channels in
models:

model(outputs,inputs)

3-116
 Subreferencing Models

In this syntax, outputs and inputs specify channel indexes or channel names.

To select all output or all input channels, use a colon (:). To select no channels, specify
an empty matrix ([]). If you need to reference several channel names, use a cell array of
strings.

For example, to create a new model m2 from m from inputs 1 ('power') and 4 ('speed')
to output number 3 ('position'), use either of the following equivalent commands:

m2 = m('position',{'power','speed'})

or
m2 = m(3,[1 4])

For a single-output model, you can use the following syntax to subreference specific input
channels without ambiguity:
m3 = m(inputs)

Similarly, for a single-input model, you can use the following syntax to subreference
specific output channels:
m4 = m(outputs)

Separation of Measured and Noise Components of Models

For linear models, the general symbolic model description is given by:

y = Gu + He

G is an operator that takes the measured inputs u to the outputs and captures the
system dynamics.

H is an operator that describes the properties of the additive output disturbance and
takes the hypothetical (unmeasured) noise source inputs to the outputs. H represents the
noise model. When you specify to estimate a noise model, the resulting model include one
noise channel e at the input for each output in your system.

Thus, linear, parametric models represent input-output relationships for two kinds of
input channels: measured inputs and (unmeasured) noise inputs. For example, consider
the ARX model given by one of the following equations:

3-117
3 Linear Model Identification

A( q) y( t) = B(q) u ( t - nk ) + e( t)

or

B( q) 1
y( t) = u (t ) + e(t)
A( q) A( q)

In this case, the dynamic model is the relationship between the measured input u and
output y, G = B( q) A ( q) . The noise model is the contribution of the input noise e to the

output y, given by H = 1 A ( q) .

Suppose that the model m contains both a dynamic model G and a noise model H.
To create a new model that only has G and no noise contribution, simply set its
NoiseVariance property value to zero value.

To create a new model by subreferencing H due to unmeasured inputs, use the following
syntax:
m_H = m(:,[])

This operation creates a time-series model from m by ignoring the measured input.

The covariance matrix of e is given by the model property NoiseVariance, which is the
matrix L :

L = LLT

The covariance matrix of e is related to v, as follows:

e = Lv

where v is white noise with an identity covariance matrix representing independent noise
sources with unit variances.

Treating Noise Channels as Measured Inputs

To study noise contributions in more detail, it might be useful to convert the noise
channels to measured channels using noisecnv:

3-118
 Subreferencing Models

m_GH = noisecnv(m)

This operation creates a model m_GH that represents both measured inputs u and noise
inputs e, treating both sources as measured signals. m_GH is a model from u and e to y,
describing the transfer functions G and H.

Converting noise channels to measured inputs loses information about the variance of
the innovations e. For example, step response due to the noise channels does not take
into consideration the magnitude of the noise contributions. To include this variance
information, normalize e such that v becomes white noise with an identity covariance
matrix, where

e = Lv

To normalize e, use the following command:

m_GH = noisecnv(m,'Norm')

This command creates a model where u and v are treated as measured signals, as
follows:

È u˘
y( t) = Gu( t) + HLv = [ G HL] Í ˙
Îv˚

For example, the scaling by L causes the step responses from v to y to reflect the size of
the disturbance influence.

The converted noise sources are named in a way that relates the noise channel to the
corresponding output. Unnormalized noise sources e are assigned names such as 'e@y1',
'e@y2', ..., 'e@yn', where 'e@yn' refers to the noise input associated with the output
yn. Similarly, normalized noise sources v, are named 'v@y1', 'v@y2', ..., 'v@yn'.

If you want to create a model that has only the noise channels of an identified model as
its measured inputs, use the noise2meas command. It results in a model with y(t) = He
or y(t) = HLv, where e or v is treated as a measured input.

Note: When you plot models in the app that include noise sources, you can select to view
the response of the noise model corresponding to specific outputs. For more information,
see “Selecting Measured and Noise Channels in Plots” on page 14-14.

3-119
3 Linear Model Identification

Concatenating Models
In this section...
“About Concatenating Models” on page 3-120
“Limitation on Supported Models” on page 3-120
“Horizontal Concatenation of Model Objects” on page 3-121
“Vertical Concatenation of Model Objects” on page 3-121
“Concatenating Noise Spectrum Data of idfrd Objects” on page 3-122
“See Also” on page 3-123

About Concatenating Models

You can perform horizontal and vertical concatenation of linear model objects to grow the
number of inputs or outputs in the model.

When you concatenate identified models, such as idtf, idpoly, idproc, and idss
model objects, the resulting model combines the parameters of the individual models.
However, the estimated parameter covariance is lost. If you want to translate the
covariance information during concatenation, use translatecov.

Concatenation is not supported for idgrey models; convert them to idss models first if
you want to perform concatenation.

You can also concatenate nonparametric models, which contain the estimated impulse-
response (idtf object) and frequency-response (idfrd object) of a system.

In case of idfrd models, concatenation combines information in the ResponseData

properties of the individual model objects. ResponseData is an ny-by-nu-by-nf array
that stores the response of the system, where ny is the number of output channels, nu is
the number of input channels, and nf is the number of frequency values. The (j,i,:)
vector of the resulting response data represents the frequency response from the ith
input to the jth output at all frequencies.

Limitation on Supported Models

Concatenation is supported for linear models only.

3-120
 Concatenating Models

Horizontal Concatenation of Model Objects

Horizontal concatenation of model objects requires that they have the same outputs.
If the output channel names are different and their dimensions are the same, the
concatenation operation resets the output names to their default values.

The following syntax creates a new model object m that contains the horizontal
concatenation of m1,m2,...,mN:
m = [m1,m2,...,mN]

m takes all of the inputs of m1,m2,...,mN to the same outputs as in the original models.
The following diagram is a graphical representation of horizontal concatenation of the
models.

u1 y1 y1
Model 1 u3 Model 2
u2 y2 y2

u1 y1
Combined Horizonal Concatenation Same
Inputs u2 of Model 1 and Model 2 Outputs
y2
u3

Vertical Concatenation of Model Objects

Vertical concatenation combines output channels of specified models. Vertical
concatenation of model objects requires that they have the same inputs. If the input
channel names are different and their dimensions are the same, the concatenation
operation resets the input channel names to their default ('') values.

The following syntax creates a new model object m that contains the vertical
concatenation of m1,m2,...,mN:
m = [m1;m2;... ;mN]

m takes the same inputs in the original models to all of the output of m1,m2,...,mN. The
following diagram is a graphical representation of vertical concatenation of frequency-
response data.

3-121
3 Linear Model Identification

u1 y1 u1
Model 1 Model 2 y3
u2 y2 u2

y1
Same u1 Vertical Concatenation Combined
y2
Inputs u2 of Model 1 and Model 2 Outputs
y3

Concatenating Noise Spectrum Data of idfrd Objects

When idfrd models are obtained as a result of estimation (such as using spa), the
SpectrumData property is not empty and contains the power spectra and cross spectra
of the output noise in the system. For each output channel, this toolbox estimates
one noise channel to explain the difference between the output of the model and the
measured output.

When the SpectrumData property of individual idfrd objects is not empty, horizontal
and vertical concatenation handle SpectrumData, as follows.

In case of horizontal concatenation, there is no meaningful way to combine the

SpectrumData of individual idfrd objects, and the resulting SpectrumData property
is empty. An empty property results because each idfrd object has its own set of noise
channels, where the number of noise channels equals the number of outputs. When
the resulting idfrd object contains the same output channels as each of the individual
idfrd objects, it cannot accommodate the noise data from all the idfrd objects.

In case of vertical concatenation, this toolbox concatenates individual noise models

diagonally. The following shows that m.SpectrumData is a block diagonal matrix of the
power spectra and cross spectra of the output noise in the system:

Ê m1.s
Á 0 ˆ˜
m.s = Á O ˜
ÁÁ ˜
Ë 0 mN .s ˜¯

s in m.s is the abbreviation for the SpectrumData property name.

3-122
 Concatenating Models

See Also
If you have the Control System Toolbox product, see “Combining Model Objects” on page
12-5 about additional functionality for combining models.

3-123
3 Linear Model Identification

Merging Models
You can merge models of the same structure to obtain a single model with parameters
that are statistically weighed means of the parameters of the individual models. When
computing the merged model, the covariance matrices of the individual models determine
the weights of the parameters.

You can perform the merge operation for the idtf, idgrey, idpoly, idproc, and idss
model objects.

Note: Each merge operation merges the same type of model object.

Merging models is an alternative to merging data sets into a single multiexperiment

data set, and then estimating a model for the merged data. Whereas merging data sets
assumes that the signal-to-noise ratios are about the same in the two experiments,
merging models allows greater variations in model uncertainty, which might result from
greater disturbances in an experiment.

When the experimental conditions are about the same, merge the data instead of models.
This approach is more efficient and typically involves better-conditioned calculations. For
more information about merging data sets into a multiexperiment data set, see “Create
Multiexperiment Data at the Command Line” on page 2-60.

For more information about merging models, see the merge reference page.

3-124
 Building and Estimating Process Models Using System Identification Toolbox™

Building and Estimating Process Models Using System

Identification Toolbox™
This example shows how to build simple process models using System Identification
Toolbox™. Techniques for creating these models and estimating their parameters using
experimental data is described. This example requires Simulink®.

Introduction

This example illustrates how to build simple process models often used in process
industry. Simple, low-order continuous-time transfer functions are usually employed to
describe process behavior. Such models are described by IDPROC objects which represent
the transfer function in a pole-zero-gain form.

Process models are of the basic type 'Static Gain + Time Constant + Time Delay'. They
may be represented as:

or as an integrating process:

where the user can determine the number of real poles (0, 1, 2 or 3), as well as the
presence of a zero in the numerator, the presence of an integrator term (1/s) and the
presence of a time delay (Td). In addition, an underdamped (complex) pair of poles may
replace the real poles.

Representation of Process Models using IDPROC Objects

IDPROC objects define process models by using the letters P (for process model), D (for
time delay), Z (for a zero) and I (for integrator). An integer will denote the number of
poles. The models are generated by calling idproc with a string identifier using these
letters.

This should be clear from the following examples.

3-125
3 Linear Model Identification

idproc('P1') % transfer function with only one pole (no zeros or delay)
idproc('P2DIZ') % model with 2 poles, delay integrator and delay
idproc('P0ID') % model with no poles, but an integrator and a delay

ans =
Process model with transfer function:
Kp
G(s) = ----------
1+Tp1*s

Kp = NaN
Tp1 = NaN

Parameterization:
'P1'
Number of free coefficients: 2
Use "getpvec", "getcov" for parameters and their uncertainties.

Status:
Created by direct construction or transformation. Not estimated.

ans =
Process model with transfer function:
1+Tz*s
G(s) = Kp * ------------------- * exp(-Td*s)
s(1+Tp1*s)(1+Tp2*s)

Kp = NaN
Tp1 = NaN
Tp2 = NaN
Td = NaN
Tz = NaN

Parameterization:
'P2DIZ'
Number of free coefficients: 5
Use "getpvec", "getcov" for parameters and their uncertainties.

Status:
Created by direct construction or transformation. Not estimated.

ans =
Process model with transfer function:
Kp

3-126
 Building and Estimating Process Models Using System Identification Toolbox™

G(s) = --- * exp(-Td*s)

s

Kp = NaN
Td = NaN

Parameterization:
'P0DI'
Number of free coefficients: 2
Use "getpvec", "getcov" for parameters and their uncertainties.

Status:
Created by direct construction or transformation. Not estimated.

Creating an IDPROC Object (using a Simulink® Model as Example)

Consider the system described by the following SIMULINK model:

open_system('iddempr1')
set_param('iddempr1/Random Number','seed','0')

The red part is the system, the blue part is the controller and the reference signal is a
swept sinusoid (a chirp signal). The data sampling time is set to 0.5 seconds. As observed,
the system is a continuous-time transfer function, and can hence be described using
model objects in System Identification Toolbox, such as idss, idpoly or idproc.

Let us describe the system using idpoly and idproc objects. Using idpoly object, the
system may be described as:

m0 = idpoly(1,0.1,1,1,[1 0.5],'Ts',0,'InputDelay',1.57,'NoiseVariance',0.01);

3-127
3 Linear Model Identification

The IDPOLY form used above is useful for describing transfer functions of arbitrary
orders. Since the system we are considering here is quite simple (one pole and no zeros),
and is continuous-time, we may use the simpler IDPROC object to capture its dynamics:

m0p = idproc('p1d','Kp',0.2,'Tp1',2,'Td',1.57) % one pole+delay, with initial values

% for gain, pole and delay specified.

m0p =
Process model with transfer function:
Kp
G(s) = ---------- * exp(-Td*s)
1+Tp1*s

Kp = 0.2
Tp1 = 2
Td = 1.57

Parameterization:
'P1D'
Number of free coefficients: 3
Use "getpvec", "getcov" for parameters and their uncertainties.

Status:
Created by direct construction or transformation. Not estimated.

Estimating Parameters of IDPROC Models

Once a system is described by a model object, such as IDPROC, it may be used for
estimation of its parameters using measurement data. As an example, we consider the
problem of estimation of parameters of the Simulink model's system (red portion) using
simulation data. We begin by acquiring data for estimation:

sim('iddempr1')
dat1e = iddata(y,u,0.5); % The IDDATA object for storing measurement data

Let us look at the data:

plot(dat1e)

3-128
 Building and Estimating Process Models Using System Identification Toolbox™

We can identify a process model using procest command, by providing the same
structure information specified to create IDPROC models. For example, the 1-pole+delay
model may be estimated by calling procest as follows:
m1 = procest(dat1e,'p1d'); % estimation of idproc model using data 'dat1e'.

% Check the result of estimation:

m1

m1 =
Process model with transfer function:
Kp
G(s) = ---------- * exp(-Td*s)
1+Tp1*s

3-129
3 Linear Model Identification

Kp = 0.20045
Tp1 = 2.0431
Td = 1.499

Parameterization:
'P1D'
Number of free coefficients: 3
Use "getpvec", "getcov" for parameters and their uncertainties.

Status:
Estimated using PROCEST on time domain data "dat1e".
Fit to estimation data: 87.34%
FPE: 0.01068, MSE: 0.01062

To get information about uncertainties, use

present(m1)

m1 =
Process model with transfer function:
Kp
G(s) = ---------- * exp(-Td*s)
1+Tp1*s

Kp = 0.20045 +/- 0.00077275

Tp1 = 2.0431 +/- 0.061216
Td = 1.499 +/- 0.040854

Parameterization:
'P1D'
Number of free coefficients: 3
Use "getpvec", "getcov" for parameters and their uncertainties.

Status:
Termination condition: Near (local) minimum, (norm(g) < tol).
Number of iterations: 4, Number of function evaluations: 9

Estimated using PROCEST on time domain data "dat1e".

Fit to estimation data: 87.34%
FPE: 0.01068, MSE: 0.01062
More information in model's "Report" property.

3-130
 Building and Estimating Process Models Using System Identification Toolbox™

The model parameters, K, Tp1 and Td are now shown with one standard deviation
uncertainty range.

Computing Time and Frequency Response of IDPROC Models

The model m1 estimated above is an IDPROC model object to which all of the toolbox's
model commands can be applied:

step(m1,m0) %step response of models m1 (estimated) and m0 (actual)

legend('m1 (estimated parameters)','m0 (known parameters)','location','northwest')

Bode response with confidence region corresponding to 3 standard deviations may be

computed by doing:

3-131
3 Linear Model Identification

h = bodeplot(m1,m0);
showConfidence(h,3)

Similarly, the measurement data may be compared to the models outputs using compare
as follows:

compare(dat1e,m0,m1)

3-132
 Building and Estimating Process Models Using System Identification Toolbox™

Other operations such as sim, impulse, c2d are also available, just as they are for other
model objects.

bdclose('iddempr1')

Accommodating the Effect of Intersample Behavior in Estimation

It may be important (at least for slow sampling) to consider the intersample behavior of
the input data. To illustrate this, let us study the same system as before, but without the
sample-and-hold circuit:

open_system('iddempr5')

3-133
3 Linear Model Identification

Simulate this system with the same sampling interval:

sim('iddempr5')
dat1f = iddata(y,u,0.5); % The IDDATA object for the simulated data

We estimate an IDPROC model using data1f while also imposing an upper bound on
the allowable value delay. We will use 'lm' as search method and also choose to view the
estimation progress.

m2_init = idproc('P1D');
m2_init.Structure.Td.Maximum = 2;
opt = procestOptions('SearchMethod','lm','Display','on');
m2 = procest(dat1f,m2_init,opt);
m2

m2 =
Process model with transfer function:
Kp
G(s) = ---------- * exp(-Td*s)
1+Tp1*s

Kp = 0.20038
Tp1 = 2.01
Td = 1.31

Parameterization:
'P1D'
Number of free coefficients: 3

3-134
 Building and Estimating Process Models Using System Identification Toolbox™

Use "getpvec", "getcov" for parameters and their uncertainties.

Status:
Estimated using PROCEST on time domain data "dat1f".
Fit to estimation data: 87.26%
FPE: 0.01067, MSE: 0.01061

This model has a slightly less precise estimate of the delay than the previous one, m1:

[m0p.Td, m1.Td, m2.Td]

step(m0,m1,m2)
legend('m0 (actual)','m1 (estimated with ZOH)','m2 (estimated without ZOH)','location',

ans =

1.5700 1.4990 1.3100

3-135
3 Linear Model Identification

However, by telling the estimation process that the intersample behavior is first-order-
hold (an approximation to the true continuous) input, we do better:

dat1f.InterSample = 'foh';
m3 = procest(dat1f,m2_init,opt);

Compare the four models m0 (true) m1 (obtained from zoh input) m2 (obtained for
continuous input, with zoh assumption) and m3 (obtained for the same input, but with
foh assumption)

[m0p.Td, m1.Td, m2.Td, m3.Td]

compare(dat1e,m0,m1,m2,m3)
step(m0,m1,m2,m3)
legend('m0','m1','m2','m3')

3-136
 Building and Estimating Process Models Using System Identification Toolbox™

bdclose('iddempr5')

ans =

1.5700 1.4990 1.3100 1.5570

3-137
3 Linear Model Identification

Modeling a System Operating in Closed Loop

Let us now consider a more complex process, with integration, that is operated in closed
loop:

open_system('iddempr2')

3-138
 Building and Estimating Process Models Using System Identification Toolbox™

The true system can be represented by:

m0 = idproc('P2ZDI','Kp',1,'Tp1',1,'Tp2',5,'Tz',3,'Td',2.2);

The process is controlled by a PD regulator with limited input amplitude and a zero order
hold device. The sampling interval is 1 second.

set_param('iddempr2/Random Number','seed','0')
sim('iddempr2')
dat2 = iddata(y,u,1); % IDDATA object for estimation

Two different simulations are made, the first for estimation and the second one for
validation purposes.

set_param('iddempr2/Random Number','seed','13')
sim('iddempr2')
dat2v = iddata(y,u,1); % IDDATA object for validation purpose

Let us look at the data (estimation and validation).

plot(dat2,dat2v)
legend('dat2 (estimation)','dat2v (validation)')

3-139
3 Linear Model Identification

Let us now perform estimation using dat2.

m2_init = idproc('P2ZDI');
m2_init.Structure.Td.Maximum = 5;
m2_init.Structure.Tp1.Maximum = 2;
opt = procestOptions('SearchMethod','lsq','Display','on');
opt.SearchOption.MaxIter = 100;
m2 = procest(dat2, m2_init, opt)

m2 =
Process model with transfer function:
1+Tz*s
G(s) = Kp * ------------------- * exp(-Td*s)
s(1+Tp1*s)(1+Tp2*s)

3-140
 Building and Estimating Process Models Using System Identification Toolbox™

Kp = 1.0028
Tp1 = 1.1175
Tp2 = 5.9022
Td = 2.133
Tz = 3.8084

Parameterization:
'P2DIZ'
Number of free coefficients: 5
Use "getpvec", "getcov" for parameters and their uncertainties.

Status:
Estimated using PROCEST on time domain data "dat2".
Fit to estimation data: 91.86%
FPE: 0.1041, MSE: 0.1005

compare(dat2v,m2,m0) % Gives very good agreement with data

3-141
3 Linear Model Identification

bode(m2,m0)
legend({'m2 (est)','m0 (actual)'},'location','west')

3-142
 Building and Estimating Process Models Using System Identification Toolbox™

impulse(m2,m0)
legend({'m2 (est)','m0 (actual)'})

3-143
3 Linear Model Identification

Compare also with the parameters of the true system:

present(m2)
[getpvec(m0), getpvec(m2)]

m2 =
Process model with transfer function:
1+Tz*s
G(s) = Kp * ------------------- * exp(-Td*s)
s(1+Tp1*s)(1+Tp2*s)

Kp = 1.0028 +/- 0.016478

Tp1 = 1.1175 +/- 2.114
Tp2 = 5.9022 +/- 4.1107

3-144
 Building and Estimating Process Models Using System Identification Toolbox™

Td = 2.133 +/- 0.61131

Tz = 3.8084 +/- 5.4513

Parameterization:
'P2DIZ'
Number of free coefficients: 5
Use "getpvec", "getcov" for parameters and their uncertainties.

Status:
Termination condition: Change in cost was less than the specified tolerance.
Number of iterations: 32, Number of function evaluations: 33

Estimated using PROCEST on time domain data "dat2".

Fit to estimation data: 91.86%
FPE: 0.1041, MSE: 0.1005
More information in model's "Report" property.

ans =

1.0000 1.0028
1.0000 1.1175
5.0000 5.9022
2.2000 2.1330
3.0000 3.8084

A word of caution. Identification of several real time constants may sometimes be an ill-
conditioned problem, especially if the data are collected in closed loop.

To illustrate this, let us estimate a model based on the validation data:

m2v = procest(dat2v, m2_init, opt)
[getpvec(m0), getpvec(m2), getpvec(m2v)]

m2v =
Process model with transfer function:
1+Tz*s
G(s) = Kp * ------------------- * exp(-Td*s)
s(1+Tp1*s)(1+Tp2*s)

Kp = 1.0059
Tp1 = 0.22619
Tp2 = 3.6898
Td = 0.943

3-145
3 Linear Model Identification

Tz = -0.33583

Parameterization:
'P2DIZ'
Number of free coefficients: 5
Use "getpvec", "getcov" for parameters and their uncertainties.

Status:
Estimated using PROCEST on time domain data "dat2v".
Fit to estimation data: 91.98%
FPE: 0.1032, MSE: 0.09958

ans =

1.0000 1.0028 1.0059

1.0000 1.1175 0.2262
5.0000 5.9022 3.6898
2.2000 2.1330 0.9430
3.0000 3.8084 -0.3358

This model has much worse parameter values. On the other hand, it performs nearly
identically to the true system m0 when tested on the other data set dat2:

compare(dat2,m0,m2,m2v)

3-146
 Building and Estimating Process Models Using System Identification Toolbox™

Fixing Known Parameters During Estimation

Suppose we know from other sources that one time constant is 1:

m2v.Structure.Tp1.Value = 1;
m2v.Structure.Tp1.Free = false;

We can fix this value, while estimating the other parameters:

m2v = procest(dat2v,m2v)
%

m2v =
Process model with transfer function:

3-147
3 Linear Model Identification

1+Tz*s
G(s) = Kp * ------------------- * exp(-Td*s)
s(1+Tp1*s)(1+Tp2*s)

Kp = 1.0111
Tp1 = 1
Tp2 = 5.2957
Td = 2.194
Tz = 3.2249

Parameterization:
'P2DIZ'
Number of free coefficients: 4
Use "getpvec", "getcov" for parameters and their uncertainties.

Status:
Estimated using PROCEST on time domain data "dat2v".
Fit to estimation data: 92.05%
FPE: 0.09989, MSE: 0.09794

As observed, fixing Tp1 to its known value dramatically improves the estimates of the
remaining parameters in model m2v.

This also indicates that simple approximation should do well on the data:
m1x_init = idproc('P2D'); % simpler structure (no zero, no integrator)
m1x_init.Structure.Td.Maximum = 2;
m1x = procest(dat2v, m1x_init)
compare(dat2,m0,m2,m2v,m1x)

m1x =
Process model with transfer function:
Kp
G(s) = ----------------- * exp(-Td*s)
(1+Tp1*s)(1+Tp2*s)

Kp = 14815
Tp1 = 15070
Tp2 = 3.1505
Td = 1.833

Parameterization:
'P2D'
Number of free coefficients: 4

3-148
 Building and Estimating Process Models Using System Identification Toolbox™

Use "getpvec", "getcov" for parameters and their uncertainties.

Status:
Estimated using PROCEST on time domain data "dat2v".
Fit to estimation data: 91.23%
FPE: 0.1209, MSE: 0.119

Thus, the simpler model is able to estimate system output pretty well. However, m1x
does not contain any integration, so the open loop long time range behavior will be quite
different:

step(m0,m2,m2v,m1x)
legend('m0','m2','m2v','m1x')
bdclose('iddempr2')

3-149
3 Linear Model Identification

Additional Information

For more information on identification of dynamic systems with System Identification

Toolbox visit the System Identification Toolbox product information page.

3-150
 Determining Model Order and Delay

Determining Model Order and Delay

Estimation requires you to specify the model order and delay. Many times, these values
are not known. You can determine the model order and delay in one of the following
ways:

• Guess their values by visually inspecting the data or based on the prior knowledge of
the system.
• Estimate delay as a part of idproc or idtf model estimation. These models
treat delay as an estimable parameter and you can determine their values by
the estimation commands procest and tfest, respectively. However automatic
estimation of delays can cause errors. Therefore, it is recommended that you analyze
the data for delays in advance.
• To estimate delays, you can also use one of the following tools:

• Estimate delay using delayest. The choice of the order of the underlying ARX
model and the lower/upper bound on the value of the delay to be estimated
influence the value returned by delayest.
• Compute impulse response using impulseest. Plot the impulse response with
a confidence interval of sufficient standard deviations (usually 3). The delay is
indicated by the number of response samples that are inside the statistically zero
region (marked by the confidence bound) before the response goes outside that
region.
• Select the model order in n4sid by specifying the model order as a vector.
• Choose the model order of an ARX model using arxstruc or ivstruc and
selstruc. These command select the number of poles, zeros and delay.

See “Model Structure Selection: Determining Model Order and Input Delay” on page
3-152 for an example of using these tools.

3-151
3 Linear Model Identification

Model Structure Selection: Determining Model Order and Input

Delay
This example shows some methods for choosing and configuring the model structure.
Estimation of a model using measurement data requires selection of a model structure
(such as state-space or transfer function) and its order (e.g., number of poles and zeros)
in advance. This choice is influenced by prior knowledge about the system being modeled,
but can also be motivated by an analysis of data itself. This example describes some
options for determining model orders and input delay.

Introduction

Choosing a model structure is usually the first step towards its estimation. There are
various possibilities for structure - state-space, transfer functions and polynomial forms
such as ARX, ARMAX, OE, BJ etc. If you do not have detailed prior knowledge of your
system, such as its noise characteristics and indication of feedback, the choice of a
reasonable structure may not be obvious. Also for a given choice of structure, the order
of the model needs to be specified before the corresponding parameters are estimated.
System Identification Toolbox™ offers some tools to assist in the task of model order
selection.

The choice of a model order is also influenced by the amount of delay. A good idea of
the input delay simplifies the task of figuring out the orders of other model coefficients.
Discussed below are some options for input delay determination and model structure and
order selection.

Choosing and Preparing Example Data for Analysis

This example uses the hair dryer data, also used by iddemo1 ("Estimating Simple Models
from Real Laboratory Process Data"). The process consists of air being fanned through a
tube. The air is heated at the inlet of the tube, and the input is the voltage applied to the
heater. The output is the temperature at the outlet of the tube.

Let us begin by loading the measurement data and doing some basic preprocessing:

load dry2

Form a data set for estimation of the first half, and a reference set for validation
purposes of the second half:

ze = dry2(1:500);

3-152
 Model Structure Selection: Determining Model Order and Input Delay

zr = dry2(501:1000);

Detrend each of the sets:

ze = detrend(ze);
zr = detrend(zr);

Let us look at a portion of the estimation data:

plot(ze(200:350))

Estimating Input Delay

There are various options available for determining the time delay from input to output.
These are:

3-153
3 Linear Model Identification

• Using the DELAYEST utility.

• Using a non-parametric estimate of the impulse response, using IMPULSEEST.
• Using the state-space model estimator N4SID with a number of different orders and
finding the delay of the 'best' one.

Using delayest:

Let us discuss the above options in detail. Function delayest returns an estimate of
the delay for a given choice of orders of numerator and denominator polynomials. This
function evaluates an ARX structure:

y(t) + a1*y(t-1) + ... + ana*y(t-na) = b1*u(t-nk) + ...+bnb*u(t-nb-

nk+1)

with various delays and chooses the delay value that seems to return the best fit. In this
process, chosen values of na and nb are used.

delay = delayest(ze) % na = nb = 2 is used, by default

delay =

A value of 3 is returned by default. But this value may change a bit if the assumed orders
of numerator and denominator polynomials (2 here) is changed. For example:

delay = delayest(ze,5,4)

delay =

returns a value of 2. To gain insight into how delayest works, let us evaluate the
loss function for various choices of delays explicitly. We select a second order model
(na=nb=2), which is the default for delayest, and try out every time delay between 1
and 10. The loss function for the different models are computed using the validation data
set:

3-154
 Model Structure Selection: Determining Model Order and Input Delay

V = arxstruc(ze,zr,struc(2,2,1:10));

We now select that delay that gives the best fit for the validation data:

[nn,Vm] = selstruc(V,0); % nn is given as [na nb nk]

The chosen structure was:

nn

nn =

2 2 3

which show the best model has a delay of nn(3) = 3.

We can also check how the fit depends on the delay. This information is returned in the
second output Vm. The logarithms of a quadratic loss function are given as the first row,
while the indexes na, nb and nk are given as a column below the corresponding loss
function.

Vm

Vm =

Columns 1 through 7

-0.1283 -1.3142 -1.8787 -0.2339 0.0084 0.0900 0.1957

2.0000 2.0000 2.0000 2.0000 2.0000 2.0000 2.0000
2.0000 2.0000 2.0000 2.0000 2.0000 2.0000 2.0000
1.0000 2.0000 3.0000 4.0000 5.0000 6.0000 7.0000

Columns 8 through 10

0.2082 0.1728 0.1631

2.0000 2.0000 2.0000
2.0000 2.0000 2.0000
8.0000 9.0000 10.0000

The choice of 3 delays is thus rather clear, since the corresponding

3-155
3 Linear Model Identification

loss is minimum.

Using impulse

To gain a better insight into the dynamics, let us compute the impulse response of the
system. We will use the function impulseest to compute a non-parametric impulse
response model. We plot this response with a confidence interval represented by 3
standard deviations.

FIRModel = impulseest(ze);
clf
h = impulseplot(FIRModel);
showConfidence(h,3)

3-156
 Model Structure Selection: Determining Model Order and Input Delay

The filled light-blue region shows the confidence interval for the insignificant response in
this estimation. There is a clear indication that the impulse response "takes off" (leaves
the uncertainty region) after 3 samples. This points to a delay of three intervals.

Using n4sid based state-space evaluation

We may also estimate a family of parametric models to find the delay corresponding
to the "best" model. In case of state-space models, a range of orders may be evaluated
simultaneously and the best order picked from a Hankel Singular Value plot. Execute the
following command to invoke n4sid in an interactive mode:

m = n4sid(ze,1:15); % All orders between 1 and 15.

3-157
3 Linear Model Identification

The plot indicates an order of 3 as the best value. For this choice, let us compute the
impulse response of the model m:

m = n4sid(ze, 3);
showConfidence(impulseplot(m),3)

As with non-parametric impulse response, there is a clear indication that the delay from
input to output is of three samples.

Choosing a Reasonable Model Structure

In lack of any prior knowledge, it is advisable to try out various available choices and
use the one that seems to work the best. State-space models may be a good starting point

3-158
 Model Structure Selection: Determining Model Order and Input Delay

since only the number of states needs to be specified in order to estimate a model. Also,
a range of orders may be evaluated quickly, using n4sid, for determining the best order,
as described in the next section. For polynomial models, a similar advantage is realized
using the arx estimator. Output-error (OE) models may also be good choice for a starting
polynomial model because of their simplicity.

Determining Model Order

Once you have decided upon a model structure to use, the next task is to determine the
order(s). In general, the aim should be to not use a model order higher than necessary.
This can be determined by analyzing the improvement in %fit as a function of model
order. When doing this, it is advisable to use a separate, independent dataset for
validation. Choosing an independent validation data set (zr in our example) would
improve the detection of over-fitting.

In addition to a progressive analysis of multiple model orders, explicit determination of

optimum orders can be performed for some model structures. Functions arxstruc and
selstruc may be used for choosing the best order for ARX models. For our example,
let us check the fit for all 100 combinations of up to 10 b-parameters and up to 10 a-
parameters, all with a delay value of 3:

V = arxstruc(ze,zr,struc(1:10,1:10,3));

The best fit for the validation data set is obtained for:

nn = selstruc(V,0)

nn =

10 4 3

Let us check how much the fit is improved for the higher order models. For this, we
use the function selstruc with only one input. In this case, a plot showing the fit as a
function of the number of parameters used is generated. The user is also prompted to
enter the number of parameters. The routine then selects a structure with these many
parameters that gives the best fit. Note that several different model structures use the
same number of parameters. Execute the following command to choose a model order
interactively:

nns = selstruc(V) %invoke selstruc in an interactive mode

3-159
3 Linear Model Identification

The best fit is thus obtained for nn = [4 4 3], while we see that the improved fit compared
to nn = [2 2 3] is rather marginal.

We may also approach this problem from the direction of reducing a higher order model.
If the order is higher than necessary, then the extra parameters are basically used to
"model" the measurement noise. These "extra" poles are estimated with a lower level of
accuracy (large confidence interval). If their are cancelled by a zero located nearby, then
it is an indication that this pole-zero pair may not be required to capture the essential
dynamics of the system.

3-160
 Model Structure Selection: Determining Model Order and Input Delay

For our example, let us compute a 4th order model:

th4 = arx(ze,[4 4 3]);

Let us check the pole-zero configuration for this model. We can also include confidence
regions for the poles and zeros corresponding to 3 standard deviations, in order to
determine how accurately they are estimated and also how close the poles and zeros are
to each other.
h = iopzplot(th4);
showConfidence(h,3)

The confidence intervals for the two complex-conjugate poles and zeros overlap,
indicating they are likely to cancel each other. Hence, a second order model might be
adequate. Based on this evidence, let us compute a 2nd order ARX model:

3-161
3 Linear Model Identification

th2 = arx(ze,[2 2 3]);

We can test how well this model (th2) is capable of reproducing the validation data set.
To compare the simulated output from the two models with the actual output (plotting
the mid 200 data points) we use the compare utility:

compare(zr(150:350),th2,th4)

The plot indicates that there was no significant loss of accuracy in reducing the order
from 4 to 2. We can also check the residuals ("leftovers") of this model, i.e., what is left
unexplained by the model.

e = resid(ze,th2);
plot(e(:,1,[])), title('The residuals')

3-162
 Model Structure Selection: Determining Model Order and Input Delay

We see that the residuals are quite small compared to the signal level of the output, that
they are reasonably well (although not perfectly) uncorrelated with the input and among
themselves. We can thus be (provisionally) satisfied with the model th2.

Let us now check if we can determine the model order for a state-space structure. As
before, we know the delay is 3 samples. We can try all orders from 1 to 15 with a total lag
of 3 samples in n4sid. Execute the following command to try various orders and choose
one interactively.

ms = n4sid(ze,[1:15],'InputDelay',2); %n4sid estimation with variable orders

3-163
3 Linear Model Identification

The "InputDelay" was set to 2 because by default n4sid estimates a model with no
feedthrough (which accounts for one sample lag between input and output). The default
order, indicated in the figure above, is 3, that is in good agreement with our earlier
findings. Finally, we compare how the state-space model ms and the ARX model th2
compare in reproducing the measured output of the validation data:

ms = n4sid(ze,3,'InputDelay',2);
compare(zr,ms,th2)

3-164
 Model Structure Selection: Determining Model Order and Input Delay

The comparison plot indicates that the two models are practically identical.

Conclusions

This example described some options for choosing a reasonable model order. Determining
delay in advance can simplify the task of choosing orders. With ARX and state-space
structures, we have some special tools (arx and n4sid estimators) for automatically
evaluating a whole set of model orders, and choosing the best one among them. The
information revealed by this exercise (using utilities such as arxstruc, selstruc,
n4sid and delayest) could be used as a starting point when estimating models of other
structures, such as BJ and ARMAX.

Additional Information

3-165
3 Linear Model Identification

For more information on identification of dynamic systems with System Identification

Toolbox visit the System Identification Toolbox product information page.

3-166
 Frequency Domain Identification: Estimating Models Using Frequency Domain Data

Frequency Domain Identification: Estimating Models Using

Frequency Domain Data
This example shows how to estimate models using frequency domain data. The
estimation and validation of models using frequency domain data work the same way as
they do with time domain data. This provides a great amount of flexibility in estimation
and analysis of models using time and frequency domain as well as spectral (FRF) data.
You may simultaneously estimate models using data in both domains, compare and
combine these models. A model estimated using time domain data may be validated
using spectral data or vice-versa.

Frequency domain data can not be used for estimation or validation of nonlinear models.

Introduction

Frequency domain experimental data are common in many applications. It could be that
the data was collected as frequency response data (frequency functions: FRF) from the
process using a frequency analyzer. It could also be that it is more practical to work with
the input's and output's Fourier transforms (FFT of time-domain data), for example to
handle periodic or band-limited data. (A band-limited continuous time signal has no
frequency components above the Nyquist frequency). In System Identification Toolbox,
frequency domain I/O data are represented the same way as time-domain data, i.e.,
using iddata objects. The 'Domain' property of the object must be set to 'Frequency'.
Frequency response data are represented as complex vectors or as magnitude/phase
vectors as a function of frequency. IDFRD objects in the toolbox are used to encapsulate
FRFs, where a user specifies the complex response data and a frequency vector. Such
IDDATA or IDFRD objects (and also FRD objects of Control System Toolbox) may be
used seamlessly with any estimation routine (such as procest, tfest etc).

Inspecting Frequency Domain Data

Let us begin by loading some frequency domain data:

load demofr

This MAT-file contains frequency response data at frequencies W, with the amplitude
response AMP and the phase response PHA. Let us first have a look at the data:

subplot(211), loglog(W,AMP),title('Amplitude Response')

subplot(212), semilogx(W,PHA),title('Phase Response')

3-167
3 Linear Model Identification

This experimental data will now be stored as an IDFRD object. First transform
amplitude and phase to a complex valued response:

zfr = AMP.*exp(1i*PHA*pi/180);
Ts = 0.1;
gfr = idfrd(zfr,W,Ts);

Ts is the sampling interval of the underlying data. If the data corresponds to continuous
time, for example since the input has been band-limited, use Ts = 0.

Note: If you have the Control System Toolbox™, you could use an FRD object instead of
the IDFRD object. IDFRD has options for more information, like disturbance spectra and
uncertainty measures which are not available in FRD objects.

3-168
 Frequency Domain Identification: Estimating Models Using Frequency Domain Data

The IDFRD object gfr now contains the data, and it can be plotted and analyzed in
different ways. To view the data, we may use plot or bode:
clf
bode(gfr), legend('gfr')

Estimating Models Using Frequency Response (FRF) Data

To estimate models, you can now use gfr as a data set with all the commands of the
toolbox in a transparent fashion. The only restriction is that noise models cannot be built.
This means that for polynomial models only OE (output-error models) apply, and for
state-space models, you have to fix K = 0.
m1 = oe(gfr,[2 2 1]) % Discrete-time Output error (transfer function) model
ms = ssest(gfr) % Continuous-time state-space model with default choice of order

3-169
3 Linear Model Identification

mproc = procest(gfr,'P2UDZ') % 2nd-order, continuous-time model with underdamped poles

compare(gfr,m1,ms,mproc)

m1 =
Discrete-time OE model: y(t) = [B(z)/F(z)]u(t) + e(t)
B(z) = 0.9986 z^-1 + 0.4968 z^-2

F(z) = 1 - 1.499 z^-1 + 0.6998 z^-2

Sample time: 0.1 seconds

Parameterization:
Polynomial orders: nb=2 nf=2 nk=1
Number of free coefficients: 4
Use "polydata", "getpvec", "getcov" for parameters and their uncertainties.

Status:
Estimated using OE on frequency response data "gfr".
Fit to estimation data: 88.04%
FPE: 0.2492, MSE: 0.2492

ms =
Continuous-time identified state-space model:
dx/dt = A x(t) + B u(t) + K e(t)
y(t) = C x(t) + D u(t) + e(t)

A =
x1 x2
x1 -0.9009 6.635
x2 -3.307 -2.668

B =
u1
x1 -32.9
x2 -28.33

C =
x1 x2
y1 -0.5073 0.499

D =
u1
y1 0

3-170
 Frequency Domain Identification: Estimating Models Using Frequency Domain Data

K =
y1
x1 0
x2 0

Parameterization:
FREE form (all coefficients in A, B, C free).
Feedthrough: none
Disturbance component: none
Number of free coefficients: 8
Use "idssdata", "getpvec", "getcov" for parameters and their uncertainties.

Status:
Estimated using SSEST on frequency response data "gfr".
Fit to estimation data: 73.16%
FPE: 0.2502, MSE: 1.254

mproc =
Process model with transfer function:
1+Tz*s
G(s) = Kp * ---------------------- * exp(-Td*s)
1+2*Zeta*Tw*s+(Tw*s)^2

Kp = 7.4619
Tw = 0.20245
Zeta = 0.36242
Td = 0
Tz = 0.013617

Parameterization:
'P2DUZ'
Number of free coefficients: 5
Use "getpvec", "getcov" for parameters and their uncertainties.

Status:
Estimated using PROCEST on frequency response data "gfr".
Fit to estimation data: 73.2%
FPE: 0.2495, MSE: 1.25

3-171
3 Linear Model Identification

As shown above a variety of linear model types may be estimated in both continuous and
discrete time domains, using spectral data. These models may be validated using, time-
domain data. The time-domain I/O data set ztime, for example, is collected from the
same system, and can be used for validation of m1, ms and mproc:

compare(ztime,m1,ms,mproc) %validation in a different domain

3-172
 Frequency Domain Identification: Estimating Models Using Frequency Domain Data

We may also look at the residuals to affirm the quality of the model using the validation
data ztime. As observed, the residuals are almost white:

resid(mproc,ztime) % Residuals plot

3-173
3 Linear Model Identification

Condensing Data Using SPAFDR

An important reason to work with frequency response data is that it is easy to condense
the information with little loss. The command SPAFDR allows you to compute smoothed
response data over limited frequencies, for example with logarithmic spacing. Here is
an example where the gfr data is condensed to 100 logarithmically spaced frequency
values. With a similar technique, also the original time domain data can be condensed:

sgfr = spafdr(gfr) % spectral estimation with frequency-dependent resolution

sz = spafdr(ztime); % spectral estimation using time-domain data
clf
bode(gfr,sgfr,sz)
axis([pi/100 10*pi, -272 105])
legend('gfr (raw data)','sgfr','sz','location','southwest')

3-174
 Frequency Domain Identification: Estimating Models Using Frequency Domain Data

sgfr =
IDFRD model.
Contains Frequency Response Data for 1 output(s) and 1 input(s), and the spectra for di
Response data and disturbance spectra are available at 100 frequency points, ranging fr

Sample time: 0.1 seconds

Output channels: 'y1'
Input channels: 'u1'
Status:
Estimated using SPAFDR on frequency response data "gfr".

3-175
3 Linear Model Identification

The Bode plots show that the information in the smoothed data has been taken well care
of. Now, these data records with 100 points can very well be used for model estimation.
For example:
msm = oe(sgfr,[2 2 1]);
compare(ztime,msm,m1) % msm has the same accuracy as M1 (based on 1000 points)

Estimation Using Frequency-Domain I/O Data

It may be that the measurements are available as Fourier transforms of inputs and
output. Such frequency domain data from the system are given as the signals Y and U. In
loglog plots they look like
Wfd = (0:500)'*10*pi/500;
subplot(211),loglog(Wfd,abs(Y)),title('The amplitude of the output')

3-176
 Frequency Domain Identification: Estimating Models Using Frequency Domain Data

subplot(212),loglog(Wfd,abs(U)),title('The amplitude of the input')

The frequency response data is essentially the ratio between Y and U. To collect the
frequency domain data as an IDDATA object, do as follows:
ZFD = iddata(Y,U,'ts',0.1,'Domain','Frequency','Freq',Wfd)

ZFD =

Frequency domain data set with responses at 501 frequencies,

ranging from 0 to 31.416 rad/seconds
Sample time: 0.1 seconds

Outputs Unit (if specified)

3-177
3 Linear Model Identification

y1

Inputs Unit (if specified)

u1

Now, again the frequency domain data set ZFD can be used as data in all estimation
routines, just as time domain data and frequency response data:
mf = ssest(ZFD) % SSEST picks best order in 1:10 range when called this way
mfr = ssregest(ZFD) % an alternative regularized reduction based state-space estimator
compare(ztime,mf,mfr,m1)

mf =
Continuous-time identified state-space model:
dx/dt = A x(t) + B u(t) + K e(t)
y(t) = C x(t) + D u(t) + e(t)

A =
x1 x2
x1 -1.501 6.791
x2 -3.115 -2.059

B =
u1
x1 -28.11
x2 -33.39

C =
x1 x2
y1 -0.5844 0.4129

D =
u1
y1 0

K =
y1
x1 0
x2 0

Parameterization:
FREE form (all coefficients in A, B, C free).
Feedthrough: none

3-178
 Frequency Domain Identification: Estimating Models Using Frequency Domain Data

Disturbance component: none

Number of free coefficients: 8
Use "idssdata", "getpvec", "getcov" for parameters and their uncertainties.

Status:
Estimated using SSEST on frequency domain data "ZFD".
Fit to estimation data: 97.21%
FPE: 0.04263, MSE: 0.04179

mfr =
Discrete-time identified state-space model:
x(t+Ts) = A x(t) + B u(t) + K e(t)
y(t) = C x(t) + D u(t) + e(t)

A =
x1 x2 x3 x4 x5 x6
x1 0.6552 0.6194 -0.212 0.001125 0.05928 0.08817
x2 -0.189 0.3955 0.6197 0.1951 0.2022 -0.08355
x3 -0.3024 0.6054 -0.3126 0.02585 -0.1829 -0.162
x4 -0.06436 0.1201 0.07096 -0.679 0.5621 -0.06486
x5 -0.05577 0.00528 -0.09426 -0.08431 0.1222 0.5888
x6 0.005306 -0.05481 0.3242 -0.3403 -0.2923 -0.2853
x7 -0.0005408 0.1019 0.02489 -0.09813 -0.3741 0.05454
x8 0.01313 0.06723 -0.03633 0.07937 -0.2934 0.3659
x9 -0.01436 0.0138 -0.133 0.1117 0.259 -0.01073
x10 0.007467 -0.09578 0.09325 0.0836 0.2193 -0.5335

x7 x8 x9 x10
x1 0.06715 0.06958 0.06185 0.1452
x2 0.1775 0.1755 0.03826 0.05137
x3 -0.2076 0.06537 -0.17 0.06656
x4 -0.2574 -0.1622 0.0801 -0.3327
x5 0.7376 0.2966 -0.4042 0.499
x6 -0.06971 0.3705 -0.11 0.292
x7 0.1632 0.1927 0.3252 -0.7328
x8 -0.4043 -0.02296 -0.4008 -0.2135
x9 -0.463 0.822 -0.0677 0.02849
x10 0.1881 0.2637 -0.2241 -0.6783

B =
u1
x1 1.555
x2 0.3412
x3 -0.01226

3-179
3 Linear Model Identification

x4 0.01437
x5 0.08268
x6 0.03217
x7 0.06806
x8 0.02006
x9 -0.04152
x10 0.02122

C =
x1 x2 x3 x4 x5 x6 x7
y1 1.18 -2.337 -0.7734 -0.2832 0.2353 0.8738 -0.6108

x8 x9 x10
y1 0.653 1.103 -0.5883

D =
u1
y1 0

K =
y1
x1 -0.02321
x2 -0.07354
x3 -0.08669
x4 0.009556
x5 -0.03322
x6 -0.03809
x7 0.05905
x8 -0.02889
x9 0.009478
x10 0.0161

Sample time: 0.1 seconds

Parameterization:
FREE form (all coefficients in A, B, C free).
Feedthrough: none
Disturbance component: estimate
Number of free coefficients: 130
Use "idssdata", "getpvec", "getcov" for parameters and their uncertainties.

Status:
Estimated using SSREGEST on frequency domain data "ZFD".
Fit to estimation data: 66.73% (prediction focus)

3-180
 Frequency Domain Identification: Estimating Models Using Frequency Domain Data

FPE: 6.697, MSE: 5.953

Transformations Between Data Representations (Time - Frequency)

Time and frequency domain input-output data sets can be transformed to either domain
by using FFT and IFFT. These commands are adapted to IDDATA objects:

dataf = fft(ztime)
datat = ifft(dataf)

dataf =

3-181
3 Linear Model Identification

Frequency domain data set with responses at 501 frequencies,

ranging from 0 to 31.416 rad/seconds
Sample time: 0.1 seconds

Outputs Unit (if specified)

y1

Inputs Unit (if specified)

u1

datat =

Time domain data set with 1000 samples.

Sample time: 0.1 seconds

Outputs Unit (if specified)

y1

Inputs Unit (if specified)

u1

Time and frequency domain input-output data can be transformed to frequency response
data by SPAFDR, SPA and ETFE:

g1 = spafdr(ztime)
g2 = spafdr(ZFD);
clf;
bode(g1,g2)

g1 =
IDFRD model.
Contains Frequency Response Data for 1 output(s) and 1 input(s), and the spectra for di
Response data and disturbance spectra are available at 100 frequency points, ranging fr

Sample time: 0.1 seconds

Output channels: 'y1'
Input channels: 'u1'
Status:
Estimated using SPAFDR on time domain data "ztime".

3-182
 Frequency Domain Identification: Estimating Models Using Frequency Domain Data

Frequency response data can also be transformed to more smoothed data (less resolution
and less data) by SPAFDR and SPA;

g3 = spafdr(gfr);

Frequency response data can be transformed to frequency domain input-output signals

by the command IDDATA:

gfd = iddata(g3)

gfd =

Frequency domain data set with responses at 100 frequencies,

3-183
3 Linear Model Identification

ranging from 0.031416 to 31.416 rad/seconds

Sample time: 0.1 seconds

Outputs Unit (if specified)

y1

Inputs Unit (if specified)

u1

Using Continuous-time Frequency-domain Data to Estimate Continuous-time Models

Time domain data can naturally only be stored and dealt with as discrete-time, sampled
data. Frequency domain data have the advantage that continuous time data can be
represented correctly. Suppose that the underlying continuous time signals have no
frequency information above the Nyquist frequency, e.g. because they are sampled
fast, or the input has no frequency component above the Nyquist frequency. Then the
Discrete Fourier transforms (DFT) of the data also are the Fourier transforms of the
continuous time signals, at the chosen frequencies. They can therefore be used to directly
fit continuous time models. In fact, this is the correct way of using band-limited data for
model fit.

This will be illustrated by the following example.

Consider the continuous time system:

m0 = idpoly(1,1,1,1,[1 1 1],'ts',0)

m0 =
Continuous-time OE model: y(t) = [B(s)/F(s)]u(t) + e(t)
B(s) = 1

F(s) = s^2 + s + 1

Parameterization:
Polynomial orders: nb=1 nf=2 nk=0
Number of free coefficients: 3

3-184
 Frequency Domain Identification: Estimating Models Using Frequency Domain Data

Use "polydata", "getpvec", "getcov" for parameters and their uncertainties.

Status:
Created by direct construction or transformation. Not estimated.

Choose an input with low frequency contents that is fast sampled:

rng(235,'twister');
u = idinput(500,'sine',[0 0.2]);
u = iddata([],u,0.1,'intersamp','bl');

0.1 is the sampling interval, and 'bl' indicates that the input is band-limited, i.e.
in continuous time it consists of sinusoids with frequencies below half the sampling
frequency. Correct simulation of such a system should be done in the frequency domain:

uf = fft(u);
uf.ts = 0; % Denoting that the data is continuous time
yf = sim(m0,uf);
%
% Add some noise to the data:
yf.y = yf.y + 0.05*(randn(size(yf.y))+1i*randn(size(yf.y)));
dataf = [yf uf] % This is now a continuous time frequency domain data set.

dataf =

Frequency domain data set with responses at 251 frequencies,

ranging from 0 to 31.416 rad/seconds
Sample time: 0 seconds

Outputs Unit (if specified)

y1

Inputs Unit (if specified)

u1

Look at the data:

plot(dataf)
axis([0 10 0 2.5])

3-185
3 Linear Model Identification

Using dataf for estimation will by default give continuous time models: State-space:

m4 = ssest(dataf,2); %Second order continuous-time model

For a polynomial model with nb = 2 numerator coefficient and nf = 2 estimated

denominator coefficients use:

nb = 2;
nf = 2;
m5 = oe(dataf,[nb nf])

m5 =
Continuous-time OE model: y(t) = [B(s)/F(s)]u(t) + e(t)

3-186
 Frequency Domain Identification: Estimating Models Using Frequency Domain Data

B(s) = -0.01761 s + 1

F(s) = s^2 + 0.9873 s + 0.9902

Parameterization:
Polynomial orders: nb=2 nf=2 nk=0
Number of free coefficients: 4
Use "polydata", "getpvec", "getcov" for parameters and their uncertainties.

Status:
Estimated using OE on frequency domain data "dataf".
Fit to estimation data: 70.15%
FPE: 0.00482, MSE: 0.004713

Compare step responses with uncertainty of the true system m0 and the models m4 and
m5. The confidence intervals are shown with patches in the plot.

clf
h = stepplot(m0,m4,m5);
showConfidence(h,1)
legend('show','location','southeast')

3-187
3 Linear Model Identification

Although it was not necessary in this case, it is generally advised to focus the fit to a
limited frequency band (low pass filter the data) when estimating using continuous time
data. The system has a bandwidth of about 3 rad/s, and was excited by sinusoids up to
6.2 rad/s. A reasonable frequency range to focus the fit to is then [0 7] rad/s:

m6 = ssest(dataf,2,ssestOptions('Focus',[0 7])) % state space model

m6 =
Continuous-time identified state-space model:
dx/dt = A x(t) + B u(t) + K e(t)
y(t) = C x(t) + D u(t) + e(t)

A =

3-188
 Frequency Domain Identification: Estimating Models Using Frequency Domain Data

x1 x2
x1 -0.2714 -1.809
x2 0.4407 -0.7282

B =
u1
x1 0.5873
x2 -0.3025

C =
x1 x2
y1 0.7565 1.526

D =
u1
y1 0

K =
y1
x1 0
x2 0

Parameterization:
FREE form (all coefficients in A, B, C free).
Feedthrough: none
Disturbance component: none
Number of free coefficients: 8
Use "idssdata", "getpvec", "getcov" for parameters and their uncertainties.

Status:
Estimated using SSEST on frequency domain data "dataf".
Fit to estimation data: 87.08% (filter focus)
FPE: 0.004213, MSE: 0.003682

m7 = oe(dataf,[1 2],oeOptions('Focus',[0 7])) % polynomial model of Output Error struct

m7 =
Continuous-time OE model: y(t) = [B(s)/F(s)]u(t) + e(t)
B(s) = 0.9866

F(s) = s^2 + 0.9791 s + 0.9761

Parameterization:
Polynomial orders: nb=1 nf=2 nk=0

3-189
3 Linear Model Identification

Number of free coefficients: 3

Use "polydata", "getpvec", "getcov" for parameters and their uncertainties.

Status:
Estimated using OE on frequency domain data "dataf".
Fit to estimation data: 87.04% (filter focus)
FPE: 0.004008, MSE: 0.003708

opt = procestOptions('SearchMethod','lsqnonlin','Focus',[0 7]); % Requires Optimization

m8 = procest(dataf,'P2UZ',opt) % process model with underdamped poles

m8 =
Process model with transfer function:
1+Tz*s
G(s) = Kp * ----------------------
1+2*Zeta*Tw*s+(Tw*s)^2

Kp = 1.0124
Tw = 1.0019
Zeta = 0.5021
Tz = -0.017474

Parameterization:
'P2UZ'
Number of free coefficients: 4
Use "getpvec", "getcov" for parameters and their uncertainties.

Status:
Estimated using PROCEST on frequency domain data "dataf".
Fit to estimation data: 87.08% (filter focus)
FPE: 0.003947, MSE: 0.003682

opt = tfestOptions('SearchMethod','lsqnonlin','Focus',[0 7]); % Requires Optimization T

m9 = tfest(dataf,2,opt) % transfer function with 2 poles

m9 =

From input "u1" to output "y1":

-0.01647 s + 1.003
-----------------------
s^2 + 0.9949 s + 0.9922

Continuous-time identified transfer function.

3-190
 Frequency Domain Identification: Estimating Models Using Frequency Domain Data

Parameterization:
Number of poles: 2 Number of zeros: 1
Number of free coefficients: 4
Use "tfdata", "getpvec", "getcov" for parameters and their uncertainties.

Status:
Estimated using TFEST on frequency domain data "dataf".
Fit to estimation data: 87.08% (filter focus)
FPE: 0.004067, MSE: 0.003684

h = stepplot(m0,m6,m7,m8,m9);
showConfidence(h,1)
legend('show')

3-191
3 Linear Model Identification

Conclusions

We saw how time, frequency and spectral data can seamlessly be used to estimate a
variety of linear models in both continuous and discrete time domains. The models may
be validated and compared in domains different from the ones they were estimated in.
The data formats (time, frequency and spectrum) are interconvertible, using methods
such as fft, ifft, spafdr and spa. Furthermore, direct, continuous-time estimation
is achievable by using tfest, ssest and procest estimation routines. The seamless
use of data in any domain for estimation and analysis is an important feature of System
Identification Toolbox.

Additional Information

For more information on identification of dynamic systems with System Identification

Toolbox visit the System Identification Toolbox product information page.

3-192
 Building Structured and User-Defined Models Using System Identification Toolbox™

Building Structured and User-Defined Models Using System

Identification Toolbox™
This example shows how to estimate parameters in user-defined model structures. Such
structures are specified by IDGREY (linear state-space) or IDNLGREY (nonlinear state-
space) models. We shall investigate how to assign structure, fix parameters and create
dependencies among them.

Experiment Data

We shall investigate data produced by a (simulated) dc-motor. We first load the data:

load dcmdata
who

Your variables are:

text u y

The matrix y contains the two outputs: y1 is the angular position of the motor shaft and
y2 is the angular velocity. There are 400 data samples and the sampling interval is 0.1
seconds. The input is contained in the vector u. It is the input voltage to the motor.

z = iddata(y,u,0.1); % The IDDATA object

z.InputName = 'Voltage';
z.OutputName = {'Angle';'AngVel'};
plot(z(:,1,:))

3-193
3 Linear Model Identification

Figure: Measurement Data: Voltage to Angle

plot(z(:,2,:))

3-194
 Building Structured and User-Defined Models Using System Identification Toolbox™

Figure: Measurement Data: Voltage to Angle

Model Structure Selection

d/dt x = A x + B u + K e
y = C x + D u + e

We shall build a model of the dc-motor. The dynamics of the motor is well known. If we
choose x1 as the angular position and x2 as the angular velocity it is easy to set up a
state-space model of the following character neglecting disturbances: (see Example 4.1 in
Ljung(1999):

| 0 1 | | 0 |
d/dt x = | | x + | | u

3-195
3 Linear Model Identification

| 0 -th1 | | th2 |

| 1 0 |
y = | | x
| 0 1 |

The parameter th1 is here the inverse time-constant of the motor and th2 is such that
th2/th1 is the static gain from input to the angular velocity. (See Ljung(1987) for how
th1 and th2 relate to the physical parameters of the motor). We shall estimate these
two parameters from the observed data. The model structure (parameterized state space)
described above can be represented in MATLAB® using IDSS and IDGREY models.
These models let you perform estimation of parameters using experimental data.

Specification of a Nominal (Initial) Model

If we guess that th1=1 and th2 = 0.28 we obtain the nominal or initial model

A = [0 1; 0 -1]; % initial guess for A(2,2) is -1

B = [0; 0.28]; % initial guess for B(2) is 0.28
C = eye(2);
D = zeros(2,1);

and we package this into an IDSS model object:

ms = idss(A,B,C,D);

The model is characterized by its matrices, their values, which elements are free (to be
estimated) and upper and lower limits of those:

ms.Structure.a

ans =

Name: 'a'
Value: [2x2 double]
Minimum: [2x2 double]
Maximum: [2x2 double]
Free: [2x2 logical]
Scale: [2x2 double]
Info: [2x2 struct]

1x1 param.Continuous

3-196
 Building Structured and User-Defined Models Using System Identification Toolbox™

ms.Structure.a.Value
ms.Structure.a.Free

ans =

0 1
0 -1

ans =

1 1
1 1

Specification of Free (Independent) Parameters Using IDSS Models

So we should now mark that it is only A(2,2) and B(2,1) that are free parameters to be
estimated.

ms.Structure.a.Free = [0 0; 0 1];
ms.Structure.b.Free = [0; 1];
ms.Structure.c.Free = 0; % scalar expansion used
ms.Structure.d.Free = 0;
ms.Ts = 0; % This defines the model to be continuous

The Initial Model

ms % Initial model

ms =
Continuous-time identified state-space model:
dx/dt = A x(t) + B u(t) + K e(t)
y(t) = C x(t) + D u(t) + e(t)

A =
x1 x2
x1 0 1
x2 0 -1

B =

3-197
3 Linear Model Identification

u1
x1 0
x2 0.28

C =
x1 x2
y1 1 0
y2 0 1

D =
u1
y1 0
y2 0

K =
y1 y2
x1 0 0
x2 0 0

Parameterization:
STRUCTURED form (some fixed coefficients in A, B, C).
Feedthrough: none
Disturbance component: none
Number of free coefficients: 2
Use "idssdata", "getpvec", "getcov" for parameters and their uncertainties.

Status:
Created by direct construction or transformation. Not estimated.

Estimation of Free Parameters of the IDSS Model

The prediction error (maximum likelihood) estimate of the parameters is now computed
by:
dcmodel = ssest(z,ms,ssestOptions('Display','on'));
dcmodel

dcmodel =
Continuous-time identified state-space model:
dx/dt = A x(t) + B u(t) + K e(t)
y(t) = C x(t) + D u(t) + e(t)

A =
x1 x2

3-198
 Building Structured and User-Defined Models Using System Identification Toolbox™

x1 0 1
x2 0 -4.013

B =
Voltage
x1 0
x2 1.002

C =
x1 x2
Angle 1 0
AngVel 0 1

D =
Voltage
Angle 0
AngVel 0

K =
Angle AngVel
x1 0 0
x2 0 0

Parameterization:
STRUCTURED form (some fixed coefficients in A, B, C).
Feedthrough: none
Disturbance component: none
Number of free coefficients: 2
Use "idssdata", "getpvec", "getcov" for parameters and their uncertainties.

Status:
Estimated using SSEST on time domain data "z".
Fit to estimation data: [98.35;84.42]%
FPE: 0.001071, MSE: 0.1192

The estimated values of the parameters are quite close to those used when the data were
simulated (-4 and 1). To evaluate the model's quality we can simulate the model with the
actual input by and compare it with the actual output.

compare(z,dcmodel);

3-199
3 Linear Model Identification

We can now, for example plot zeros and poles and their uncertainty regions. We will
draw the regions corresponding to 3 standard deviations, since the model is quite
accurate. Note that the pole at the origin is absolutely certain, since it is part of the
model structure; the integrator from angular velocity to position.

clf
showConfidence(iopzplot(dcmodel),3)

3-200
 Building Structured and User-Defined Models Using System Identification Toolbox™

Now, we may make various modifications. The 1,2-element of the A-matrix (fixed to 1)
tells us that x2 is the derivative of x1. Suppose that the sensors are not calibrated, so
that there may be an unknown proportionality constant. To include the estimation of
such a constant we just "let loose" A(1,2) and re-estimate:
dcmodel2 = dcmodel;
dcmodel2.Structure.a.Free(1,2) = 1;
dcmodel2 = pem(z,dcmodel2,ssestOptions('Display','on'));

The resulting model is

dcmodel2

dcmodel2 =

3-201
3 Linear Model Identification

Continuous-time identified state-space model:

dx/dt = A x(t) + B u(t) + K e(t)
y(t) = C x(t) + D u(t) + e(t)

A =
x1 x2
x1 0 0.9975
x2 0 -4.011

B =
Voltage
x1 0
x2 1.004

C =
x1 x2
Angle 1 0
AngVel 0 1

D =
Voltage
Angle 0
AngVel 0

K =
Angle AngVel
x1 0 0
x2 0 0

Parameterization:
STRUCTURED form (some fixed coefficients in A, B, C).
Feedthrough: none
Disturbance component: none
Number of free coefficients: 3
Use "idssdata", "getpvec", "getcov" for parameters and their uncertainties.

Status:
Estimated using SSEST on time domain data "z".
Fit to estimation data: [98.35;84.42]%
FPE: 0.001076, MSE: 0.1192

We find that the estimated A(1,2) is close to 1. To compare the two model we use the
compare command:
compare(z,dcmodel,dcmodel2)

3-202
 Building Structured and User-Defined Models Using System Identification Toolbox™

Specification of Models with Coupled Parameters Using IDGREY Objects

Suppose that we accurately know the static gain of the dc-motor (from input voltage to
angular velocity, e.g. from a previous step-response experiment. If the static gain is G,
and the time constant of the motor is t, then the state-space model becomes

|0 1| | 0 |
d/dt x = | |x + | | u
|0 -1/t| | G/t |

|1 0|
y = | | x
|0 1|

3-203
3 Linear Model Identification

With G known, there is a dependence between the entries in the different matrices. In
order to describe that, the earlier used way with "Free" parameters will not be sufficient.
We thus have to write a MATLAB file which produces the A, B, C, and D, and optionally
also the K and X0 matrices as outputs, for each given parameter vector as input. It also
takes auxiliary arguments as inputs, so that the user can change certain things in the
model structure, without having to edit the file. In this case we let the known static gain
G be entered as such an argument. The file that has been written has the name 'motor.m'.

type motor

function [A,B,C,D,K,X0] = motor(par,ts,aux)

%MOTOR ODE file representing the dynamics of a motor.
%
% [A,B,C,D,K,X0] = MOTOR(Tau,Ts,G)
% returns the State Space matrices of the DC-motor with
% time-constant Tau (Tau = par) and known static gain G. The sample
% time is Ts.
%
% This file returns continuous-time representation if input argument Ts
% is zero. If Ts>0, a discrete-time representation is returned. To make
% the IDGREY model that uses this file aware of this flexibility, set the
% value of Structure.FcnType property to 'cd'. This flexibility is useful
% for conversion between continuous and discrete domains required for
% estimation and simulation.
%
% See also IDGREY, IDDEMO7.

% L. Ljung
% Copyright 1986-2011 The MathWorks, Inc.

t = par(1);
G = aux(1);

A = [0 1;0 -1/t];
B = [0;G/t];
C = eye(2);
D = [0;0];
K = zeros(2);
X0 = [0;0];
if ts>0 % Sample the model with sample time Ts
s = expm([[A B]*ts; zeros(1,3)]);
A = s(1:2,1:2);
B = s(1:2,3);

3-204
 Building Structured and User-Defined Models Using System Identification Toolbox™

end

We now create an IDGREY model object corresponding to this model structure: The
assumed time constant will be

par_guess = 1;

We also give the value 0.25 to the auxiliary variable G (gain) and sampling interval.

aux = 0.25;
dcmm = idgrey('motor',par_guess,'cd',aux,0);

The time constant is now estimated by

dcmm = greyest(z,dcmm,greyestOptions('Display','on'));

We have thus now estimated the time constant of the motor directly. Its value is in good
agreement with the previous estimate.

dcmm

dcmm =
Continuous-time linear grey box model defined by "motor" function:
dx/dt = A x(t) + B u(t) + K e(t)
y(t) = C x(t) + D u(t) + e(t)

A =
x1 x2
x1 0 1
x2 0 -4.006

B =
Voltage
x1 0
x2 1.001

C =
x1 x2
Angle 1 0
AngVel 0 1

D =
Voltage

3-205
3 Linear Model Identification

Angle 0
AngVel 0

K =
Angle AngVel
x1 0 0
x2 0 0

Model parameters:
Par1 = 0.2496

Parameterization:
ODE Function: motor
(parameterizes both continuous- and discrete-time equations)
Disturbance component: parameterized by the ODE function
Initial state: parameterized by the ODE function
Number of free coefficients: 1
Use "getpvec", "getcov" for parameters and their uncertainties.

Status:
Estimated using GREYEST on time domain data "z".
Fit to estimation data: [98.35;84.42]%
FPE: 0.00107, MSE: 0.1193

With this model we can now proceed to test various aspects as before. The syntax of all
the commands is identical to the previous case. For example, we can compare the idgrey
model with the other state-space model:

compare(z,dcmm,dcmodel)

3-206
 Building Structured and User-Defined Models Using System Identification Toolbox™

They are clearly very close.

Estimating Multivariate ARX Models

The state-space part of the toolbox also handles multivariate (several outputs) ARX
models. By a multivariate ARX-model we mean the following:

A(q) y(t) = B(q) u(t) + e(t)

Here A(q) is a ny | ny matrix whose entries are polynomials in the delay operator 1/q.
The k-l element is denoted by:

3-207
3 Linear Model Identification

where:

It is thus a polynomial in 1/q of degree nakl.

Similarly B(q) is a ny | nu matrix, whose kj-element is:

There is thus a delay of nkkj from input number j to output number k. The most
common way to create those would be to use the ARX-command. The orders are specified
as: nn = [na nb nk] with na being a ny-by-ny matrix whose kj-entry is nakj; nb
and nk are defined similarly.

Let's test some ARX-models on the dc-data. First we could simply build a general second
order model:
dcarx1 = arx(z,'na',[2,2;2,2],'nb',[2;2],'nk',[1;1])

dcarx1 =
Discrete-time ARX model:
Model for output "Angle": A(z)y_1(t) = - A_i(z)y_i(t) + B(z)u(t) + e_1(t)
A(z) = 1 - 0.5545 z^-1 - 0.4454 z^-2

A_2(z) = -0.03548 z^-1 - 0.06405 z^-2

B(z) = 0.004243 z^-1 + 0.006589 z^-2

Model for output "AngVel": A(z)y_2(t) = - A_i(z)y_i(t) + B(z)u(t) + e_2(t)

A(z) = 1 - 0.2005 z^-1 - 0.2924 z^-2

A_1(z) = 0.01849 z^-1 - 0.01937 z^-2

B(z) = 0.08642 z^-1 + 0.03877 z^-2

Sample time: 0.1 seconds

Parameterization:
Polynomial orders: na=[2 2;2 2] nb=[2;2] nk=[1;1]
Number of free coefficients: 12
Use "polydata", "getpvec", "getcov" for parameters and their uncertainties.

3-208
 Building Structured and User-Defined Models Using System Identification Toolbox™

Status:
Estimated using ARX on time domain data "z".
Fit to estimation data: [97.87;83.44]% (prediction focus)
FPE: 0.002197, MSE: 0.1398

The result, dcarx1, is stored as an IDPOLY model, and all previous commands apply.
We could for example explicitly list the ARX-polynomials by:
dcarx1.a

ans =

[1x3 double] [1x3 double]

[1x3 double] [1x3 double]

as cell arrays where e.g. the {1,2} element of dcarx1.a is the polynomial A(1,2) described
earlier, relating y2 to y1.

We could also test a structure, where we know that y1 is obtained by filtering y2 through
a first order filter. (The angle is the integral of the angular velocity). We could then also
postulate a first order dynamics from input to output number 2:
na = [1 1; 0 1];
nb = [0 ; 1];
nk = [1 ; 1];
dcarx2 = arx(z,[na nb nk])

dcarx2 =
Discrete-time ARX model:
Model for output "Angle": A(z)y_1(t) = - A_i(z)y_i(t) + B(z)u(t) + e_1(t)
A(z) = 1 - 0.9992 z^-1

A_2(z) = -0.09595 z^-1

B(z) = 0

Model for output "AngVel": A(z)y_2(t) = B(z)u(t) + e_2(t)

A(z) = 1 - 0.6254 z^-1

B(z) = 0.08973 z^-1

3-209
3 Linear Model Identification

Sample time: 0.1 seconds

Parameterization:
Polynomial orders: na=[1 1;0 1] nb=[0;1] nk=[1;1]
Number of free coefficients: 4
Use "polydata", "getpvec", "getcov" for parameters and their uncertainties.

Status:
Estimated using ARX on time domain data "z".
Fit to estimation data: [97.52;81.46]% (prediction focus)
FPE: 0.003468, MSE: 0.177

To compare the different models obtained we use

compare(z,dcmodel,dcmm,dcarx2)

3-210
 Building Structured and User-Defined Models Using System Identification Toolbox™

Finally, we could compare the bodeplots obtained from the input to output one for the
different models by using bode: First output:

dcmm2 = idss(dcmm); % convert to IDSS for subreferencing

bode(dcmodel(1,1),'r',dcmm2(1,1),'b',dcarx2(1,1),'g')

Second output:

bode(dcmodel(2,1),'r',dcmm2(2,1),'b',dcarx2(2,1),'g')

3-211
3 Linear Model Identification

The two first models are more or less in exact agreement. The ARX-models are not
so good, due to the bias caused by the non-white equation error noise. (We had white
measurement noise in the simulations).

Conclusions

Estimation of models with pre-selected structures can be performed using System

Identification toolbox. In state-space form, parameters may be fixed to their known
values or constrained to lie within a prescribed range. If relationship between
parameters or other constraints need to be specified, IDGREY objects may be used.
IDGREY models evaluate a user-specified MATLAB file for estimating state-space
system parameters. Multi-variate ARX models offer another option for quickly estimating
multi-output models with user-specified structure.

3-212
 Building Structured and User-Defined Models Using System Identification Toolbox™

Additional Information

For more information on identification of dynamic systems with System Identification

Toolbox visit the System Identification Toolbox product information page.

3-213
3-214
 4

Identifying State-Space Models

• “What Are State-Space Models?” on page 4-2

• “Data Supported by State-Space Models” on page 4-5
• “Supported State-Space Parameterizations” on page 4-6
• “Estimate State-Space Model With Order Selection” on page 4-7
• “Estimate State-Space Models in System Identification App” on page 4-12
• “Estimate State-Space Models at the Command Line” on page 4-21
• “Estimate State-Space Models with Free-Parameterization” on page 4-27
• “Estimate State-Space Models with Canonical Parameterization” on page 4-28
• “Estimate State-Space Models with Structured Parameterization” on page 4-30
• “Estimate the State-Space Equivalent of ARMAX and OE Models” on page 4-37
• “Assigning Estimation Weightings” on page 4-39
• “Specifying Initial States for Iterative Estimation Algorithms” on page 4-40
• “State-Space Model Estimation Methods” on page 4-41
4 Identifying State-Space Models

What Are State-Space Models?

In this section...
“Definition of State-Space Models” on page 4-2
“Continuous-Time Representation” on page 4-2
“Discrete-Time Representation” on page 4-3
“Relationship Between Continuous-Time and Discrete-Time State Matrices” on page
4-3
“State-Space Representation of Transfer Functions” on page 4-4

Definition of State-Space Models

State-space models are models that use state variables to describe a system by a set of
first-order differential or difference equations, rather than by one or more nth-order
differential or difference equations. State variables x(t) can be reconstructed from the
measured input-output data, but are not themselves measured during an experiment.

The state-space model structure is a good choice for quick estimation because it requires
you to specify only one input, the model order, n. The model order is an integer equal to
the dimension of x(t) and relates to, but is not necessarily equal to, the number of delayed
inputs and outputs used in the corresponding linear difference equation.

Continuous-Time Representation
It is often easier to define a parameterized state-space model in continuous time because
physical laws are most often described in terms of differential equations. In continuous-
time, the state-space description has the following form:

% (t)
x& (t) = Fx(t) + Gu(t) + Kw
y( t) = Hx(t) + Du( t) + w(t)
x(0) = x0

The matrices F, G, H, and D contain elements with physical significance—for example,

material constants. x0 specifies the initial states.

4-2
 What Are State-Space Models?

Note: K% = 0 gives the state-space representation of an Output-Error model. For more

information, see “What Are Polynomial Models?” on page 3-40.

You can estimate continuous-time state-space model using both time- and frequency-
domain data.

Discrete-Time Representation
The discrete-time state-space model structure is often written in the innovations form
that describes noise:

x(kT + T ) = Ax( kT ) + Bu(kT ) + Ke( kT )

y( kT ) = Cx( kT ) + Du(kT ) + e( kT )
x(0) = x0

where T is the sampling interval, u(kT) is the input at time instant kT, and y(kT) is the
output at time instant kT.

Note: K=0 gives the state-space representation of an Output-Error model. For more
information about Output-Error models, see “What Are Polynomial Models?” on page
3-40.

Discrete-time state-space models provide the same type of linear difference relationship
between the inputs and outputs as the “linear ARX model”, but are rearranged such that
there is only one delay in the expressions.

You cannot estimate a discrete-time state-space model using continuous-time frequency-

domain data.

Relationship Between Continuous-Time and Discrete-Time State Matrices

The relationships between the discrete state-space matrices A, B, C, D, and K and
the continuous-time state-space matrices F, G, H, D, and K% are given for piece-wise-
constant input, as follows:

4-3
4 Identifying State-Space Models

A = e FT
T
Ft
B= Úe Gdt
0
C=H

These relationships assume that the input is piece-wise-constant over time intervals
kT £ t < (k + 1) T .

The exact relationship between K and K% is complicated. However, for short sampling
intervals T, the following approximation works well:

T
Ft
K= Úe % t
Kd
0

State-Space Representation of Transfer Functions

For linear models, the general model description is given by:

y = Gu + He

G is a transfer function that takes the input u to the output y. H is a transfer function
that describes the properties of the additive output noise model.

The relationships between the transfer functions and the discrete-time state-space
matrices are given by the following equations:

G ( q) = C( qInx - A) -1 B + D
H ( q) = C(qI nx - A) -1 K + I ny

where Inx is the nx-by-nx identity matrix, Iny is the nx-by-nx identity matrix, and ny is
the dimension of y and e. The state-space representation in the continuous-time case is
similar.

4-4
 Data Supported by State-Space Models

Data Supported by State-Space Models

You can estimate linear state-space models from data with the following characteristics:

• Time- or frequency-domain data

To estimate state-space models for time-series data, see “Time-Series Model

Identification”.
• Real data or complex data
• Single-output and multiple-output

To estimate state-space models, you must first import your data into the MATLAB
workspace, as described in “Data Preparation”.

4-5
4 Identifying State-Space Models

Supported State-Space Parameterizations

System Identification Toolbox software supports the following parameterizations that
indicate which parameters are estimated and which remain fixed at specific values:

• Free parameterization results in the estimation of all elements of the system

matrices A, B, C, D, and K. See “Estimate State-Space Models with Free-
Parameterization” on page 4-27.
• Canonical parameterization represents a state-space system in a reduced-
parameter form where many entries of the A, B and C matrices are fixed to zeros and
ones. The free parameters appear in only a few of the rows and columns in the system
matrices A, B, C and D. The software supports companion, modal decomposition and
observability canonical forms. See “Estimate State-Space Models with Canonical
Parameterization” on page 4-28.
• Structured parameterization lets you specify the fixed values of specific
parameters and exclude these parameters from estimation. You choose which entries
of the system matrices to estimate and which to treat as fixed. See “Estimate State-
Space Models with Structured Parameterization” on page 4-30.
• Completely arbitrary mapping of parameters to state-space matrices. See
“Estimating Linear Grey-Box Models”.

See Also
• “Estimate State-Space Models with Free-Parameterization” on page 4-27
• “Estimate State-Space Models with Canonical Parameterization” on page 4-28
• “Estimate State-Space Models with Structured Parameterization” on page 4-30

4-6
 Estimate State-Space Model With Order Selection

Estimate State-Space Model With Order Selection

In this section...
“Estimate Model With Selected Order in the App” on page 4-7
“Estimate Model With Selected Order at the Command Line” on page 4-10
“Using the Model Order Selection Window” on page 4-10

To estimate a state-space model, you must provide a value of its order, which represents
the number of states. When you do not know the order, you can search and select an
order using the following procedures.

Estimate Model With Selected Order in the App

You must have already imported your data into the app, as described in “Represent
Data”.

To estimate model orders for a specific input delay:

1 In the System Identification app, select Estimate > State Space Models to open
the State Space Models dialog box.
2 Select the Pick best value in the range option and specify a range in the adjacent
field. The default range is 1:10.

4-7
4 Identifying State-Space Models

3 (Optional) Expand Model Structure Configuration to specify additional

attributes of the model structure when searching for best orders. Such attributes
include disturbance component, input delays, presence of feedthrough, and
parameterization.
4 Expand Estimation Options and verify that Subspace (N4SID) is selected as the
Method.
5 Click Estimate.

This action opens the Model Order Selection window, which displays the relative
measure of how much each state contributes to the input-output behavior of the
model (log of singular values of the covariance matrix). The following figure shows an
example plot.

4-8
 Estimate State-Space Model With Order Selection

6 Select the rectangle that represents the cutoff for the states on the left that provide a
significant contribution to the input-output behavior.

In the previous figure, states 1 and 2 provide the most significant contribution. The
contributions to the right of state 2 drop significantly. Click Insert to estimate a
model with this order. Red indicates the recommended choice. For information about
using the Model Order Selection window, see “Using the Model Order Selection
Window” on page 4-10.

This action adds a new model to the Model Board in the System Identification app.
The default name of the model is ss1. You can use this model as an initial guess for
estimating other state-space models, as described in “Estimate State-Space Models
in System Identification App” on page 4-12.
7 Click Close to close the window.

4-9
4 Identifying State-Space Models

Estimate Model With Selected Order at the Command Line

You can estimate a state-space model with selected order using n4sid, ssest or
ssregest.

Use the following syntax to specify the range of model orders to try for a specific input
delay:

m = n4sid(data,n1:n2);

where data is the estimation data set, n1 and n2 specify the range of orders.

The command opens the Model Order Selection window. For information about using this
plot, see “Using the Model Order Selection Window” on page 4-10.

Alternatively, use ssest or ssregest:

m1 = ssest(data,nn)
m2 = ssregest(data,nn)

where nn = [n1,n2,...,nN] specifies the vector or range of orders you want to try.

n4sid and ssregest estimate a model whose sample time matches that of data
by default, hence a discrete-time model for time-domain data. ssest estimates a
continuous-time model by default. You can change the default setting by including the Ts
name-value pair input arguments in the estimation command. For example, to estimate a
discrete-time model of optimal order, assuming Data.Ts>0, type:

model = ssest(data,nn,'Ts',data.Ts);

or

model = ssregest(data,nn,'Ts',data.Ts);

To automatically select the best order without opening the Model Order Selection
window, type m = n4sid(data,'best'), m = ssest(data,'best') or m =
ssregest(data,'best').

Using the Model Order Selection Window

The following figure shows a sample Model Order Selection window.

4-10
 Estimate State-Space Model With Order Selection

You use this plot to decide which states provide a significant relative contribution to
the input-output behavior, and which states provide the smallest contribution. Based
on this plot, select the rectangle that represents the cutoff for the states on the left that
provide a significant contribution to the input-output behavior. The recommended choice
is shown in red. To learn how to generate this plot, see “Estimate Model With Selected
Order in the App” on page 4-7 or “Estimate Model With Selected Order at the
Command Line” on page 4-10.

The horizontal axis corresponds to the model order n. The vertical axis, called Log of
Singular values, shows the singular values of a covariance matrix constructed from the
observed data.

For example, in the previous figure, states 1 and 2 provide the most significant
contribution. However, the contributions of the states to the right of state 2 drop
significantly. This sharp decrease in the log of the singular values after n=2 indicates
that using two states is sufficient to get an accurate model.

4-11
4 Identifying State-Space Models

Estimate State-Space Models in System Identification App

Prerequisites
• Import data into the System Identification app. See “Represent Data”. For supported
data formats, see “Data Supported by State-Space Models” on page 4-5.
• Perform data preprocessing. To improve the accuracy of your model, you should
detrend your data. See “Ways to Prepare Data for System Identification” on page 2-6.

1 Select Estimate > State Space Models.

The State Space Models dialog box opens.

4-12
 Prerequisites

Tip For more information on the options in the dialog box, click Help.
2
Specify a model name by clicking adjacent to Model name. The name of the
model must be unique in the Model Board.
3 Select the Specify value option (if not already selected) and specify the model order
in the edit field. Model order refers to the number of states in the state-space model.

Tip When you do not know the model order, search and select for an order. For more
information, see “Estimate Model With Selected Order in the App” on page 4-7.
4 Select the Continuous-time or Discrete-time option to specify whether to estimate
a continuous- or discrete-time model.

You cannot estimate a discrete-time model if the working data is continuous-time

frequency-domain data.
5 Expand the Model Structure Selection section to configure the model structure,
such as canonical form, whether to estimate the disturbance component (K matrix),
and specification of feedthrough and input delays.

4-13
4 Identifying State-Space Models

For more information about the type of state-space parameterization, see “Supported
State-Space Parameterizations” on page 4-6.
6 Expand the Estimation Options section to select the estimation method, and
configure the cost function.

Select one of the following Estimation Method from the drop-down list and
configure the options. For more information about these methods, see “State-Space
Model Estimation Methods” on page 4-41.

4-14
 Subspace (N4SID)

Subspace (N4SID)

a In the N4Weight drop-down list, specify the weighting scheme used for
singular-value decomposition by the N4SID algorithm.

The N4SID algorithm is used both by the subspace and Prediction Error
Minimization (PEM) methods.
b In the N4Horizon field, specify the forward and backward prediction horizons
used by the N4SID algorithm.

The N4SID algorithm is used both by the subspace and PEM methods.
c In the Focus drop-down list, select how to weigh the relative importance of
the fit at different frequencies. For more information about each option, see
“Assigning Estimation Weightings” on page 4-39.
d Select the Allow unstable models check box to specify whether to allow the
estimation process to use parameter values that may lead to unstable models.

Setting this option is same as setting the estimation option Focus to

'prediction' at the command line. An unstable model is delivered only if it

4-15
4 Identifying State-Space Models

produces a better fit to the data than other stable models computed during the
estimation process.
e Select the Estimate covariance check box if you want the algorithm to
compute parameter uncertainties.

Effects of such uncertainties are displayed on plots as model confidence regions.

Skipping uncertainty computation reduces computation time for complex models
and large data sets.
f Select the Display progress check box to open a progress viewer window
during estimation.
g In the Initial state list, specify how you want the algorithm to treat initial
states. For more information about the available options, see “Specifying Initial
States for Iterative Estimation Algorithms” on page 4-40.

Prediction Error Minimization (PEM)

• In the N4Weight drop-down list, specify the weighting scheme used for singular-
value decomposition by the N4SID algorithm.

The N4SID algorithm is used both by the subspace and Prediction Error
Minimization (PEM) methods.
4-16
 Prediction Error Minimization (PEM)

• In the N4Horizon field, specify the forward and backward prediction horizons
used by the N4SID algorithm.

The N4SID algorithm is used both by the subspace and PEM methods.
• In the Focus drop-down list, select how to weigh the relative importance of the fit
at different frequencies. For more information about each option, see “Assigning
Estimation Weightings” on page 4-39.
• Select the Allow unstable models check box to specify whether to allow the
estimation process to use parameter values that may lead to unstable models.

Setting this option is same as setting the estimation option Focus to

'prediction' at the command line. An unstable model is delivered only if it
produces a better fit to the data than other stable models computed during the
estimation process.
• Select the Estimate covariance check box if you want the algorithm to compute
parameter uncertainties.

Effects of such uncertainties are displayed on plots as model confidence regions.

Skipping uncertainty computation reduces computation time for complex models
and large data sets.
• Select the Display progress check box to open a progress viewer window during
estimation.
• In the Initial state list, specify how you want the algorithm to treat initial
states. For more information about the available options, see “Specifying Initial
States for Iterative Estimation Algorithms” on page 4-40.

Tip If you get an inaccurate fit, try setting a specific method for handling initial
states rather than choosing it automatically.
• Click Regularization to obtain regularized estimates of model parameters.
Specify the regularization constants in the Regularization Options dialog box. To
learn more, see “Regularized Estimates of Model Parameters”.
• Specify Iteration Options to specify options for controlling the iterations. The
Options for Iterative Minimization dialog box opens.

4-17
4 Identifying State-Space Models

Iteration Options
In the Options for Iterative Minimization dialog box, you can specify the following
iteration options:

• Search Method — Method used by the iterative search algorithm. Search

method is auto by default. The descent direction is calculated using gn
(Gauss-Newton), gna (Adaptive Gauss-Newton), lm (Levenberg-Marquardt),
lsqnonlin (Trust-Region Reflective Newton), and grad (Gradient Search)
successively at each iteration until a sufficient reduction in error is achieved.
• Output weighting — Weighting applied to the loss function to be minimized.
Use this option for multi-output estimations only. Specify as 'noise' or a
positive semidefinite matrix of size equal the number of outputs.
• Maximum number of iterations — Maximum number of iterations to use
during search.
• Termination tolerance — Tolerance value when the iterations should
terminate.

4-18
 Regularized Reduction

• Error threshold for outlier penalty — Robustification of the quadratic

criterion of fit.

Regularized Reduction

• In the Regularization Kernel drop-down list, select the regularizing kernel to

use for regularized estimation of the underlying ARX model. To learn more, see
“Regularized Estimates of Model Parameters”.
• In the ARX Orders field, specify the order of the underlying ARX model. By
default, the orders are automatically computed by the estimation algorithm. If
you specify a value, it is recommended that you use a large value for nb order. To
learn more about ARX orders, see arx.
• In the Focus drop-down list, select how to weigh the relative importance of the fit
at different frequencies. For more information about each option, see “Assigning
Estimation Weightings” on page 4-39.
• In the Reduction Method drop-down list, specify the reduction method:

4-19
4 Identifying State-Space Models

• Truncate — Discards the specified states without altering the remaining

states. This method tends to product a better approximation in the frequency
domain, but the DC gains are not guaranteed to match.
• MatchDC — Discards the specified states and alters the remaining states to
preserve the DC gain.
• Select the Estimate covariance check box if you want the algorithm to compute
parameter uncertainties.

Effects of such uncertainties are displayed on plots as model confidence regions.

Skipping uncertainty computation reduces computation time for complex models
and large data sets.
• Select the Display progress check box to open a progress viewer window during
estimation.
• In the Initial state list, specify how you want the algorithm to treat initial
states. For more information about the available options, see “Specifying Initial
States for Iterative Estimation Algorithms” on page 4-40.

Tip If you get an inaccurate fit, try setting a specific method for handling initial
states rather than choosing it automatically.

The estimation process uses parameter values that always lead to a stable model.
7 Click Estimate to estimate the model. A new model gets added to the System
Identification app.

Next Steps
• Validate the model by selecting the appropriate response type in the Model Views
area of the app. For more information about validating models, see “Validating
Models After Estimation” on page 10-2.
• Export the model to the MATLAB workspace for further analysis by dragging it to the
To Workspace rectangle in the app.

4-20
 Estimate State-Space Models at the Command Line

Estimate State-Space Models at the Command Line

In this section...
“Black Box vs. Structured State-Space Model Estimation” on page 4-21
“Estimating State-Space Models Using ssest, ssregest and n4sid” on page 4-22
“Choosing the Structure of A, B, C Matrices” on page 4-23
“Choosing Between Continuous-Time and Discrete-Time Representations” on page
4-23
“Choosing to Estimate D, K, and X0 Matrices” on page 4-24

Black Box vs. Structured State-Space Model Estimation

You can estimate state-space models in two ways at the command line, depending upon
your prior knowledge of the nature of the system and your requirements.

• “Black Box Estimation” on page 4-21

• “Structured Estimation” on page 4-21

Black Box Estimation

In this approach, you specify the model order, and, optionally, additional model structure
attributes that configure the overall structure of the state-space matrices. You call
ssest, ssregest or n4sid with data and model order as primary input arguments, and
use name-value pairs to specify any additional attributes, such as model sample time,
presence of feedthrough, absence of noise component, etc. You do not work directly with
the coefficients of the A, B, C, D, K, and X0 matrices.

Structured Estimation

In this approach, you create and configure an idss model that contains the initial values
for all the system matrices. You use the Structure property of the idss model to
specify all the parameter constraints. For example, you can designate certain coefficients
of system matrices as fixed and impose minimum/maximum bounds on the values of
the others. For quick configuration of the parameterization and whether to estimate
feedthrough and disturbance dynamics, use ssform.

After configuring the idss model with desired constraints, you specify this model as
an input argument to the ssest command. You cannot use n4sid or ssregest for
structured estimation.

4-21
4 Identifying State-Space Models

Note:

• The structured estimation approach is also referred to as grey-box modeling.

However, in this toolbox, the “grey box modeling” terminology is used only when
referring to idgrey and idnlgrey models.

• Using the structured estimation approach, you cannot specify relationships among
state-space coefficients. Each coefficient is essentially considered to be independent of
others. For imposing dependencies, or to use more complex forms of parameterization,
use the idgrey model and greyest estimator.

Estimating State-Space Models Using ssest, ssregest and n4sid

Prerequisites
• Represent input-output data as an iddata object or frequency-response data as an
frd or idfrd object. See “Representing Time- and Frequency-Domain Data Using
iddata Objects” on page 2-50. For supported data formats, see “Data Supported by
State-Space Models” on page 4-5.
• Perform data preprocessing. To improve the accuracy of results when using time-
domain data, you can detrend the data or specify the input/output offset levels as
estimation options. See “Ways to Prepare Data for System Identification” on page 2-6.
• Select a model order. When you do not know the model order, search and select for
an order. For more information, see “Estimate Model With Selected Order at the
Command Line” on page 4-10.

You can estimate continuous-time and discrete-time state-space models using the
iterative estimation command ssest that minimizes the prediction errors to obtain
maximum-likelihood values.

Use the following general syntax to both configure and estimate state-space models:

m = ssest(data,n,opt,Name,Value)

where data is the estimation data, n is the model order. opt contains the options for
configuring the estimation of the state-space models. These options include the handling
of the initial conditions, input and output offsets, estimation focus and search algorithm
options. opt can be followed by name and value pair input arguments that specify

4-22
 Prerequisites

optional model structure attributes such as the presence of feedthrough, the canonical
form of the model, and input delay.

As an alternative to ssest, you can use the noniterative subspace estimators n4sid or
ssregest:

m = n4sid(data,n,opt,Name,Value)
m = ssregest(data,n,opt,Name,Value)

Unless you specify the sample time as a name-value pair input argument, n4sid and
ssregest estimate a discrete-time model, while ssest estimates a continuous-time
model.

Note: ssest uses n4sid to initialize the state-space matrices, and takes longer than
n4sid to estimate a model but typically provides better fit to data.

For information about validating your model, see “Validating Models After Estimation”
on page 10-2

Choosing the Structure of A, B, C Matrices

By default, all entries of the A, B, and C state-space matrices are treated as free
parameters. Using the Form name and value pair input argument of ssest , you can
choose various canonical forms, such as the companion and modal forms, that use fewer
parameters.

For more information about estimating a specific state-space parameterization, see:

• “Estimate State-Space Models with Free-Parameterization” on page 4-27

• “Estimate State-Space Models with Canonical Parameterization” on page 4-28
• “Estimate State-Space Models with Structured Parameterization” on page 4-30

Choosing Between Continuous-Time and Discrete-Time Representations

For estimation of state-space models, you have the option of switching the model sample
time between zero and that of the estimation data. You can do this using the Ts name
and value pair input argument.

4-23
4 Identifying State-Space Models

• By default, ssest estimates a continuous-time model. If you are using data set
with nonzero sample time, data, which includes all time domain data, you can also
estimate a discrete-time model by using:

model = ssest(data,nx,'Ts',data.Ts);

If you are using continuous-time frequency-domain data, you cannot estimate a

discrete-time model.
• By default, n4sid and ssregest estimate a model whose sample time matches that
of the data. Thus, for time-domain data, n4sid and ssregest deliver a discrete-time
model. You can estimate a continuous-time model by using:

model = n4sid(data,nx,'Ts',0);

or

model = ssregest(data,nx,'Ts',0);

Choosing to Estimate D, K, and X0 Matrices

For state-space models with any parameterization, you can specify whether to estimate
the D, K and X0 matrices, which represent the input-to-output feedthrough, noise model
and the initial states, respectively.

For state-space models with structured parameterization, you can also specify to
estimate the D matrix. However, for free and canonical forms, the structure of the
D matrix is set based on your choice of 'Feedthrough' name and value pair input
argument.

D Matrix

By default, the D matrix is not estimated and its value is fixed to zero, except for static
models.

• Black box estimation: Use the Feedthrough name and value pair input argument
to denote the presence or absence of feedthrough from individual inputs. For example,
in case of a two input model such that there is feedthrough from only the second
input, use model = n4sid(data,n,'Feedthrough',[false true]);.
• Structured estimation: Configure the values of the init_sys.Structure.d,
where init_sys is an idss model that represents the desired model structure. To
force no feedthrough for the i-th input, set:

4-24
 Prerequisites

init_sys.Structure.d.Value(:,i) = 0;
init_sys.Structure.d.Free = true;
init_sys.Structure.d.Free(:,i) = false;

The first line specifies the value of the i-th column of D as zero. The next line specifies
all the elements of D as free, estimable parameters. The last line specifies that the i-
th column of the D matrix is fixed for estimation.

Alternatively, use ssform with 'Feedthrough' name-value pair..

K Matrix

K represents the noise matrix of the model, such that the noise component of the model
is:.

x& = Ax + Ke
yn = Cx + e

For frequency-domain data, no noise model is estimated and K is set to 0. For time-
domain data, K is estimated by default in the black box estimation setup. yn is the
contribution of the disturbances to the model output.

• Black box estimation: Use the DisturbanceModel name and value pair input
argument to indicate if the disturbance component is fixed to zero (specify Value
= ‘none’) or estimated as a free parameter (specify Value = ‘estimate’). For
example, use model = n4sid(data,n,'DisturbanceModel','none').
• Structured estimation: Configure the value of the init_sys.Structure.k
parameter, where init_sys is an idss model that represents the desired model
structure. You can fix some K matrix coefficients to known values and prescribe
minimum/maximum bounds for free coefficients. For example, to estimate only the
first column of the K matrix for a two output model:

kpar = init_sys.Structure.k;
kpar.Free(:,1) = true;
kpar.Free(:,2) = false;
kpar.Value(:,2) = 0; % second column value is fixed to zero
init_sys.Structure.k = kpar;

Alternatively, use ssform.

4-25
4 Identifying State-Space Models

When not sure how to easily fix or free all coefficients of K, initially you can omit
estimating the noise parameters in K to focus on achieving a reasonable model for the
system dynamics. After estimating the dynamic model, you can use ssest to refine the
model while configuring the K parameters to be free. For example:
init_sys = ssest(data, n,'DisturbanceModel','none');
init_sys.Structure.k.Free = true;
sys = ssest(data, init_sys);

where init_sys is the dynamic model without noise.

To set K to zero in an existing model, you can set its Value to 0 and Free flag to false:

m.Structure.k.Value = 0;
m.Structure.k.Free = false;

X0 Matrices

The initial state vector X0 is obtained as the by-product of model estimation. The
n4sid, ssest and ssregest commands return the value of X0 as their second output
arguments. You can choose how to handle initial conditions during model estimation
by using the InitialState estimation option. Use n4sidOptions (for n4sid),
ssestOptions (for ssest) or ssregestOptions (for ssregest) to create the
estimation option set. For example, in order to hold the initial states to zero during
estimation using n4sid:

opt = n4sidOptions;
opt.InitialState = 'zero';
[m,X0] = n4sid(data,n,opt);

The returned X0 variable is a zero vector of length n.

When you estimate models using multiexperiment data, the X0 matrix contains as many
columns as data experiments.

For a complete list of values for the InitialStates option, see “Specifying Initial
States for Iterative Estimation Algorithms” on page 4-40.

4-26
 Estimate State-Space Models with Free-Parameterization

Estimate State-Space Models with Free-Parameterization

The default parameterization of the state-space matrices A, B, C, D, and K is free; that
is, any elements in the matrices are adjustable by the estimation routines. Because
the parameterization of A, B, and C is free, a basis for the state-space realization is
automatically selected to give well-conditioned calculations.

To estimate the disturbance model K, you must use time-domain data.

Suppose that you have no knowledge about the internal structure of the discrete-time
state-space model. To quickly get started, use the following syntax:
m = ssest(data)

or
m = ssregest(data)

where data is your estimation data. ssest estimates a continuous-time state-space

model for an automatically selected order between 1 and 10. ssregest estimates a
discrete-time model.

To find a model of a specific order n, use the following syntax:

m = ssest(data,n)

or
m = ssregest(dat,n)

The iterative algorithm ssest is initialized by the subspace method n4sid. You can use
n4sid directly, as an alternative to ssest:

m = n4sid(data)

which automatically estimates a discrete-time model of the best order in the 1:10 range.

4-27
4 Identifying State-Space Models

Estimate State-Space Models with Canonical Parameterization

In this section...
“What Is Canonical Parameterization?” on page 4-28
“Estimating Canonical State-Space Models” on page 4-28

What Is Canonical Parameterization?

Canonical parameterization represents a state-space system in a reduced parameter
form where many elements of A, B and C matrices are fixed to zeros and ones. The free
parameters appear in only a few of the rows and columns in state-space matrices A, B, C,
D, and K. The free parameters are identifiable — they can be estimated to unique values.
The remaining matrix elements are fixed to zeros and ones.

The software supports the following canonical forms:

• Companion form: The characteristic polynomial appears in the rightmost column of

the A matrix.
• Modal decomposition form: The state matrix A is block diagonal, with each block
corresponding to a cluster of nearby modes.

Note: The modal form has a certain symmetry in its block diagonal elements. If you
update the parameters of a model of this form (as a structured estimation using
ssest), the symmetry is not preserved, even though the updated model is still block-
diagonal.
• Observability canonical form: The free parameters appear only in select rows of
the A matrix and in the B and K matrices.

For more information about the distribution of free parameters in the observability
canonical form, see the Appendix 4A, pp 132-134, on identifiability of black-box
multivariable model structures in System Identification: Theory for the User, Second
Edition, by Lennart Ljung, Prentice Hall PTR, 1999 (equation 4A.16).

Estimating Canonical State-Space Models

You can estimate state-space models with chosen parameterization at the command line.

4-28
 Estimate State-Space Models with Canonical Parameterization

For example, to specify an observability canonical form, use the 'Form' name and value
pair input argument, as follows:
m = ssest(data,n,'Form','canonical')

Similarly, set 'Form' as 'modal' or 'companion' to specify modal decomposition and

companion canonical forms, respectively.

If you have time-domain data, the preceding command estimates a continuous-time

model. If you want a discrete-time model, specify the data sample time using the 'Ts'
name and value pair input argument:
md = ssest(data, n,'Form','canonical','Ts',data.Ts)

If you have continuous-time frequency-domain data, you can only estimate a continuous-
time model.

4-29
4 Identifying State-Space Models

Estimate State-Space Models with Structured Parameterization

In this section...
“What Is Structured Parameterization?” on page 4-30
“Specify the State-Space Model Structure” on page 4-30
“Are Grey-Box Models Similar to State-Space Models with Structured
Parameterization?” on page 4-32
“Estimate Structured Discrete-Time State-Space Models” on page 4-33
“Estimate Structured Continuous-Time State-Space Models” on page 4-34

What Is Structured Parameterization?

Structured parameterization lets you exclude specific parameters from estimation by
setting these parameters to specific values. This approach is useful when you can derive
state-space matrices from physical principles and provide initial parameter values based
on physical insight. You can use this approach to discover what happens if you fix specific
parameter values or if you free certain parameters.

There are two stages to the structured estimation procedure:

1 Specify the state-space model structure, as described in “Specify the State-Space

Model Structure” on page 4-30
2 Estimate the free model parameters, as described in “Estimate State-Space Models
at the Command Line” on page 4-21

This approach differs from estimating models with free and canonical parameterizations,
where it is not necessary to specify initial parameter values before the estimation.
For free parameterization, there is no structure to specify because it is assumed to be
unknown. For canonical parameterization, the structure is fixed to a specific form.

Note: To estimate structured state-space models in the System Identification app, define
the corresponding model structures at the command line and import them into the
System Identification app.

Specify the State-Space Model Structure

To specify the state-space model structure:

4-30
 More about Structure

1 Use idss to create a state-space model. For example:

A = [0 1; 0 -1];
B = [0; 0.28];
C = eye(2);
D = zeros(2,1);
m = idss(A,B,C,D,K,'Ts',T)

creates a discrete-time state-space structure, where A, B, C, D, and K specify the

initial values for the free parameters. T is the sampling interval.
2 Use the Structure property of the model to specify which parameters to estimate
and which to set to specific values.

More about Structure

Structure contains parameters for the five state-space matrices, A, B, C, D, and K.

For each parameter, you can set the following attributes:

• Value — Parameter values. For example, sys.Structure.a.Value contains

the initial or estimated values of the A matrix.

NaN represents unknown parameter values.

Each property sys.a, sys.b, sys.c, and sys.d is an alias to the corresponding
Value entry in the Structure property of sys. For example, sys.a is an alias to
the value of the property sys.Structure.a.Value
• Minimum — Minimum value that the parameter can assume during estimation.
For example, sys.Structure.k.Minimum = 0 constrains all entries in the K
matrix to be greater than or equal to zero.
• Maximum — Maximum value that the parameter can assume during estimation.
• Free — Boolean specifying whether the parameter is a free estimation
variable. If you want to fix the value of a parameter during estimation, set
the corresponding Free = false. For example, if A is a 3-by-3 matrix,
sys.Structure.a.Free = eyes(3) fixes all of the off-diagonal entries in A, to
the values specified in sys.Structure.a.Value. In this case, only the diagonal
entries in A are estimable.
• Scale — Scale of the parameter’s value. Scale is not used in estimation.

4-31
4 Identifying State-Space Models

• Info — Structure array for storing parameter units and labels. The structure
has Label and Unit fields.

Use these fields for your convenience, to store strings that describe parameter
units and labels.

For example, if you want to fix A(1,2)=A(2,1)=0, use:

m.Structure.a.Value(1,2) = 0;
m.Structure.a.Value(2,1) = 0;
m.Structure.a.Free(1,2) = false;
m.Structure.a.Free(2,1) = false;

The estimation algorithm only estimates the parameters in A for which

m.Structure.a.Free is true.

Use physical insight, whenever possible, to initialize the parameters for the iterative
search algorithm. Because it is possible that the numerical minimization gets stuck
in a local minimum, try several different initialization values for the parameters.
For random initialization, use init. When the model structure contains parameters
with different orders of magnitude, try to scale the variables so that the parameters
are all roughly the same magnitude.

Alternatively, to quickly configure the parameterization and whether to estimate

feedthrough and disturbance dynamics, use ssform.
3 Use ssest to estimate the model, as described in “Estimate State-Space Models at
the Command Line” on page 4-21.

The iterative search computes gradients of the prediction errors with respect to the
parameters using numerical differentiation. The step size is specified by the nuderst
command. The default step size is equal to 10–4 times the absolute value of a parameter
or equal to 10–7, whichever is larger. To specify a different step size, edit the nuderst
MATLAB file.

Are Grey-Box Models Similar to State-Space Models with Structured

Parameterization?
You estimate state-space models with structured parameterization when you know
some parameters of a linear system and need to estimate the others. These models
are therefore similar to grey-box models. However, in this toolbox, the "grey box

4-32
 More about Structure

modeling" terminology is used only when referring to idgrey and idnlgrey models.
In these models, you can specify complete linear or nonlinear models with complicated
relationships between the unknown parameters.

If you have independent unknown matrix elements in a linear state-space model

structure, then it is easier and quicker to use state-space models with structured
parameterizations. For imposing dependencies, or to use more complex forms of
parameterization, use the idgrey model and the associated greyest estimator. For
more information, see “Grey-Box Model Estimation”.

Estimate Structured Discrete-Time State-Space Models

This example shows how to estimate the unknown parameters of a discrete-time model.

In this example, you estimate in the following discrete-time model:

Suppose that the nominal values of the unknown parameters ( ) are -1, 2,
3, 4,and 5, respectively.

The discrete-time state-space model structure is defined by the following equation:

Construct the parameter matrices and initialize the parameter values using the nominal
parameter values.
A = [1,-1;0,1];
B = [2;3];
C = [1,0];
D = 0;
K = [4;5];

Construct the state-space model object.

4-33
4 Identifying State-Space Models

m = idss(A,B,C,D,K);

Specify the parameter values in the structure matrices that you do not want to estimate.
S = m.Structure;
S.a.Free(1,1) = false;
S.a.Free(2,:) = false;
S.c.Free = false;
m.Structure = S;

D is initialized, by default, as a fixed value, and K and B are initialized as free values.
Suppose you want to fix the initial states to known zero values. To enforce this, configure
the InitialState estimation option.
opt = ssestOptions;
opt.InitialState = 'zero';

Load estimation data.

load iddata1 z1;

Estimate the model structure.

m = ssest(z1,m,opt);

where z1 is name of the iddata object. The data can be time-domain or frequency-
domain data. The iterative search starts with the nominal values in the A, B, C, D, and K
matrices.

Estimate Structured Continuous-Time State-Space Models

This example shows how to estimate the unknown parameters of a continuous-time
model.

In this example, you estimate in the following continuous-time model:

4-34
 More about Structure

This equation represents an electrical motor, where is the angular position

of the motor shaft, and is the angular velocity. The parameter is the
inverse time constant of the motor, and is the static gain from the input to the
angular velocity.

The motor is at rest at t=0 , but its angular position is unknown. Suppose that the
approximate nominal values of the unknown parameters are and .

The variance of the errors in the position measurement is 0.01 , and the variance in the
angular velocity measurements is 0.1 . For more information about this example, see
the section on state-space models in System Identification: Theory for the User, Second
Edition, by Lennart Ljung, Prentice Hall PTR, 1999.

The continuous-time state-space model structure is defined by the following equation:

Construct the parameter matrices and initialize the parameter values using the nominal
parameter values.

A = [0 1;0 -1];
B = [0;0.25];
C = eye(2);
D = [0;0];
K = zeros(2,2);
x0 = [0;0];

The matrices correspond to continuous-time representation. However, to be consistent

with the idss object property name, this example uses A, B, and C instead of F, G, and H.

Construct the continuous-time state-space model object.

m = idss(A,B,C,D,K,'Ts',0);

Specify the parameter values in the structure matrices that you do not want to estimate.

S = m.Structure;
S.a.Free(1,:) = false;
S.a.Free(2,1) = false;

4-35
4 Identifying State-Space Models

S.b.Free(1) = false;
S.c.Free = false;
S.d.Free = false;
S.k.Free = false;
m.Structure = S;
m.NoiseVariance = [0.01 0; 0 0.1];

The initial state is partially unknown. Use the InitialState option of the
ssestOptions option set to configure the estimation behavior of X0.

opt = ssestOptions;
opt.InitialState = idpar(x0);
opt.InitialState.Free(2) = false;

Estimate the model structure.

load(fullfile(matlabroot, 'toolbox', 'ident', 'iddemos', 'data', 'dcmotordata'));
z = iddata(y, u, 0.1);
m = ssest(z, m, opt);

The iterative search for a minimum is initialized by the parameters in the nominal model
m . The continuous-time model is sampled using the same sampling interval as the data
during estimation.

Simulate this system using the sampling interval T=0.1 for input u and the noise
realization e.

e = randn(300,2);
u1 = idinput(300);
simdat = iddata([],u1,'Ts',0.1);
simopt = simOptions('AddNoise', true, 'NoiseData', e);
y1 = sim(m,simdat,simopt);

The continuous system is sampled using Ts=0.1 for simulation purposes. The noise
sequence is scaled according to the matrix m.NoiseVariance .

If you discover that the motor was not initially at rest, you can estimate by setting
the second element of the InitialState parameter to be free.

opt.InitialState.Free(2) = true;
m_new = ssest(z, m, opt);

4-36
 Estimate the State-Space Equivalent of ARMAX and OE Models

Estimate the State-Space Equivalent of ARMAX and OE Models

This example shows how to estimate ARMAX and OE-form models using the state-space
estimation approach.

You can estimate the equivalent of multiple-output ARMAX and Output-Error (OE)
models using state-space model structures:

• For an “ARMAX” model, specify to estimate the K matrix for the state-space model
• For an “OE” model, set K = 0

Convert the resulting model into idpoly models to see them in the commonly defined
ARMAX or Output-Error forms.

Load measured data.

load iddata1 z1

Estimate state-space models.

mss_noK = n4sid(z1, 2,'DisturbanceModel','none');
mss = n4sid(z1,2);

mss_noK is a second order state-space model with no disturbance model used during
estimation. mss is also a second order state-space model, but with an estimated noise
component. Both models use the measured data set z1 for estimation.

Convert the state-space models to polynomial models.

mOE = idpoly(mss_noK);
mARMAX = idpoly(mss);

Converting to polynomial models results in the parameter covariance information for mOE
and mARMAX to be lost.

You can use one of the following to recompute the covariance:

• Zero-iteration update using the same estimation data.

• translatecov as a Gauss approximation formula based translation of covariance of
mss_noK and mss into covariance of mOE and mARMAX.

Reestimate mOE and mARMAX for the parameters of the polynomial model using a zero
iteration update.

4-37
4 Identifying State-Space Models

opt = polyestOptions;
opt.SearchOption.MaxIter = 0;

mOE = polyest(z1,mOE);
mARMAX = polyest(z1,mARMAX);

The options object, opt, specifies a zero iteration update for mOE and mARMAX.
Consequently, the model parameters remain unchanged and only their covariance
information is updated.

Alternatively, you can use translatecov to convert the estimated models into
polynomial form.

fcn = @(x)idpoly(x);
mOE = translatecov(fcn, mss_noK);
mARMAX = translatecov(fcn, mss);

Because polyest and translatecov use different computation algorithms, the

covariance data obtained by running a zero-iteration update may not match that
obtained using translatecov.

View the uncertainties of the model parameters.

present(mOE)
present(mARMAX)

You can use a state-space model with K = 0 (Output-Error (OE) form) for initializing a
Hammerstein-Wiener estimation at the command line. This initialization may improve
the fit of the model. See “Using Linear Model for Hammerstein-Wiener Estimation” on
page 5-64.

For more information about ARMAX and OE models, see “Identifying Input-Output
Polynomial Models” on page 3-40.

4-38
 Assigning Estimation Weightings

Assigning Estimation Weightings

You can specify how the estimation algorithm weighs the fit at various frequencies. This
information supports the estimation procedures “Estimate State-Space Models in System
Identification App” on page 4-12 and “Estimate State-Space Models at the Command
Line” on page 4-21.

In the System Identification app. Set Focus to one of the following options:

• Prediction — Uses the inverse of the noise model H to weigh the relative
importance of how closely to fit the data in various frequency ranges. Corresponds to
minimizing one-step-ahead prediction, which typically favors the fit over a short time
interval. Optimized for output prediction applications.
• Simulation — Uses the input spectrum to weigh the relative importance of the
fit in a specific frequency range. Does not use the noise model to weigh the relative
importance of how closely to fit the data in various frequency ranges. Optimized for
output simulation applications.
• Stability — Estimates the best stable model. For more information about model
stability, see “Unstable Models” on page 10-88.
• Filter — Specify a custom filter to open the Estimation Focus dialog box, where
you can enter a filter, as described in “Simple Passband Filter” on page 2-128 or
“Defining a Custom Filter” on page 2-129. This prefiltering applies only for estimating
the dynamics from input to output. The disturbance model is determined from the
estimation data.

At the command line. Specify the focus as an estimation option using the same options
as in the app. For example, use this command to emphasize the fit between the 5 and 8
rad/s:
opt = ssestOptions;
opt.Focus = [5 8];
model = ssest(data,4,opt);

For more information on the 'Focus' option, see the reference page for ssestOptions
and ssregestOptions.

4-39
4 Identifying State-Space Models

Specifying Initial States for Iterative Estimation Algorithms

When you estimate state-space models, you can specify how the algorithm treats initial
states. This information supports the estimation procedures “Estimate State-Space
Models in System Identification App” on page 4-12 and “Estimate State-Space Models at
the Command Line” on page 4-21.

In the System Identification app. Set Initial state to one of the following options:

• Auto — Automatically chooses Zero, Estimate, or Backcast based on the

estimation data. If initial states have negligible effect on the prediction errors, the
initial states are set to zero to optimize algorithm performance.
• Zero — Sets all initial states to zero.
• Estimate — Treats the initial states as an unknown vector of parameters and
estimates these states from the data.
• Backcast — Estimates initial states using a backward filtering method (least-
squares fit).

At the command line. Specify the method for handling initial states using the
InitialState estimation option. For example, to estimate a fourth-order state-space
model and set the initial states to be estimated from the data:
opt = ssestOptions('InitialState','estimate');
m = ssest(data, 4, opt)

For a complete list of values for the InitialState model property, see the
ssestOptions, n4sidOptions and ssregestOptions reference pages.

Note: For the n4sid algorithm, 'auto' and 'backcast' are equivalent to
'estimate'.

4-40
 State-Space Model Estimation Methods

State-Space Model Estimation Methods

You can estimate state-space models using one of the following estimation methods:

• N4SID — This is a noniterative, subspace method. The method works on both time-
domain and frequency-domain data and is typically faster than the SSEST algorithm.
You can choose the subspace algorithms such as CVA, SSARX, or MOESP using the
n4Weight option. You can also use this method to get an initial model (see n4sid),
and then refine the initial estimate using the iterative prediction-error method
ssest.

For more information about this algorithm, see [1].

• SSEST — This is an iterative method and uses prediction error minimization
algorithm. The method works on both time-domain and frequency-domain data. For
black-box estimation, the method initializes the model parameters using n4sid and
then it updates the parameters using an iterative search o minimize the prediction
errors. You can also use this method for structured estimation using an initial model
with initial values of one or more parameters fixed in value.

For more information on this algorithm, see [2].

• SSREGEST — This is a noniterative method. The method works on discrete time-
domain data and frequency-domain data. It first estimates a high-order regularized
ARX or FIR model, converts it to a state-space model and then performs balanced
reduction on it. This method provides improved accuracy on short, noisy data sets.

With all the estimation methods, you have the option of specifying how to handle initial
state, delays, feedthrough behavior and disturbance component of the model.

References

[1] van Overschee, P., and B. De Moor. Subspace Identification of Linear Systems: Theory,
Implementation, Applications. Springer Publishing: 1996.

[2] Ljung, L. System Identification: Theory For the User, Second Edition, Upper Saddle
River, N.J: Prentice Hall, 1999.

[3] T. Chen, H. Ohlsson, and L. Ljung. “On the Estimation of Transfer Functions,
Regularizations and Gaussian Processes - Revisited”, Automatica, Volume 48,
August 2012.

4-41
4-42
 5

Nonlinear Black-Box Model

Identification

• “About Nonlinear Model Identification” on page 5-2

• “Preparing Data for Nonlinear Identification” on page 5-6
• “Identifying Nonlinear ARX Models” on page 5-7
• “Identifying Hammerstein-Wiener Models” on page 5-48
• “Linear Approximation of Nonlinear Black-Box Models” on page 5-81
5 Nonlinear Black-Box Model Identification

About Nonlinear Model Identification

In this section...
“What Are Nonlinear Models?” on page 5-2
“When to Fit Nonlinear Models” on page 5-2
“Available Nonlinear Models” on page 5-4

What Are Nonlinear Models?

Dynamic models in System Identification Toolbox software are mathematical
relationships between the system's inputs u(t) and outputs y(t). You can use these
relationships to compute the current output from previous inputs and outputs. The
general form of a model in discrete time is:
y(t) = f(u(t - 1), y(t - 1), u(t - 2), y(t - 2), . . .)

Such a model is nonlinear if the function f is a nonlinear function. f may represent

arbitrary nonlinearities, such as switches and saturations.

When to Fit Nonlinear Models

In practice, all systems are nonlinear and the output is a nonlinear function of the input
variables. However, a linear model is often sufficient to accurately describe the system
dynamics. In most cases, you should first try to fit linear models.

Here are some scenarios when you might need the additional flexibility of nonlinear
models:

• “Linear Model Is Not Good Enough” on page 5-2

• “Physical System Is Weakly Nonlinear” on page 5-3
• “Physical System Is Inherently Nonlinear” on page 5-3
• “Linear and Nonlinear Dynamics Are Captured Separately” on page 5-4

Linear Model Is Not Good Enough

You might need nonlinear models when a linear model provides a poor fit to the
measured output signals and cannot be improved by changing the model structure or

5-2
 About Nonlinear Model Identification

order. Nonlinear models have more flexibility in capturing complex phenomena than the
linear models of similar orders.

Physical System Is Weakly Nonlinear

From physical insight or data analysis, you might know that a system is weakly
nonlinear. In such cases, you can estimate a linear model and then use this model as an
initial model for nonlinear estimation. Nonlinear estimation can improve the fit by using
nonlinear components of the model structure to capture the dynamics not explained
by the linear model. For more information, see “Using Linear Model for Nonlinear
ARX Estimation” on page 5-28 and “Using Linear Model for Hammerstein-Wiener
Estimation” on page 5-64.

Physical System Is Inherently Nonlinear

You might have physical insight that your system is nonlinear. Certain phenomena are
inherently nonlinear in nature, including dry friction in mechanical systems, actuator
power saturation, and sensor nonlinearities in electro-mechanical systems. You can
try modeling such systems using the Hammerstein-Wiener model structure, which lets
you interconnect linear models with static nonlinearities. For more information, see
“Identifying Hammerstein-Wiener Models” on page 5-48.

Nonlinear models might be necessary to represent systems that operate over a range of
operating points. In some cases, you might fit several linear models, where each model
is accurate at specific operating conditions. You can also try using the nonlinear ARX
model structure with tree partitions to model such systems. For more information, see
“Identifying Nonlinear ARX Models” on page 5-7.

If you know the nonlinear equations describing a system, you can represent this system
as a nonlinear grey-box model and estimate the coefficients from experimental data. In
this case, the coefficients are the parameters of the model. For more information, see
“Grey-Box Model Estimation”.

Before fitting a nonlinear model, try transforming your input and output variables
such that the relationship between the transformed variables becomes linear. For
example, you might be dealing with a system that has current and voltage as inputs
to an immersion heater, and the temperature of the heated liquid as an output. In this
case, the output depends on the inputs via the power of the heater, which is equal to the
product of current and voltage. Instead of fitting a nonlinear model to two-input and one-
output data, you can create a new input variable by taking the product of current and
voltage and then fitting a linear model to the single-input/single-output data.

5-3
5 Nonlinear Black-Box Model Identification

Linear and Nonlinear Dynamics Are Captured Separately

You might have multiple data sets that capture the linear and nonlinear dynamics
separately. For example, one data set with low amplitude input (excites the linear
dynamics only) and another data set with high amplitude input (excites the nonlinear
dynamics). In such cases, first estimate a linear model using the first data set. Next,
use the model as an initial model to estimate a nonlinear model using the second data
set. For more information, see “Using Linear Model for Nonlinear ARX Estimation” on
page 5-28 and “Using Linear Model for Hammerstein-Wiener Estimation” on page
5-64.

Available Nonlinear Models

System Identification Toolbox supports these nonlinear models:

• “Nonlinear ARX Models” on page 5-4

• “Hammerstein-Wiener Models” on page 5-4
• “Nonlinear State-Space Models” on page 5-5

Nonlinear ARX Models

Nonlinear ARX models extend the linear ARX models to the nonlinear case and have this
structure:
y(t) = f(y(t - 1), ..., y(t - na), u(t - nk), ..., u(t -nk -nb + 1))

where the function f depends on a finite number of previous inputs u and outputs y. na
is the number of past output terms and nb is the number of past input terms used to
predict the current output. nk is the delay from the input to the output, specified as the
number of samples.

Typically, you use nonlinear ARX models as black-box structures. The nonlinear function
of the nonlinear ARX model is a flexible nonlinearity estimator with parameters that
need not have physical significance.

System Identification Toolbox software uses idnlarx objects to represent nonlinear ARX
models. For more information about estimation, see “Nonlinear ARX Models”.

Hammerstein-Wiener Models

Hammerstein-Wiener models describe dynamic systems using one or two static nonlinear
blocks in series with a linear block. The linear block is a discrete transfer function and
represents the dynamic component of the model.

5-4
 About Nonlinear Model Identification

You can use the Hammerstein-Wiener structure to capture physical nonlinear effects in
sensors and actuators that affect the input and output of a linear system, such as dead
zones and saturation. Alternatively, use Hammerstein-Wiener structures as black box
structures that do not represent physical insight into system processes.

System Identification Toolbox software uses idnlhw objects to represent Hammerstein-

Wiener models. For more information about estimation, see “Hammerstein-Wiener
Models”.

Nonlinear State-Space Models

Nonlinear state-space models have this representation:

x& ( t ) = F ( x ( t ) , u ( t ) )
y ( t ) = H ( x ( t) ,u ( t ) )

where F and H can have any parameterization. A nonlinear ordinary differential

equation of high order can be represented as a set of first order equations. You use the
idnlgrey object to specify the structures of such models based on physical insight about
your system. The parameters of such models typically have physical interpretations.

For more information about estimating nonlinear state-space models, see “Grey-Box
Model Estimation”.

5-5
5 Nonlinear Black-Box Model Identification

Preparing Data for Nonlinear Identification

Estimating nonlinear ARX and Hammerstein-Wiener models requires uniformly sampled
time-domain data. Your data can have one or more input and output channels.

For time-series data, you can only fit nonlinear ARX models and “nonlinear state-space
models”.

Tip Whenever possible, use different data sets for model estimation and validation.

Before estimating models, import your data into the MATLAB workspace and do one of
the following:

• In the System Identification app. Import data into the app, as described in
“Represent Data”.
• At the command line. Represent your data as an iddata object, as described in the
corresponding reference page.

You can analyze data quality and preprocess data by interpolating missing values,
filtering to emphasize a specific frequency range, or resampling using a different sample
time (see “Ways to Prepare Data for System Identification” on page 2-6).

Data detrending can be useful in certain cases, such as before modeling the relationship
between the change in input and the change in output about an operating point.
However, most applications do not require you to remove offsets and linear trends from
the data before nonlinear modeling.

5-6
 Identifying Nonlinear ARX Models

Identifying Nonlinear ARX Models

In this section...
“Nonlinear ARX Model Extends the Linear ARX Structure” on page 5-7
“Structure of Nonlinear ARX Models” on page 5-8
“Nonlinearity Estimators for Nonlinear ARX Models” on page 5-9
“Ways to Configure Nonlinear ARX Estimation” on page 5-11
“How to Estimate Nonlinear ARX Models in the System Identification App” on page
5-14
“How to Estimate Nonlinear ARX Models at the Command Line” on page 5-17
“Using Linear Model for Nonlinear ARX Estimation” on page 5-28
“Validating Nonlinear ARX Models” on page 5-33
“Using Nonlinear ARX Models” on page 5-39
“How the Software Computes Nonlinear ARX Model Output” on page 5-40

Nonlinear ARX Model Extends the Linear ARX Structure

A nonlinear ARX model can be understood as an extension of a linear model. A linear
SISO ARX model has this structure:

y( t) + a1 y(t - 1) + a2 y(t - 2) + ... + ana y(t - na) =

b1u(t) + b2u(t - 1) + ... + bnbu( t - nb + 1) + e ( t)

where the input delay nk is zero to simplify the notation.

This structure implies that the current output y(t) is predicted as a weighted sum of past
output values and current and past input values. Rewriting the equation as a product:

yp (t) = [ - a1, - a2 ,..., -ana , b1 , b2 ,.., bnb ] *

[ y(t - 1), y(t - 2),..., y(t - na), u( t),u( t - 1),..., u( t - nb - 1) ]T

where y( t - 1), y( t - 2),..., y(t - na), u(t), u( t - 1),..., u(t - nb - 1) are delayed input and output
variables, called regressors. The linear ARX model thus predicts the current output yp as
a weighted sum of its regressors.

5-7
5 Nonlinear Black-Box Model Identification

This structure can be extended to create a nonlinear form as:

• Instead of the weighted sum that represents a linear mapping, the nonlinear ARX
model has a more flexible nonlinear mapping function:

yp (t) = f ( y(t - 1), y( t - 2), y( t - 3),..., u(t), u(t - 1), u( t - 2),..)

where f is a nonlinear function. Inputs to f are model regressors. When you specify
the nonlinear ARX model structure, you can choose one of several available nonlinear
mapping functions in this toolbox (see “Nonlinearity Estimators for Nonlinear ARX
Models” on page 5-9).
• Nonlinear ARX regressors can be both delayed input-output variables and more
complex, nonlinear expressions of delayed input and output variables. Examples of
such nonlinear regressors are y(t-1)2, u(t-1)*y(t-2), tan(u(t-1)), and u(t-1)*y(t-3).

Structure of Nonlinear ARX Models

This block diagram represents the structure of a nonlinear ARX model in a simulation
scenario:

Nonlinearity Estimator

u Nonlinear
Regressors Function y
u(t),u(t-1),y(t-1), ...
Linear
Function

The nonlinear ARX model computes the output y in two stages:

1 Computes regressors from the current and past input values and past output data.

In the simplest case, regressors are delayed inputs and outputs, such as u(t-1) and
y(t-3)—called standard regressors. You can also specify custom regressors, which
are nonlinear functions of delayed inputs and outputs. For example, tan(u(t-1)) or
u(t-1)*y(t-3).

5-8
 Identifying Nonlinear ARX Models

By default, all regressors are inputs to both the linear and the nonlinear function
blocks of the nonlinearity estimator. You can choose a subset of regressors as inputs
to the nonlinear function block.
2 The nonlinearity estimator block maps the regressors to the model output using
a combination of nonlinear and linear functions. You can select from available
nonlinearity estimators, such as tree-partition networks, wavelet networks, and
multi-layer neural networks. You can also exclude either the linear or the nonlinear
function block from the nonlinearity estimator.

The nonlinearity estimator block can include linear and nonlinear blocks in parallel. For
example:

F( x) = LT ( x - r) + d + g ( Q( x - r) )

x is a vector of the regressors. LT ( x) + d is the output of the linear function block and is
affine when d≠0. d is a scalar offset. g ( Q( x - r) ) represents the output of the nonlinear
function block. r is the mean of the regressors x. Q is a projection matrix that makes
the calculations well conditioned. The exact form of F(x) depends on your choice of the
nonlinearity estimator.

Estimating a nonlinear ARX model computes the model parameter values, such as L, r,
d, Q, and other parameters specifying g. Resulting models are idnlarx objects that store
all model data, including model regressors and parameters of the nonlinearity estimator.
See the idnlarx reference page for more information.

Nonlinearity Estimators for Nonlinear ARX Models

System Identification Toolbox software provides several nonlinearity estimators F(x) for
nonlinear ARX models. For more information about F(x), see “Structure of Nonlinear
ARX Models” on page 5-8.

Each nonlinearity estimator corresponds to an object class in this toolbox. When you
estimate nonlinear ARX models in the app, System Identification Toolbox creates and
configures objects based on these classes. You can also create and configure nonlinearity
estimators at the command line.

Most nonlinearity estimators represent the nonlinear function as a summed series of

nonlinear units, such as wavelet networks or sigmoid functions. You can configure the

5-9
5 Nonlinear Black-Box Model Identification

number of nonlinear units n for estimation. For a detailed description of each estimator,
see the references page of the corresponding nonlinearity class.

Nonlinearity Class Structure Comments

Wavelet wavenet n By default, the
network g( x) = Â a k k (b k ( x - g k ) ) estimation algorithm
(default) k= 1 determines the
number of units n
where k( s) is the wavelet function. automatically.

One layer sigmoidnet n Default number of

sigmoid g( x) = Â ak k (b k ( x - gk )) units n is 10.
network k= 1

-1
where k ( s) = e s + 1
( ) is the sigmoid

function. bk is a row vector such that

bk ( x - g k ) is a scalar.
Tree partition treepartition Piecewise linear function over partitions The estimation
of the regressor space defined by a binary algorithm determines
tree. the number of units
automatically.
Try using tree
partitions for
modeling data
collected over a
range of operating
conditions.
F is linear in x linear This estimator produces a model that Use to specify
is similar to the linear ARX model, custom regressors
but offers the additional flexibility of as the nonlinearity
specifying custom regressors. estimator
and exclude a
nonlinearity mapping
function.

5-10
 Identifying Nonlinear ARX Models

Nonlinearity Class Structure Comments

Multilayered neuralnet Uses as a network object created using  
neural the Neural Network Toolbox™ software.
network
Custom customnet Similar to sigmoid network but you (For advanced use)
network specify k( s) . Uses the unit
(user-defined) function that you
specify.

Ways to Configure Nonlinear ARX Estimation

• “Configurable Elements of Nonlinear ARX Structure” on page 5-11
• “Default Nonlinear ARX Structure” on page 5-12
• “Nonlinear ARX Order and Delay” on page 5-13
• “Estimation Algorithm for Nonlinear ARX Models” on page 5-13

Configurable Elements of Nonlinear ARX Structure

You can adjust various elements of the nonlinear ARX model structure and fit different
models to your data.

Configure model regressors by:

• Specifying model order and delay, which creates the set of standard regressors.

For a definition, see “Nonlinear ARX Order and Delay” on page 5-13.
• Creating custom regressors.

Custom regressors are arbitrary functions of past inputs and outputs, such as
products, powers, and other MATLAB expressions of input and output variables.
You can specify custom regressors in addition to or instead of standard regressors for
greater flexibility in modeling your data.
• Including a subset of regressors in the nonlinear function of the nonlinear estimator
block.

Selecting which regressors are inputs to the nonlinear function reduces model
complexity and keeps the estimation well-conditioned.
• Initializing using a linear ARX model.

5-11
5 Nonlinear Black-Box Model Identification

You can perform this operation only at the command line. The initialization
configures the nonlinear ARX model to use standard regressors, which the toolbox
computes using the orders and delays of the linear model. See “Using Linear Model
for Nonlinear ARX Estimation” on page 5-28.

Configure the nonlinearity estimator block by:

• Specifying and configuring the nonlinear function, including the number of units.
• Excluding the nonlinear function from the nonlinear estimator such that
F(x)= LT ( x) + d .
• Excluding the linear function from the nonlinear estimator such that
F(x)= g ( Q( x - r) ) .

Note: You cannot exclude the linear function from tree partitions and neural
networks.

See these topics for detailed steps to change the model structure:

• “How to Estimate Nonlinear ARX Models in the System Identification App” on page
5-14
• “How to Estimate Nonlinear ARX Models at the Command Line” on page 5-17

Default Nonlinear ARX Structure

Estimate a nonlinear ARX model with default configuration by one of the following:

• Specifying only model order and input delay. Specifying the order automatically
creates standard regressors.
• Specifying a linear ARX model. The linear model sets the model orders and linear
function of the nonlinear model. You can perform this operation only at the command
line.

By default:

• The nonlinearity estimator is a wavelet network (see the wavenet reference page).

This nonlinearity often provides satisfactory results and uses a fast estimation
method.

5-12
 Identifying Nonlinear ARX Models

• All of the standard regressors are inputs to the linear and nonlinear functions of the
wavelet network.

Nonlinear ARX Order and Delay

The order and delay of nonlinear ARX models are positive integers:

• na — Number of past output terms used to predict the current output.

• nb — Number of past input terms used to predict the current output.
• nk — Delay from input to the output in terms of the number of samples.

The meaning of na, nb, and nk is similar to linear ARX model parameters. Orders are
scalars for SISO data, and matrices for MIMO data. If you are not sure how to specify the
order and delay, you can estimate them as described in “Preliminary Step – Estimating
Model Orders and Input Delays” on page 3-46. Such an estimate is based on linear ARX
models and only provides initial guidance—the best orders for a linear ARX model might
not be the best orders for a nonlinear ARX model.

System Identification Toolbox software computes standard regressors using model

orders.

For example, if you specify this order and delay for a SISO model with input u and
output y:
na=2, nb=3, and nk=5

the toolbox computes standard regressors y(t-2), y(t-1), u(t-5), u(t-6), and u(t-7).

You can specify custom regressors in addition to standard regressors, as described in

“How to Estimate Nonlinear ARX Models in the System Identification App” on page
5-14 and “How to Estimate Nonlinear ARX Models at the Command Line” on page
5-17.

Estimation Algorithm for Nonlinear ARX Models

The estimation algorithm depends on your choice of nonlinearity estimator and other
properties of the idnlarx class. You can set algorithm properties both in the app and at
the command line.
Focus property of idnlarx class

By default, estimating nonlinear ARX models minimizes one-step prediction errors,

which corresponds to Focus value of Prediction.

5-13
5 Nonlinear Black-Box Model Identification

If you want a model that is optimized for reproducing simulation behavior, try setting
the Focus value to Simulation. In this case, you cannot use treepartition and
neuralnet because these nonlinearity estimators are not differentiable. Minimization
of simulation error requires differentiable nonlinear functions. Simulation error
minimization takes more time than one-step-ahead prediction error minimization.
Common algorithm properties of idnlarx class

• MaxIter — Maximum number of iterations.

• SearchMethod — Search method for minimization of prediction or simulation errors,
such as Gauss-Newton and Levenberg-Marquardt line search, and Trust-region
reflective Newton approach. By default, the algorithm uses a combination of these
methods.
• Tolerance — Condition for terminating iterative search when the expected
improvement of the parameter values is less than a specified value.
• Display — Shows progress of iterative minimization in the MATLAB Command
Window.

How to Estimate Nonlinear ARX Models in the System Identification App

Prerequisites
• Learn about the nonlinear ARX model structure (see “Structure of Nonlinear ARX
Models” on page 5-8).
• Import data into the System Identification app (see “Preparing Data for Nonlinear
Identification” on page 5-6).
• (Optional) Choose a nonlinearity estimator in “Nonlinearity Estimators for Nonlinear
ARX Models” on page 5-9.

1 In the System Identification app, select Estimate > Nonlinear models to open the
Nonlinear Models dialog box.
2 In the Configure tab, verify that Nonlinear ARX is selected in the Model type
list.
3 (Optional) Edit the Model name by clicking . The name of the model should be
unique to all nonlinear ARX models in the System Identification app.
4 (Optional) If you want to refine the parameters of a previously estimated model or
configure the model structure to match that of an existing model:

5-14
 Prerequisites

a Click Initialize. A Initial Model Specification dialog box opens.

b In the Initial Model drop-down list, select a nonlinear ARX model.

The model must be in the Model Board of the System Identification app and the
input/output dimensions of this initial model must match that of the estimation
data, selected as Working Data in the app.
c Click OK.

The model structure as well as the parameter values are updated to match that
of the selected model.

Clicking Estimate causes the estimation to use the parameters of the initial model
as the starting point.

Note: When you select an initial model, you can optionally update the estimation
algorithm settings to match those used for the initial model by selecting the Inherit
the model’s algorithm properties option.
5 Keep the default settings in the Nonlinear Models dialog box that specify the model
structure and the algorithm, or modify these settings:

Note: For more information about available options, click Help in the Nonlinear
Models dialog box to open the app help.

What to Configure Options in Nonlinear Models Comment

GUI
Model order In the Regressors tab, Model order na is the
edit the No. of Terms output number of terms
corresponding to each and nb is the input number
input and output channel. of terms.
Input delay In the Regressors If you do not know the
tab, edit the Delay input delay value, click
corresponding to an input Infer Input Delay. This
channel. action opens the Infer
Input Delay dialog box

5-15
5 Nonlinear Black-Box Model Identification

What to Configure Options in Nonlinear Models Comment

GUI
to suggest possible delay
values.
Regressors In the Regressors tab, This action opens the
click Edit Regressors. Model Regressors dialog
box. Use this dialog box to
create custom regressors
or to include specific
regressors in the nonlinear
block.
Nonlinearity estimator In the Model Properties To use all standard and
tab. custom regressors in the
linear block only, you can
exclude the nonlinear block
by setting Nonlinearity to
None.
Estimation algorithm In the Estimate tab, click  
Algorithm Options.
6 To obtain regularized estimates of model parameters, in the Estimate tab,
click Estimation Options. Specify the regularization constants in the
Regularization_Tradeoff_Constant and Regularization_Weighting fields. To
learn more, see “Regularized Estimates of Model Parameters”.
7 Click Estimate to add this model to the System Identification app.

The Estimate tab displays the estimation progress and results.

8 Validate the model response by selecting the desired plot in the Model Views area
of the System Identification app. For more information about validating models, see
“How to Plot Nonlinear ARX Plots Using the App” on page 5-34.

If you get a poor fit, try changing the model structure or algorithm configuration in
step 5.

You can export the model to the MATLAB workspace by dragging it to To Workspace in
the System Identification app.

5-16
 Prerequisites

How to Estimate Nonlinear ARX Models at the Command Line

Prerequisites
• Learn about the nonlinear ARX model structure in “Structure of Nonlinear ARX
Models” on page 5-8.
• Prepare your data, as described in “Preparing Data for Nonlinear Identification” on
page 5-6.
• (Optional) Estimate model orders and delays the same way you would for linear ARX
models. See “Preliminary Step – Estimating Model Orders and Input Delays” on page
3-46.
• (Optional) Choose a nonlinearity estimator in “Nonlinearity Estimators for Nonlinear
ARX Models” on page 5-9.
• (Optional) Estimate or construct an linear ARX model for initialization of nonlinear
ARX model. See “Using Linear Model for Nonlinear ARX Estimation” on page
5-28.

Estimate model using nlarx.

Use nlarx to both construct and estimate a nonlinear ARX model. After each estimation,
validate the model by comparing it to other models and simulating or predicting the
model response.

Basic Estimation

Start with the simplest estimation using m = nlarx(data,[na nb nk]). For example:
load iddata1;
% na = nb = 2 and nk = 1
m = nlarx(z1,[2 2 1])

m =
Nonlinear ARX model with 1 output and 1 input
Inputs: u1
Outputs: y1
Standard regressors corresponding to the orders
na = 2, nb = 2, nk = 1

5-17
5 Nonlinear Black-Box Model Identification

No custom regressor
Nonlinear regressors:
y1(t-1)
y1(t-2)
u1(t-1)
u1(t-2)
Nonlinearity: wavenet with 1 unit
Loss function: 1.8973
Sample time: 0.1
Estimated model (NLARX)

The second input argument [na nb nk] specify the model orders and delays. By default,
the nonlinearity estimator is the wavelet network (see the wavenet reference page),
which takes all standard regressors as inputs to its linear and nonlinear functions. m is
an idnlarx object.

For MIMO systems, nb, nf, and nk are ny-by-nu matrices. See the nlarx reference page
for more information about MIMO estimation.

Specify a different nonlinearity estimator (for example, sigmoid network).

M = nlarx(z1,[2 2 1],'sigmoid');

Set the Focus property of the idnlarx object estimation to simulation error
minimization.

M = nlarx(z1,[2 2 1],'sigmoid','Focus','simulation');

Configure model regressors.

Standard Regressors

Change the model order to create a model structure with different model regressors,
which are delayed input and output variables that are inputs to the nonlinearity
estimator.

Custom Regressors

Explore including custom regressors in the nonlinear ARX model structure. Custom
regressors are in addition to the standard model regressors (see “Nonlinear ARX Order
and Delay” on page 5-13).

5-18
 Configure model regressors.

Use polyreg or customreg to construct custom regressors in terms of model input-

output variables. You can specify custom regressors using the CustomRegressors
property of the idnlarx class or addreg to append custom regressors to an existing
model.

For example, generate regressors as polynomial functions of inputs and outputs:

load iddata1
m = nlarx(z1,[2 2 1],'sigmoidnet');
getreg(m) % displays all regressors
% Generate polynomial regressors up to order 2:
reg = polyreg(m)

Regressors:
y1(t-1)
y1(t-2)
u1(t-1)
u1(t-2)

4x1 array of Custom Regressors with fields: Function, Arguments, Delays, Vectorized.

Append polynomial regressors to CustomRegressors.

m = addreg(m,reg);
getreg(m)

Regressors:
y1(t-1)
y1(t-2)
u1(t-1)
u1(t-2)
y1(t-1).^2
y1(t-2).^2
u1(t-1).^2
u1(t-2).^2

m now includes polynomial regressors.

You can also specify arbitrary functions of input and output variables. For example:

load iddata1
m = nlarx(z1,[2 2 1],'sigmoidnet','CustomReg',{'y1(t-1)^2','y1(t-2)*u1(t-3)'});
getreg(m) % displays all regressors

5-19
5 Nonlinear Black-Box Model Identification

Regressors:
y1(t-1)
y1(t-2)
u1(t-1)
u1(t-2)
y1(t-1)^2
y1(t-2)*u1(t-3)

Append polynomial regressors to CustomRegressors.

m = addreg(m,reg);
getreg(m) % polynomial regressors

Regressors:
y1(t-1)
y1(t-2)
u1(t-1)
u1(t-2)
y1(t-1)^2
y1(t-2)*u1(t-3)
y1(t-1).^2
y1(t-2).^2
u1(t-1).^2
u1(t-2).^2

Manipulate custom regressors using the CustomRegressors property of the idnlarx

class. For example, to get the function handle of the first custom regressor in the array:

CReg1 = m.CustomReg(1).Function;

To view the regressor expression as a string, use:

m.CustomReg(1).Display

ans =

y1(t-1)^2

You can exclude all standard regressors and use only custom regressors in the model
structure by setting na=nb=nk=0:

5-20
 Linear and nonlinear regressors.

m = nlarx(z1,[0 0 0],'linear','CustomReg',{'y1(t-1)^2','y1(t-2)*u1(t-3)'});

In advanced applications, you can specify advanced estimation options for nonlinearity
estimators. For example, wavenet and treepartition classes provide the Options
property for setting such estimation options.

Linear and nonlinear regressors.

By default, all model regressors enter as inputs to both linear and nonlinear function
blocks of the nonlinearity estimator. To reduce model complexity and keep the estimation
well-conditioned, use a subset of regressors as inputs to the nonlinear function of the
nonlinear estimator block. For example, specify a nonlinear ARX model to be linear in
past outputs and nonlinear in past inputs.
m = nlarx(z1,[2 2 1]); % all standard regressors are
% inputs to the nonlinear function
getreg(m); % lists all standard regressors
m = nlarx(z1,[4 4 1],sigmoidnet,'nlreg',[5 6 7 8]);

Regressors:
y1(t-1)
y1(t-2)
u1(t-1)
u1(t-2)

This example uses getreg to determine the index of each regressor from the complete
list of all model regressors. Only regressor numbers 5 through 8 are inputs to the
nonlinear function - getreg shows that these regressors are functions of the input
variable u1. nlreg is an abbreviation for the NonlinearRegressors property of the
idnlarx class. Alternatively, include only input regressors in the nonlinear function
block using:
m = nlarx(z1,[4 4 1],sigmoidnet,'nlreg','input');

When you are not sure which regressors to include as inputs to the nonlinear function
block, specify to search during estimation for the optimum regressor combination:
m = nlarx(z1,[4 4 1],sigmoidnet,'nlreg','search');

This search typically takes a long time. You can display the search progress using m =
nlarx(z1,[4 4 1],sigmoidnet,'nlreg','search','Display','on') . After

5-21
5 Nonlinear Black-Box Model Identification

estimation, use m.NonlinearRegressors to view which regressors were selected by the

automatic regressor search.

m.NonlinearRegressors

ans =

3 5 7 8

Configure the nonlinearity estimator.

Specify the nonlinearity estimator directly in the estimation command as:

• A string of the nonlinearity name, which uses the default nonlinearity configuration.

m = nlarx(z1, [2 2 1],'sigmoidnet');

or

m = nlarx(z1,[2 2 1],'sig'); % abbreviated string

• Nonlinearity object.

m = nlarx(z1,[2 2 1],wavenet('num',5));

This estimation uses a nonlinear ARX model with a wavelet nonlinearity that has
5 units. To construct the nonlinearity object before providing it as an input to the
nonlinearity estimator:

w = wavenet('num', 5);
m = nlarx(z1,[2 2 1],w);
% or
w = wavenet;
w.NumberOfUnits = 5;
m = nlarx(z1,[2 2 1],w);

For MIMO systems, you can specify a different nonlinearity for each output. For
example, to specify sigmoidnet for the first output and wavenet for the second output:

load iddata1 z1
load iddata2 z2

5-22
 Include only nonlinear function in nonlinearity estimator.

data = [z1, z2(1:300)];

M = nlarx(data,[[1 1;1 1] [2 1;1 1] [2 1;1 1]],[sigmoidnet wavenet]);

If you want the same nonlinearity for all output channels, specify one nonlinearity.
M = nlarx(data,[[1 1;1 1] [2 1;1 1] [2 1;1 1]],sigmoidnet);

The following table summarizes values that specify nonlinearity estimators.

Nonlinearity Value (Default Nonlinearity Configuration) Class

Wavelet network 'wavenet' or 'wave' wavenet
(default)
One layer sigmoid 'sigmoidnet' or 'sigm' sigmoidnet
network
Tree partition 'treepartition' or 'tree' treepartition
F is linear in x 'linear' or [ ] linear

Additional available nonlinearities include multilayered neural networks and custom

networks that you create.

Specify a multilayered neural network using:

m = nlarx(data,[na nb nk],NNet)

where NNet is the neural network object you create using the Neural Network Toolbox
software. See the neuralnet reference page.

Specify a custom network by defining a function called gaussunit.m, as described in the

customnet reference page. Define the custom network object CNetw and estimate the
model:
CNetw = customnet(@gaussunit);
m = nlarx(data,[na nb nk],CNetw)

Include only nonlinear function in nonlinearity estimator.

If your model includes wavenet, sigmoidnet, and customnet nonlinearity estimators,
you can exclude the linear function using the LinearTerm property of the nonlinearity
estimator. The nonlinearity estimator becomes F(x)= g ( Q( x - r) ) .

5-23
5 Nonlinear Black-Box Model Identification

Nonlinearity Estimator

u
Regressors Nonlinear y
u(t),u(t-1),y(t-1), ... Function

For example:
SNL = sigmoidnet('LinearTerm','off');
m = nlarx(z1,[2 2 1],SNL);

Note: You cannot exclude the linear function from tree partition and neural network
nonlinearity estimators.

Include only linear function in nonlinearity estimator.

Configure the nonlinear ARX structure to include only the linear function in
the nonlinearity estimator by setting the nonlinearity to linear. In this case,
F(x)= LT ( x) + d is a weighted sum of model regressors plus an offset. Such models provide
a bridge between purely linear ARX models and fully flexible nonlinear models.

Nonlinearity Estimator

u
Regressors Linear y
u(t),u(t-1),y(t-1), ... Function

In the simplest case, a model with only standard regressors is linear (affine). For
example, this structure:

5-24
 Iteratively refine the model.

m = nlarx(z1,[2 2 1],'linear');

is similar to the linear ARX model:

lin_m = arx(z1,[2 2 1]);

However, the nonlinear ARX model m is more flexible than the linear ARX model lin_m
because it contains the offset term, d. This offset term provides the additional flexibility
of capturing signal offsets, which is not available in linear models.

A popular nonlinear ARX configuration in many applications uses polynomial regressors

to model system nonlinearities. In such cases, the system is considered to be a linear
combination of products of (delayed) input and output variables. Use the polyreg
command to easily generate combinations of regressor products and powers.

For example, suppose that you know the output y(t) of a system to be a linear
combination of (y(t − 1))2 and y(t − 2)*u(t − 3). To model such a system, use:

M = nlarx(z1,[0 0 0],'linear','CustomReg',{'y1(t-1)^2','y1(t-2)*u1(t-3)'});

M has no standard regressors and the nonlinearity in the model is described only by the
custom regressors.

Iteratively refine the model.

If your model structure includes nonlinearities that support iterative search (see
“Estimation Algorithm for Nonlinear ARX Models” on page 5-13), you can use pem to
refine model parameters:

m = nlarx(z1,[2 2 1],'sigmoidnet');
m2 = pem(z1,m);

You can also use nlarx to refine the original model:

m1 = nlarx(z1, [2 2 1],'sigmoidnet');
m2 = nlarx(z1,m1); % can repeatedly run this command

Check the search termination criterion m.EstimationInfo.WhyStop . If WhyStop

indicates that the estimation reached the maximum number of iterations, try repeating
the estimation and possibly specifying a larger value for the MaxIter idnlarx property:

5-25
5 Nonlinear Black-Box Model Identification

m2 = pem(z1,m1,'MaxIter',30); % runs 30 more iterations

% starting from m1

When the m.EstimationInfo.WhyStop value is Near (local) minimum,

(norm( g) < tol or No improvement along the search direction with
line search , validate your model to see if this model adequately fits the data. If
not, the solution might be stuck in a local minimum of the cost-function surface. Try
adjusting the Algorithm.Tolerance property value of the idnlarx class or the
Algorithm.SearchMethod property, and repeat the estimation.

You can also try perturbing the parameters of the last model using init (called
randomization) and refining the model using pem:

M1 = nlarx(z1, [2 2 1], 'sigm'); % original model

M1p = init(M1); % randomly perturbs parameters about nominal values
M2 = pem(z1, M1p); % estimates parameters of perturbed model

You can display the progress of the iterative search in the MATLAB Command Window
using the Display property of the idnlarx class:

M2= pem(z1,M1p,'Display','On');

What if you cannot get a satisfactory model?

If you do not get a satisfactory model after many trials with various model structures
and algorithm settings, it is possible that the data is poor. For example, your data might
be missing important input or output variables and does not sufficiently cover all the
operating points of the system.

Nonlinear black-box system identification usually requires more data than linear model
identification to gain enough information about the system.

Use nlarx to Estimate Nonlinear ARX Models

This example shows how to use nlarx to estimate a nonlinear ARX model for measured
input/output data.

Prepare the data for estimation.

load twotankdata

5-26
 What if you cannot get a satisfactory model?

z = iddata(y, u, 0.2);
ze = z(1:1000); zv = z(1001:3000);

Estimate several models using different model orders, delays, and nonlinearity settings.

m1 = nlarx(ze,[2 2 1]);
m2 = nlarx(ze,[2 2 3]);
m3 = nlarx(ze,[2 2 3],wavenet('num',8));
m4 = nlarx(ze,[2 2 3],wavenet('num',8),...
'nlr', [1 2]);

An alternative way to perform the estimation is to configure the model structure first,
and then to estimate this model.

m5 = idnlarx([2 2 3],sigmoidnet('num',14),'nlr',[1 2]);

m5 = pem(ze,m5);

Compare the resulting models by plotting the model outputs with the measured output.

compare(zv, m1,m2,m3,m4,m5)

5-27
5 Nonlinear Black-Box Model Identification

Using Linear Model for Nonlinear ARX Estimation

• “About Using Linear Models” on page 5-28
• “How to Initialize Nonlinear ARX Estimation Using Linear ARX Models” on page
5-29
• “Estimate Nonlinear ARX Models Using Linear ARX Models” on page 5-30

About Using Linear Models

You can use an ARX structure polynomial model (idpoly with only A and B as active
polynomials) for nonlinear ARX estimation.

5-28
 What if you cannot get a satisfactory model?

Tip To learn more about when to use linear models, see “When to Fit Nonlinear Models”
on page 5-2.

Typically, you create a linear ARX model using the arx command. You can provide the
linear model only at the command line when constructing (see idnlarx) or estimating
(see nlarx) a nonlinear ARX model.

The software uses the linear model for initializing the nonlinear ARX estimation:

• Assigns the linear model orders as initial values of nonlinear model orders (na and
nb properties of the idnlarx object) and delays (nk property) to compute standard
regressors in the nonlinear ARX model structure.
• Uses the A and B polynomials of the linear model to compute the linear function of
the nonlinearity estimators (LinearCoef parameter of the nonlinearity estimator
object), except for neural network nonlinearity estimator.

During estimation, the estimation algorithm uses these values to further adjust the
nonlinear model to the data. The initialization always provides a better fit to the
estimation data than the linear ARX model.

How to Initialize Nonlinear ARX Estimation Using Linear ARX Models

Estimate a nonlinear ARX model initialized using a linear model by typing

m = nlarx(data,LinARXModel)

LinARXModel is an idpoly object of ARX structure. m is an idnlarx object. data is a

time-domain iddata object.

By default, the nonlinearity estimator is the wavelet network (wavenet object). This
network takes all standard regressors computed using orders and delay of LinARXModel
as inputs to its linear and nonlinear functions. The software computes the LinearCoef
parameter of the wavenet object using the A and B polynomials of the linear ARX model.

Tip When you use the same data set, a nonlinear ARX model initialized using a linear
ARX model produces a better fit than the linear ARX model.

Specify a different nonlinearity estimator, for example a sigmoid network:

5-29
5 Nonlinear Black-Box Model Identification

m = nlarx(data,LinARXModel,'sigmoid')

Set the Focus property of the idnlarx object estimation to simulation error
minimization:

m = nlarx(data,LinARXModel,'sigmoid','Focus','simulation')

After each estimation, validate the model by comparing the simulated response to the
data. To improve the fit of the nonlinear ARX model, adjust various elements of the
nonlinear ARX structure. For more information, see “Ways to Configure Nonlinear ARX
Estimation” on page 5-11.

Estimate Nonlinear ARX Models Using Linear ARX Models

This example shows how to estimate nonlinear ARX models by using linear ARX models.

Load the estimation data.

load throttledata.mat

This command loads the data object ThrottleData into the workspace. The object
contains input and output samples collected from an engine throttle system, sampled at a
rate of 100 Hz.

A DC motor controls the opening angle of the butterfly valve in the throttle system. A
step signal (in volts) drives the DC motor. The output is the angular position (in degrees)
of the valve.

Plot the data to view and analyze the data characteristics.

plot(ThrottleData)

5-30
 What if you cannot get a satisfactory model?

In the normal operating range of 15-90 degrees, the input and output variables have a
linear relationship. You use a linear model of low order to model this relationship.

In the throttle system, a hard stop limits the valve position to 90 degrees, and a
spring brings the valve to 15 degrees when the DC motor is turned off. These physical
components introduce nonlinearities that a linear model cannot capture.

Estimate an ARX model to model the linear behavior of this single-input single-output
system in the normal operating range.

% Detrend the data because linear models cannot capture offsets.

Tr = getTrend(ThrottleData);
Tr.OutputOffset = 15;
DetrendedData = detrend(ThrottleData,Tr);

5-31
5 Nonlinear Black-Box Model Identification

% Estimate a linear ARX model with na=2, nb=1, nk=1.

opt = arxOptions('Focus','simulation');
LinearModel = arx(DetrendedData,[2 1 1],opt);

Compare the simulated model response with estimation data.

compare(DetrendedData, LinearModel)

The linear model captures the rising and settling behavior in the linear operating range
but does not account for output saturation at 90 degrees.

Estimate a nonlinear ARX model to model the output saturation.

NonlinearModel = nlarx(ThrottleData, LinearModel, 'sigmoidnet',...
'Focus', 'Simulation');

5-32
 What if you cannot get a satisfactory model?

The software uses the orders and delay of the linear model for the orders of the
nonlinear model. In addition, the software computes the linear function of sigmoidnet
nonlinearity estimator.

Compare the nonlinear model with data.

compare(ThrottleData, NonlinearModel)

The model captures the nonlinear effects (output saturation) and improves the overall fit
to data.

Validating Nonlinear ARX Models

• “About Nonlinear ARX Plots” on page 5-34

5-33
5 Nonlinear Black-Box Model Identification

• “How to Plot Nonlinear ARX Plots Using the App” on page 5-34
• “How to Validate Nonlinear ARX Models at the Command Line” on page 5-35
• “Configuring the Nonlinear ARX Plot” on page 5-38
• “Axis Limits, Legend, and 3-D Rotation” on page 5-39

About Nonlinear ARX Plots

The Nonlinear ARX plot displays the characteristics of model nonlinearities as a

function of one or two regressors. The model nonlinearity (model.Nonlinearity) is a
nonlinearity estimator function, such as wavenet, sigmoidnet, treepartition, and
uses model regressors as its inputs. The value of the nonlinearity is plotted by projecting
its response in 2 or 3-dimensional space. The plot uses one or two regressors as the plot
axes for 2- or 3-D plots, respectively and a center point (cross-section location) for the
other regressors.

Examining a nonlinear ARX plot can help you gain insight into which regressors have
the strongest effect on the model output. Understanding the relative importance of the
regressors on the output can help you decide which regressors should be included in the
nonlinear function.

Furthermore, you can create several nonlinear models for the same data set using
different nonlinearity estimators, such a wavenet network and treepartition, and
then compare the nonlinear surfaces of these models. Agreement between nonlinear
surfaces increases the confidence that these nonlinear models capture the true dynamics
of the system.

In the plot window, you can choose:

• The regressors to use on the plot axes, and specify the center points for the other
regressors in the configuration panel. For multi-output models, each output is plotted
separately.
• The output to view from the drop-down list located at the top of the plot.

How to Plot Nonlinear ARX Plots Using the App

You can plot linear and nonlinear blocks of nonlinear ARX models.

To create a nonlinear ARX plot in the System Identification app, select the Nonlinear
ARX check box in the Model Views area. For general information about creating and
working with plots, see “Working with Plots” on page 14-11.

5-34
 Compare Model Output to Measured Output

Note: The Nonlinear ARX check box is unavailable if you do not have a nonlinear ARX
model in the Model Board.

The following figure shows a sample nonlinear ARX plot.

How to Validate Nonlinear ARX Models at the Command Line

You can use the following approaches to validate nonlinear ARX models at the command
line:

Compare Model Output to Measured Output

Compare estimated models using compare. Use an independent validation data
set whenever possible. For more information about validating models, see “Model
Validation”.

For example, compare linear and nonlinear ARX models of same order:

5-35
5 Nonlinear Black-Box Model Identification

load iddata1
LM = arx(z1,[2 2 1]); % estimates linear ARX model
M = nlarx(z1,[2 2 1],'sigmoidnet'); % estimates nonlinear ARX model
compare(z1,LM,M) % compares responses of LM and M against measured data

Compare the performance of several models using the properties

M.EstimationInfo.FPE (final prediction error) and M.EstimationInfo.LossFcn
(value of loss function at estimation termination). Smaller values typically indicate better
performance. However, m.EstimationInfo.FPE values might be unreliable when the
model contains a large number of parameters relative to the estimation data size.

5-36
 Simulate and Predict Model Response

Simulate and Predict Model Response

Use sim(idnlarx) and predict to simulate and predict model response, respectively.
To compute the step response of the model, use step. See the corresponding reference
page for more information.

Analyze Residuals
Residuals are differences between the one-step-ahead predicted output from the model
and the measured output from the validation data set. Thus, residuals represent the
portion of the validation data output not explained by the model. Use resid to compute
and plot the residuals.

Plot Nonlinearity
Use plot to view the shape of the nonlinearity. For example:

plot(M)

where M is the nonlinear ARX (idnlarx) model. The plot command opens the Nonlinear
ARX Model Plot window.

If the shape of the plot looks like a plane for all the chosen regressor values,
then the model is probably linear in those regressors. In this case, you can
remove the corresponding regressors from nonlinear block by specifying the
M.NonlinearRegressors property and repeat the estimation.

You can use additional plot arguments to specify the following information:

• Include multiple nonlinear ARX models on the plot.

• Configure the regressor values for computing the nonlinearity values.

Check Iterative Search Termination Conditions

If your idnlarx model structure uses iterative search to minimize prediction or
simulation errors, use M.EstimationInfo to display the estimation termination

5-37
5 Nonlinear Black-Box Model Identification

conditions. M is the estimated idnlarx model. For example, check the WhyStop field
of the EstimationInfo property, which describes why the estimation stopped—the
algorithm might have reached the maximum number of iterations or the required
tolerance value. For more information about iterative search, see “Estimation Algorithm
for Nonlinear ARX Models” on page 5-13.

Configuring the Nonlinear ARX Plot

To include or exclude a model on the plot, click the corresponding model icon in the
System Identification app. Active models display a thick line inside the Model Board
icon.

To configure the plot:

1 If your model contains multiple outputs, select the output channel in the Select
nonlinearity at output drop-down list. Selecting the output channel displays the
nonlinearity values that correspond to this output channel.
2
If the regressor selection options are not visible, click to expand the Nonlinear
ARX Model Plot window.
3 Select Regressor 1 from the list of available regressors. In the Range field, enter
the range of values to include on the plot for this regressor. The regressor values are
plotted on the Reg1 axis.
4 Specify a second regressor for a 3-D plot by selecting one of the following types of
options:

• Select Regressor 2 to display three axes. In the Range field, enter the range of
values to include on the plot for this regressor. The regressor values are plotted
on the Reg2 axis.
• Select <none> in the Regressor 2 list to display only two axes.
5 To fix the values of the regressor that are not displayed, click Fix Values. In the Fix
Regressor Values dialog box, double-click the Value cell to edit the constant value
of the corresponding regressor. The default values are determined during model
estimation. Click OK.
6 Click Apply to update the plot.

1 To change the grid of the regressor space along each axis, Options > Set number
of samples, and enter the number of samples to use for each regressor. Click Apply
and then Close.

5-38
 Simulation and Prediction

For example, if the number of samples is 20, each regressor variable contains
20 points in its specified range. For a 3-D plots, this results in evaluating the
nonlinearity at 20 x 20 = 400 points.

Axis Limits, Legend, and 3-D Rotation

The following table summarizes the commands to modify the appearance of the
Nonlinear ARX plot.

Changing Appearance of the Nonlinear ARX Plot

Action Command
Change axis limits. Select Options > Set axis limits to open
the Axis Limits dialog box, and edit the
limits. Click Apply.
Hide or show the legend. Select Style > Legend. Select this option
again to show the legend.
(Three axes only) Select Style > Rotate 3D and drag the
Rotate in three dimensions. axes on the plot to a new orientation. To
disable three-dimensional rotation, select
Style > Rotate 3D again.
Note: Available only when you have
selected two regressors as independent
variables.

Using Nonlinear ARX Models

Simulation and Prediction

Use sim(idnlarx) to simulate the model output, and predict to predict the model
output. To compare models to measured output and to each other, use compare.

Simulation and prediction commands provide default handling of the model's initial
conditions, or initial state values. See the idnlarx reference page for a definition of the
nonlinear ARX model states.

This toolbox provides several options to facilitate how you specify initial states. For
example, you can use findstates(idnlarx) and data2state(idnlarx) to compute

5-39
5 Nonlinear Black-Box Model Identification

state values based on operating conditions or the requirement to maximize fit to

measured output.

To learn more about how sim and predict compute the model output, see “How the
Software Computes Nonlinear ARX Model Output” on page 5-40.

Linearization
Compute linear approximation of nonlinear ARX models using linearize or linapp.

linearize provides a first-order Taylor series approximation of the system about

an operation point (also called tangent linearization). linapp computes a linear
approximation of a nonlinear model for a given input data. For more information, see the
“Linear Approximation of Nonlinear Black-Box Models” on page 5-81.

You can compute the operating point for linearization using findop(idnlarx).

After computing a linear approximation of a nonlinear model, you can perform linear
analysis and control design on your model using Control System Toolbox commands. For
more information, see “Using Identified Models for Control Design Applications” and
“Create and Plot Identified Models Using Control System Toolbox Software”.

Simulation and Code Generation Using Simulink

You can import estimated Nonlinear ARX models into the Simulink software using the
Nonlinear ARX block (IDNLARX Model) from the System Identification Toolbox block
library. Import the idnlarx object from the workspace into Simulink using this block to
simulate the model output.

The IDNLARX Model block supports code generation with Simulink Coder™ software,
using both generic and embedded targets. Code generation does not work when the model
contains customnet or neuralnet nonlinearity estimator, or custom regressors.

How the Software Computes Nonlinear ARX Model Output

In most applications, sim(idnlarx) and predict are sufficient for computing the
simulated and predicted model response, respectively. This advanced topic describes

5-40
 Estimating and Exploring a Nonlinear ARX Model

how the software evaluates the output of nonlinearity estimators and uses this output to
compute the model response.

Evaluating Nonlinearities

Evaluating the predicted output of a nonlinearity for a specific regressor value x requires
that you first extract the nonlinearity F and regressors from the model:
F = get(m,'Nonlinearity') % equivalent to F = m.nl
x = getreg(m,'all',data) % computes regressors

Evaluate F(x):
y = evaluate(F,x)

where x is a row vector of regressor values.

You can also evaluate predicted output values at multiple time instants by evaluating F
for several regressor vectors simultaneously:
y = evaluate(F,[x1;x2;x3])

Low-Level Simulation and Prediction of Sigmoid Network

This example shows how the software computes the simulated and predicted output
of the model as a result of evaluating the output of its nonlinearity estimator for given
regressor values.

Estimating and Exploring a Nonlinear ARX Model

Estimate nonlinear ARX model with sigmoid network nonlinearity.
load twotankdata
estData = iddata(y,u,0.2,'Tstart',0);
M = nlarx(estData,[1 1 0],'sig');

Inspect the model properties and estimation result.

present(M)

M =
Nonlinear ARX model with 1 output and 1 input

5-41
5 Nonlinear Black-Box Model Identification

Inputs: u1
Outputs: y1
Standard regressors corresponding to the orders
na = 1, nb = 1, nk = 0
No custom regressor
Nonlinear regressors:
y1(t-1)
u1(t)
Nonlinearity: sigmoidnet with 10 units
Loss function: 4.5903e-05
Sample time: 0.2
Estimated model (NLARX)

Created: 20-Sep-2014 00:50:28

Last modified: 20-Sep-2014 00:50:29

This command provides information about input and output variables, regressors, and
nonlinearity estimator.

Inspect the nonlinearity estimator.

NL = M.Nonlinearity; % equivalent to M.nl

class(NL) % nonlinearity class
display(NL) % equivalent to NL

ans =

sigmoidnet

Sigmoid Network:
NumberOfUnits: 10
LinearTerm: 'on'
Parameters: [1x1 struct]

Inspect the sigmoid network parameter values.

NL.Parameters;

Prediction of Output
The model output is:

5-42
 Prediction of Output

y1(t)= f(y1(t-1),u1(t))

where f is the sigmoid network function. The model regressors y1(t-1) and u1(t) are
inputs to the nonlinearity estimator. Time t is a discrete variable representing kT , where
k = 0, 1,…, and T is the sampling interval. In this example, T=0.2 second.

The output prediction equation is:

yp(t)=f(y1_meas(t-1),u1_meas(t))

where yp(t) is the predicted value of the response at time t. y1_meas(t-1) and u1_meas(t)
are the measured output and input values at times t-1 and t, respectively.

Computing the predicted response includes:

• Computing regressor values from input-output data.

• Evaluating the nonlinearity for given regressor values.

To compute the predicted value of the response using initial conditions and current input:

Estimate model from data and get nonlinearity parameters.

load twotankdata
estData = iddata(y,u,0.2,'Tstart',0);
M = nlarx(estData,[1 1 0],'sig');
NL = M.Nonlinearity;

Specify zero initial states.

x0 = 0;

The model has one state because there is only one delayed term y1(t-1). The number of
states is equal to sum(getDelayInfo(M)).

Compute the predicted output at time t=0.

RegValue = [0,estData.u(1)]; % input to nonlinear function f
yp_0 = evaluate(NL,RegValue);

RegValue is the vector of regressors at t=0 . The predicted output is

yp(t=0)=f(y1_meas(t=-1),u1_meas(t=0)). In terms of MATLAB variables, this output is
f(0,estData.u(1)), where

• y1_meas(t=0) is the measured output value at t=0, which is to estData.y(1).

5-43
5 Nonlinear Black-Box Model Identification

• u1_meas(t =1) is the second input data sample estData.u(2).

Perform one-step-ahead prediction at all time values for which data is available.

RegMat = getreg(M,[],estData,x0);
yp = evaluate(NL,RegMat);

This code obtains a matrix of regressors RegMat for all the time samples using getreg.
RegMat has as many rows as there are time samples, and as many columns as there are
regressors in the model—two, in this example.

These steps are equivalent to the predicted response computed in a single step using
predict:

yp = predict(M,estData,1,'InitialState',x0);

Simulation of Output
The model output is:

y1(t)=f(y1(t-1),u1(t))

where f is the sigmoid network function. The model regressors y1(t-1) and u1(t) are
inputs to the nonlinearity estimator. Time t is a discrete variable representing kT , where
k= 0, 1,.., and T is the sampling interval. In this example, T=0.2 second.

The simulated output is:

ys(t) = f(ys(t-1),u1_meas(t))

where ys(t) is the simulated value of the response at time t. The simulation equation is
the same as the prediction equation, except that the past output value ys(t-1) results
from the simulation at the previous time step, rather than the measured output value.

Computing the simulated response includes:

• Computing regressor values from input-output data using simulated output values.
• Evaluating the nonlinearity for given regressor values.

To compute the simulated value of the response using initial conditions and current
input:

5-44
 Simulation of Output

Estimate model from data and get nonlinearity parameters.

load twotankdata
estData = iddata(y,u,0.2,'Tstart',0);
M = nlarx(estData,[1 1 0],'sig');
NL = M.Nonlinearity;

Specify zero initial states.

x0 = 0;

The model has one state because there is only one delayed term y1(t-1). The number of
states is equal to sum(getDelayInfo(M)).

Compute the simulated output at time t =0, ys(t=0).

RegValue = [0,estData.u(1)];
ys_0 = evaluate(NL,RegValue);

RegValue is the vector of regressors at t=0. ys(t=0)=f(y1(t=-1),u1_meas(t=0)). In terms of

MATLAB variables, this output is f(0,estData.u(1)), where

• y1(t=-1) is the initial state x0 (=0).

• u1_meas(t=0) is the value of the input at t =0, which is the first input data sample
estData.u(1).

Compute the simulated output at time t=1, ys(t=1).

RegValue = [ys_0,estData.u(2)];
ys_1 = evaluate(NL,RegValue);

The simulated output ys(t=1)=f(ys(t=0),u1_meas(t=1)). In terms of MATLAB variables,

this output is f(ys_0,estData.u(2)), where

• ys(t=0) is the simulated value of the output at t=0.

• u1_meas(t=1) is the second input data sample estData.u(2).

Compute the simulated output at time t=2.

RegValue = [ys_1,estData.u(3)];
ys_2 = evaluate(NL,RegValue);

Unlike for output prediction, you cannot use getreg to compute regressor values for all
time values. You must compute regressors values at each time sample separately because

5-45
5 Nonlinear Black-Box Model Identification

the output samples required for forming the regressor vector are available iteratively,
one sample at a time.

These steps are equivalent to the simulated response computed in a single step using
sim(idnlarx):

ys = sim(M,estData,x0);

Low-Level Nonlinearity Evaluation

This examples performs a low-level computation of the nonlinearity response for the
sigmoidnet network function:

where f is the sigmoid function, given by the following equation:

In F(x), the input to the sigmoid function is x-r. x is the regressor value and r is
regressor mean, computed from the estimation data. , , and are the network
parameters stored in the model property M.nl.par, where M is an idnlarx object.

Compute the output value at time t=1, when the regressor values are
x=[estData.y(1),estData.u(2)]:

Estimate model from sample data.

load twotankdata
estData = iddata(y,u,0.2,'Tstart',0);
M = nlarx(estData,[1 1 0],'sig');
NL = M.Nonlinearity;

Assign values to the parameters in the expression for F(x).

x = [estData.y(1),estData.u(2)]; % regressor values at t=1

r = NL.Parameters.RegressorMean;
P = NL.Parameters.LinearSubspace;
L = NL.Parameters.LinearCoef;

5-46
 Low-Level Nonlinearity Evaluation

d = NL.Parameters.OutputOffset;
Q = NL.Parameters.NonLinearSubspace;
aVec = NL.Parameters.OutputCoef; %[a_1; a_2; ...]
cVec = NL.Parameters.Translation; %[c_1; c_2; ...]
bMat = NL.Parameters.Dilation; %[b_1; b_2; ...]

Compute the linear portion of the response (plus offset).

yLinear = (x-r)*P*L+d;

Compute the nonlinear portion of the response.

f = @(z)1/(exp(-z)+1); % anonymous function for sigmoid unit
yNonlinear = 0;
for k = 1:length(aVec)
fInput = (x-r)*Q* bMat(:,k)+cVec(k);
yNonlinear = yNonlinear+aVec(k)*f(fInput);
end

Compute total response y = F(x) = yLinear + yNonlinear.

y = yLinear + yNonlinear;

y is equal to evaluate(NL,x).

5-47
5 Nonlinear Black-Box Model Identification

Identifying Hammerstein-Wiener Models

In this section...
“Applications of Hammerstein-Wiener Models” on page 5-48
“Structure of Hammerstein-Wiener Models” on page 5-49
“Nonlinearity Estimators for Hammerstein-Wiener Models” on page 5-50
“Ways to Configure Hammerstein-Wiener Estimation” on page 5-52
“Estimation Algorithm for Hammerstein-Wiener Models” on page 5-53
“How to Estimate Hammerstein-Wiener Models in the System Identification App” on
page 5-54
“How to Estimate Hammerstein-Wiener Models at the Command Line” on page 5-56
“Using Linear Model for Hammerstein-Wiener Estimation” on page 5-64
“Validating Hammerstein-Wiener Models” on page 5-69
“Using Hammerstein-Wiener Models” on page 5-77
“How the Software Computes Hammerstein-Wiener Model Output” on page 5-78

Applications of Hammerstein-Wiener Models

When the output of a system depends nonlinearly on its inputs, sometimes it is possible
to decompose the input-output relationship into two or more interconnected elements. In
this case, you can represent the dynamics by a linear transfer function and capture the
nonlinearities using nonlinear functions of inputs and outputs of the linear system. The
Hammerstein-Wiener model achieves this configuration as a series connection of static
nonlinear blocks with a dynamic linear block.

Hammerstein-Wiener model applications span several areas, such as modeling electro-

mechanical system and radio frequency components, audio and speech processing and
predictive control of chemical processes. These models are popular because they have
a convenient block representation, transparent relationship to linear systems, and are
easier to implement than heavy-duty nonlinear models (such as neural networks and
Volterra models).

You can use the Hammerstein-Wiener model as a black-box model structure because
it provides a flexible parameterization for nonlinear models. For example, you might
estimate a linear model and try to improve its fidelity by adding an input or output
nonlinearity to this model.

5-48
 Identifying Hammerstein-Wiener Models

You can also use a Hammerstein-Wiener model as a grey-box structure to capture

physical knowledge about process characteristics. For example, the input nonlinearity
might represent typical physical transformations in actuators and the output
nonlinearity might describe common sensor characteristics.

Structure of Hammerstein-Wiener Models

This block diagram represents the structure of a Hammerstein-Wiener model:

u(t) Input w(t) Linear x(t) Output y(t)

Nonlinearity Block Nonlinearity
f B/F h

where:

• w(t) = f(u(t)) is a nonlinear function transforming input data u(t). w(t) has the same
dimension as u(t).
• x(t) = (B/F)w(t) is a linear transfer function. x(t) has the same dimension as y(t).

where B and F are similar to polynomials in the linear Output-Error model, as

described in “What Are Polynomial Models?”.

For ny outputs and nu inputs, the linear block is a transfer function matrix containing
entries:

B j ,i (q)
Fj ,i ( q)

where j = 1,2,...,ny and i = 1,2,...,nu.

• y(t) = h(x(t)) is a nonlinear function that maps the output of the linear block to the
system output.

w(t) and x(t) are internal variables that define the input and output of the linear block,
respectively.

Because f acts on the input port of the linear block, this function is called the input
nonlinearity. Similarly, because h acts on the output port of the linear block, this function
is called the output nonlinearity. If system contains several inputs and outputs, you must
define the functions f and h for each input and output signal.

5-49
5 Nonlinear Black-Box Model Identification

You do not have to include both the input and the output nonlinearity in the
model structure. When a model contains only the input nonlinearity f, it is called a
Hammerstein model. Similarly, when the model contains only the output nonlinearity h),
it is called a Wiener model.

The nonlinearities f and h are scalar functions, one nonlinear function for each input and
output channel.

The Hammerstein-Wiener model computes the output y in three stages:

1 Computes w(t) = f(u(t)) from the input data.

w(t) is an input to the linear transfer function B/F.

The input nonlinearity is a static (memoryless) function, where the value of the
output a given time t depends only on the input value at time t.

You can configure the input nonlinearity as a sigmoid network, wavelet network,
saturation, dead zone, piecewise linear function, one-dimensional polynomial, or a
custom network. You can also remove the input nonlinearity.
2 Computes the output of the linear block using w(t) and initial conditions: x(t) = (B/
F)w(t).

You can configure the linear block by specifying the numerator B and denominator F
orders.
3 Compute the model output by transforming the output of the linear block x(t) using
the nonlinear function h: y(t) = h(x(t)).

Similar to the input nonlinearity, the output nonlinearity is a static function.

Configure the output nonlinearity in the same way as the input nonlinearity. You
can also remove the output nonlinearity, such that y(t) = x(t).

Resulting models are idnlhw objects that store all model data, including model
parameters and nonlinearity estimator. See the idnlhw reference page for more
information.

Nonlinearity Estimators for Hammerstein-Wiener Models

System Identification Toolbox software provides several scalar nonlinearity estimators
F(x) for Hammerstein-Wiener models. The nonlinearity estimators are available for both

5-50
 Identifying Hammerstein-Wiener Models

the input and output nonlinearities f and h, respectively. For more information about
F(x), see “Structure of Hammerstein-Wiener Models” on page 5-49.

Each nonlinearity estimator corresponds to an object class in this toolbox. When you
estimate Hammerstein-Wiener models in the app, System Identification Toolbox
creates and configures objects based on these classes. You can also create and configure
nonlinearity estimators at the command line. For a detailed description of each
estimator, see the references page of the corresponding nonlinearity class.

Nonlinearity Class Structure Comments

Piecewise pwlinear A piecewise linear function By default,
linear parameterized by breakpoint locations. the number of
(default) breakpoints is 10.
One layer sigmoidnet n Default number of
sigmoid g( x) = Â a k k (b k ( x - g k ) ) units n is 10.
network k= 1

k( s) is the sigmoid function

-1
(
k ( s) = e s + 1 ) . bk is a row vector such
that bk ( x - g k ) is a scalar.
Wavelet wavenet n By default, the
network g( x) = Â a k k (b k ( x - g k ) ) estimation algorithm
k= 1 determines the
number of units
where k( s) is the wavelet function. nautomatically.

Saturation saturation Parameterize hard limits on the signal Use to model known
value as upper and lower saturation saturation effects on
limits. signal amplitudes.
Dead zone deadzone Parameterize dead zones in signals as the Use to model known
duration of zero response. dead zones in signal
amplitudes.
One- poly1d Single-variable polynomial of a degree By default, the
dimensional that you specify. polynomial degree is
polynomial 1.

5-51
5 Nonlinear Black-Box Model Identification

Nonlinearity Class Structure Comments

Unit gain unitgain Excludes the input or output nonlinearity Useful for configuring
from the model structure to achieve a multi-input, multi-
Wiener or Hammerstein configuration, output (MIMO)
respectively. models to exclude
nonlinearities from
specific input and
Note: Excluding both the input and output channels.
output nonlinearities reduces the
Hammerstein-Wiener structure to a
linear transfer function.
Custom customnet Similar to sigmoid network but you (For advanced use)
network specify k( s) . Uses the unit
(user-defined) function that you
specify.

Ways to Configure Hammerstein-Wiener Estimation

Estimate a Hammerstein-Wiener model with default configuration by:

• Specifying model order and input delay:

• nb—The number of zeros plus one.

• nf—The number of poles.
• nk—The delay from input to the output in terms of the number of samples.

nb is the order of the transfer function numerator (B polynomial), and nf is the

order of the transfer function denominator (F polynomial). As you fit different
Hammerstein-Wiener models to your data, you can configure the linear block
structure by specifying a different order and delay. For MIMO systems with ny
outputs and nu inputs, nb, nf, and nk are ny-by-nu matrices.
• Initializing using one of the following discrete-time linear models:

• An input-output polynomial model of Output-Error (OE) structure (idpoly)

• A linear state-space model with no disturbance component (idss object with K=0)

You can perform this operation only at the command line. The initialization
configures the Hammerstein-Wiener model to use orders and delay of the linear

5-52
 Identifying Hammerstein-Wiener Models

model, and the B and F polynomials as the transfer function numerator and
denominator. See “Using Linear Model for Hammerstein-Wiener Estimation” on page
5-64.

By default, the input and output nonlinearity estimators are both piecewise linear
functions, parameterized by breakpoint locations (see the pwlinear reference page). You
can configure the input and output nonlinearity estimators by:

• Configuring the input and output nonlinearity properties.

• Excluding the input or output nonlinear block.

See these topics for detailed steps to change the model structure:

• “How to Estimate Hammerstein-Wiener Models in the System Identification App” on

page 5-54
• “How to Estimate Hammerstein-Wiener Models at the Command Line” on page
5-56

Estimation Algorithm for Hammerstein-Wiener Models

Estimation of Hammerstein-Wiener models uses iterative search to minimize the
simulation error between the model output and the measured output.

You can configure the estimation method using the Algorithm properties of the idnlhw
class. The most common of these properties are:

• MaxIter — Maximum number of iterations.

• SearchMethod — Search method for minimization of prediction or simulation errors,
such as Gauss-Newton and Levenberg-Marquardt line search, and Trust-region
reflective Newton approach.
• Tolerance — Condition for terminating iterative search when the expected
improvement of the parameter values is less than a specified value.
• Display — Shows progress of iterative minimization in the MATLAB Command
Window.

By default, the initial states of the model are zero and not estimated. However, you can
choose to estimate initial states during model estimation, which sometimes helps to
achieve better results.

5-53
5 Nonlinear Black-Box Model Identification

How to Estimate Hammerstein-Wiener Models in the System Identification

App

Prerequisites
• Learn about the Hammerstein-Wiener model structure (see “Structure of
Hammerstein-Wiener Models” on page 5-49).
• Import data into the System Identification app (see “Preparing Data for Nonlinear
Identification” on page 5-6).
• (Optional) Choose a nonlinearity estimator in “Nonlinearity Estimators for
Hammerstein-Wiener Models” on page 5-50.
• (Optional) Estimate or construct an OE or state-space model to use for initialization.
See “Using Linear Model for Hammerstein-Wiener Estimation” on page 5-64.

1 In the System Identification app, select Estimate > Nonlinear models to open the
Nonlinear Models dialog box.
2 In the Configure tab, select Hammerstein-Wiener from the Model type list.
3 (Optional) Edit the Model name by clicking the pencil icon. The name of the model
should be unique to all Hammerstein-Wiener models in the System Identification
app.
4 (Optional) If you want to refine a previously estimated model, click Initialize to
select a previously estimated model from the Initial Model list.

Note: Refining a previously estimated model starts with the parameter values of the
initial model and uses the same model structure. You can change these settings.

The Initial Model list includes models that:

• Exist in the System Identification app.

• Have the same number of inputs and outputs as the dimensions of the estimation
data (selected as Working Data in the System Identification app).
5 Keep the default settings in the Nonlinear Models dialog box that specify the model
structure and the algorithm, or modify these settings:

5-54
 Prerequisites

Note: For more information about available options, click Help in the Nonlinear
Models dialog box to open the app help.

What to Configure Options in Nonlinear Models Comment

GUI
Input or output In the I/O Nonlinearity If you do not know which
nonlinearity tab, select the nonlinearity to try, use the
Nonlinearity and specify (default) piecewise linear
the No. of Units. nonlinearity.

When you estimate from

binary input data, you
cannot reliably estimate
the input nonlinearity. In
this case, set Nonlinearity
for the input channel to
None.

For multiple-input
and multiple-output
systems, you can assign
nonlinearities to specific
input and output channels.
Model order and delay In the Linear Block If you do not know the
tab, specify B Order, F input delay values, click
Order, and Input Delay. Infer Input Delay. This
For MIMO systems, select action opens the Infer
the output channel and Input Delay dialog box
specify the orders and which suggests possible
delays from each input delay values.
channel.
Estimation algorithm In the Estimate tab, click You can specify to estimate
Algorithm Options. initial states.
6 To obtain regularized estimates of model parameters, in the Estimate tab,
click Estimation Options. Specify the regularization constants in the

5-55
5 Nonlinear Black-Box Model Identification

Regularization_Tradeoff_Constant and Regularization_Weighting fields. To

learn more, see “Regularized Estimates of Model Parameters”.
7 Click Estimate to add this model to the System Identification app.

The Estimate tab displays the estimation progress and results.

8 Validate the model response by selecting the desired plot in the Model Views area
of the System Identification app.

If you get a poor fit, try changing the model structure or algorithm configuration in
step 5.

You can export the model to the MATLAB workspace by dragging it to To Workspace in
the System Identification app.

How to Estimate Hammerstein-Wiener Models at the Command Line

Prerequisites
• Learn about the Hammerstein-Wiener model structure described in “Structure of
Hammerstein-Wiener Models” on page 5-49.
• Prepare your data, as described in “Preparing Data for Nonlinear Identification” on
page 5-6.
• (Optional) Choose a nonlinearity estimator in “Nonlinearity Estimators for
Hammerstein-Wiener Models” on page 5-50.
• (Optional) Estimate or construct an input-output polynomial model of Output-Error
(OE) structure (idpoly) or a state-space model with no disturbance component (idss
with K=0) for initialization of Hammerstein-Wiener model. See “Using Linear Model
for Hammerstein-Wiener Estimation” on page 5-64.

Estimate model using nlhw.

Use nlhw to both construct and estimate a Hammerstein-Wiener model. After each
estimation, validate the model by comparing it to other models and simulating or
predicting the model response.

Basic Estimation

5-56
 Configure the nonlinearity estimator.

Start with the simplest estimation using m = nlhw(data,[nb nf nk]). For example:
load iddata3;
% nb = nf = 2 and nk = 1
m = nlhw(z3,[2 2 1])

m =
Hammerstein-Wiener model with 1 output and 1 input
Linear transfer function corresponding to the orders nb = 2, nf = 2, nk = 1
Input nonlinearity: pwlinear with 10 units
Output nonlinearity: pwlinear with 10 units
Loss function: 1.3593
Sampling interval: 1
Estimated model (NLHW)

The second input argument [nb nf nk] sets the order of the linear transfer function,
where nb is the number of zeros plus 1, nf is the number of poles, and nk is the input
delay. By default, both the input and output nonlinearity estimators are piecewise linear
functions (see the pwlinear reference page). m is an idnlhw object.

For MIMO systems, nb, nf, and nk are ny-by-nu matrices. See the nlhw reference page
for more information about MIMO estimation.

Configure the nonlinearity estimator.

Specify a different nonlinearity estimator using m = nlhw(data,[nb nf
nk],InputNL,OutputNL). InputNL and OutputNL are nonlinearity estimator objects.

Note: If your input signal is binary, set InputNL to unitgain.

To use nonlinearity estimators with default settings, specify InputNL and OutputNL
using strings (such as 'wavenet' for wavelet network or 'sigmoidnet' for sigmoid
network).
load iddata3;
m = nlhw(z3,[2 2 1],'sigmoidnet','deadzone');

If you need to configure the properties of a nonlinearity estimator, use its object
representation. For example, to estimate a Hammerstein-Wiener model that uses

5-57
5 Nonlinear Black-Box Model Identification

saturation as its input nonlinearity and one-dimensional polynomial of degree 3 as its

output nonlinearity:

m = nlhw(z3,[2 2 1],'saturation',poly1d('Degree',3));

The third input 'saturation' is a string representation of the saturation nonlinearity

with default property values. poly1d('Degree',3) creates a one-dimensional
polynomial object of degree 3.

For MIMO models, specify the nonlinearities using objects unless you want to use the
same nonlinearity with default configuration for all channels.

This table summarizes values that specify the nonlinearity estimators.

Nonlinearity Value (Default Nonlinearity Configuration) Class

Piecewise linear 'pwlinear' or 'pwlin' pwlinear
(default)
One layer sigmoid 'sigmoidnet' or 'sigm' sigmoidnet
network
Wavelet network 'wavenet' or 'wave' wavenet
Saturation 'saturation' or 'sat' saturation
Dead zone 'deadzone' or 'dead' deadzone
One- 'poly1d' or 'poly' poly1d
dimensional polynomial
Unit gain 'unitgain' or [ ] unitgain

Additional available nonlinearities include custom networks that you create. Specify
a custom network by defining a function called gaussunit.m, as described in the
customnet reference page. Define the custom network object CNetw as:

CNetw = customnet(@gaussunit);
m = nlhw(z3,[2 2 1],'saturation',CNetw);

Exclude the input or output nonlinearity.

Exclude a nonlinearity for a specific channel by specifying the unitgain value for the
InputNonlinearity or OutputNonlinearity properties.

5-58
 Iteratively refine the model.

If the input signal is binary, set InputNL to unitgain.

For more information about model estimation and properties, see the nlhw and idnlhw
reference pages.

For a description of each nonlinearity estimator, see “Nonlinearity Estimators for

Hammerstein-Wiener Models” on page 5-50.

Iteratively refine the model.

Use pem to refine the original model. For example:

load iddata3;
m1 = nlhw(z3, [2 2 1],'sigmoidnet','wavenet');
% can repeatedly run this command
m2 = pem(z3,m1);

You can also use nlhw to refine the original model:

m1 = nlhw(z3, [2 2 1],'sigmoidnet','wavenet');
% can repeatedly run this command
m2 = nlhw(z3,m1);

Check the search termination criterion in m.EstimationInfo.WhyStop. If WhyStop

indicates that the estimation reached the maximum number of iterations, try repeating
the estimation and possibly specifying a larger value for the MaxIter idnlhw property.

% runs 30 more iterations starting at m1

m2 = pem(z3,m1,'MaxIter',30);

When the m.EstimationInfo.WhyStop value is Near (local) minimum,

(norm(g) < tol or No improvement along the search direction with
line search, validate your model to see if this model adequately fits the data. If
not, the solution might be stuck in a local minimum of the cost-function surface. Try
adjusting the Algorithm.Tolerance property value of the idnlhw class or the
Algorithm.SearchMethod property, and repeat the estimation.

You can also try perturbing the parameters of the last model using init (called
randomization) and refining the model using pem:

% original model

5-59
5 Nonlinear Black-Box Model Identification

M1 = nlhw(z3, [2 2 1], 'sigmoidnet','wavenet');

% randomly perturbs parameters about nominal values
M1p = init(M1);
% estimates parameters of perturbed model
M2 = pem(z3, M1p);

You can display the progress of the iterative search in the MATLAB Command Window
using the Display property of the idnlhw class.

M2= pem(z3,M1p,'Display','On');

Improve estimation results using initial states.

If your estimated Hammerstein-Wiener model provides a poor fit to measured data,
you can repeat the estimation using the initial state values estimated from the data. By
default, the initial states corresponding to the linear block of the Hammerstein-Wiener
model are zero.

To specify estimating initial states during model estimation:

load iddata3;
m = nlhw(z3,[2 2 1],sigmoidnet,[],'InitialState','e');

What if you cannot get a satisfactory model?

If you do not get a satisfactory model after many trials with various model structures
and algorithm settings, it is possible that the data is poor. For example, your data might
be missing important input or output variables and does not sufficiently cover all the
operating points of the system.

Nonlinear black-box system identification usually requires more data than linear model
identification to gain enough information about the system.

Use nlhw to Estimate Hammerstein-Wiener Models

This example shows how to use nlhw to estimate a Hammerstein-Wiener model for
measured input/output data.

Prepare the data for estimation.

5-60
 What if you cannot get a satisfactory model?

load twotankdata
z = iddata(y, u, 0.2);
ze = z(1:1000); zv = z(1001:3000);

Estimate several models using different model orders, delays, and nonlinearity settings.

m1 = nlhw(ze,[2 3 1]);
m2 = nlhw(ze,[2 2 3]);
m3 = nlhw(ze,[2 2 3], pwlinear('num',13),...
pwlinear('num',10));
m4 = nlhw(ze,[2 2 3], sigmoidnet('num',2),...
pwlinear('num',10));

An alternative way to perform the estimation is to configure the model structure first,
and then to estimate this model.

m5 = idnlhw([2 2 3], 'dead','sat');

m5 = pem(ze,m5);

Compare the resulting models by plotting the model outputs on top of the measured
output.

compare(zv,m1,m2,m3,m4,m5)

5-61
5 Nonlinear Black-Box Model Identification

Improve a Linear Model Using Hammerstein-Wiener Structure

This example shows how to use the Hammerstein-Wiener model structure to improve a
previously estimated linear model.

After estimating the linear model, insert it into the Hammerstein-Wiener structure that
includes input or output nonlinearities.

Estimate a linear model.

load iddata1
LM = arx(z1,[2 2 1]);

Extract the transfer function coefficients from the linear model.

5-62
 What if you cannot get a satisfactory model?

[Num, Den] = tfdata(LM);

Create a Hammerstein-Wiener model, where you initialize the linear block properties B
and F using Num and Den, respectively.

nb = 1; % In general, nb = ones(ny,nu)
% ny is number of outputs nu is number of inputs
nf = nb;
nk = 0; % In general, nk = zeros(ny,nu)
% ny is number of outputs nu is number of inputs
M = idnlhw([nb nf nk],[],'pwlinear');
M.b = Num;
M.f = Den;

Estimate the model coefficients, which refines the linear model coefficients in Num and
Den .

M = pem(z1, M);

Compare responses of linear and nonlinear model against measured data.

compare(z1,LM,M);

5-63
5 Nonlinear Black-Box Model Identification

Using Linear Model for Hammerstein-Wiener Estimation

• “About Using Linear Models” on page 5-64
• “How to Initialize Hammerstein-Wiener Estimation Using Linear Polynomial Output-
Error or State-Space Models” on page 5-65
• “Estimate Hammerstein-Wiener Models Using Linear OE Models” on page 5-66

About Using Linear Models

You can use a polynomial model of Output-Error (OE) structure (idpoly) or state-space
model with no disturbance component (idss model with K = 0) for Hammerstein-Wiener
estimation. The linear model must sufficiently represent the linear dynamics of your
system.

5-64
 What if you cannot get a satisfactory model?

Tip To learn more about when to use linear models, see “When to Fit Nonlinear Models”
on page 5-2.

Typically, you use the oe or n4sid command to obtain the linear model. You can provide
the linear model only at the command line when constructing (see idnlhw) or estimating
(see nlhw) a Hammerstein-Wiener model.

The software uses the linear model for initializing the Hammerstein-Wiener estimation:

• Assigns the linear model orders as initial values of nonlinear model orders (nb and nf
properties of the Hammerstein-Wiener (idnlhw) and delays (nk property).
• Sets the B and F polynomials of the linear transfer function in the Hammerstein-
Wiener model structure.

During estimation, the estimation algorithm uses these values to further adjust the
nonlinear model to the data.

How to Initialize Hammerstein-Wiener Estimation Using Linear Polynomial Output-Error or

State-Space Models

Estimate a Hammerstein-Wiener model using either a linear input-output polynomial

model of OE structure or state-space model by typing

m = nlhw(data,LinModel)

LinModel must be an idpoly model of OE structure, a state-space model (idss with K =

0) or transfer function idtf model. m is an idnlhw object. data is a time-domain iddata
object.

By default, both the input and output nonlinearity estimators are piecewise linear
functions (see pwlinear).

Specify different input and output nonlinearity, for example sigmoid and deadzone:

m = nlarx(data,LinModel, 'sigmoid','deadzone')

After each estimation, validate the model by comparing the simulated response to the
data. To improve the fit of the Hammerstein-Wiener model, adjust various elements
of the Hammerstein-Wiener structure. For more information, see “Ways to Configure
Hammerstein-Wiener Estimation” on page 5-52.

5-65
5 Nonlinear Black-Box Model Identification

Estimate Hammerstein-Wiener Models Using Linear OE Models

This example shows how to estimate Hammerstein-Wiener models using linear OE

models.

Load the estimation data.

load throttledata.mat

This command loads the data object ThrottleData into the workspace. The object
contains input and output samples collected from an engine throttle system, sampled at a
rate of 100Hz.

A DC motor controls the opening angle of the butterfly valve in the throttle system. A
step signal (in volts) drives the DC motor. The output is the angular position (in degrees)
of the valve.

Plot the data to view and analyze the data characteristics.

plot(ThrottleData)

5-66
 What if you cannot get a satisfactory model?

In the normal operating range of 15-90 degrees, the input and output variables have a
linear relationship. You use a linear model of low order to model this relationship.

In the throttle system, a hard stop limits the valve position to 90 degrees, and a
spring brings the valve to 15 degrees when the DC motor is turned off. These physical
components introduce nonlinearities that a linear model cannot capture.

Estimate a Hammerstein-Wiener model to model the linear behavior of this single-input

single-output system in the normal operating range.

% Detrend the data because linear models cannot capture offsets.

Tr = getTrend(ThrottleData);
Tr.OutputOffset = 15;
DetrendedData = detrend(ThrottleData,Tr);

5-67
5 Nonlinear Black-Box Model Identification

% Estimate a linear OE model with na=2, nb=1, nk=1.

opt = oeOptions('Focus','simulation');
LinearModel = oe(DetrendedData,[2 1 1],opt);

Compare the simulated model response with estimation data.

compare(DetrendedData, LinearModel)

The linear model captures the rising and settling behavior in the linear operating range
but does not account for output saturation at 90 degrees.

Estimate a Hammerstein-Wiener model to model the output saturation.

NonlinearModel = nlhw(ThrottleData, LinearModel, [], 'saturation');

5-68
 What if you cannot get a satisfactory model?

The software uses the orders and delay of the linear model for the orders of the nonlinear
model. In addition, the software uses the B and F polynomials of the linear transfer
function.

Compare the nonlinear model with data.

compare(ThrottleData, NonlinearModel)

Validating Hammerstein-Wiener Models

• “About Hammerstein-Wiener Plots” on page 5-70
• “How to Create Hammerstein-Wiener Plots in the App” on page 5-70

5-69
5 Nonlinear Black-Box Model Identification

• “How to Validate Hammerstein-Wiener Models at the Command Line” on page

5-71
• “Plotting Nonlinear Block Characteristics” on page 5-75
• “Plotting Linear Block Characteristics” on page 5-75

About Hammerstein-Wiener Plots

A Hammerstein-Wiener plot displays the characteristics of the linear block and the static
nonlinearities of a Hammerstein-Wiener model.

Examining a Hammerstein-Wiener plot can help you determine whether you chose an
unnecessarily complicated nonlinearity for modeling your system. For example, if you
chose a piecewise-linear nonlinearity (which is very general), but the plot indicates
saturation behavior, then you can estimate a new model using the simpler saturation
nonlinearity instead.

For multivariable systems, you can use the Hammerstein-Wiener plot to determine
whether to exclude nonlinearities for specific channels. If the nonlinearity for a specific
input or output channel does not exhibit strong nonlinear behavior, you can estimate a
new model after setting the nonlinearity at that channel to unit gain.

Explore the various plots in the plot window by clicking one of the three blocks that
represent the model:

• uNL — Input nonlinearity, representing the static nonlinearity at the input

(model.InputNonlinearity) to the Linear Block.
• Linear Block — Step, impulse, Bode and pole-zero plots of the embedded linear
model (model.LinearModel). By default, a step plot is displayed.
• yNL — Output nonlinearity, representing the static nonlinearity at the output
(model.OutputNonlinearity) of the Linear Block.

How to Create Hammerstein-Wiener Plots in the App

To create a Hammerstein-Wiener plot in the System Identification app, select the

Hamm-Wiener check box in the Model Views area. For general information about
creating and working with plots, see “Working with Plots” on page 14-11.

Note: The Hamm-Wiener check box is unavailable if you do not have a Hammerstein-
Wiener model in the Model Board.

5-70
 What if you cannot get a satisfactory model?

To include or exclude a model on the plot, click the corresponding model icon in the
System Identification app. By default, the input nonlinearity block UNL is selected. You
can select the output nonlinearity block YNL or Linear Block, as shown in the next
figure.

Selected block Supported plots

is highlighted. for linear block

Hide or show
the top pane.

After you generate a plot, you can learn more about your model by:

• “Plotting Nonlinear Block Characteristics” on page 5-75

• “Plotting Linear Block Characteristics” on page 5-75

How to Validate Hammerstein-Wiener Models at the Command Line

You can use the following approaches to validate Hammerstein-Wiener models at the
command line:

5-71
5 Nonlinear Black-Box Model Identification

Compare Model Output to Measured Output

Compare estimated models using compare. Use an independent validation data
set whenever possible. For more information about validating models, see “Model
Validation”.

For example, compare Output Error (OE) and Hammerstein-Wiener models of same
order:

load iddata1;
% Estimate linear ARX model
LM = oe(z1,[2 2 1]);
% Estimate Hammerstein-Wiener model
M = nlhw(z1,[2 2 1],'unitgain',[]);
% Compare responses of LM and M against measured data
compare(z1,LM,M);

5-72
 Compare Model Output to Measured Output

Compare the performance of several models using the properties

M.EstimationInfo.FPE (final prediction error) and M.EstimationInfo.LossFcn
(value of loss function at estimation termination). Smaller values typically indicate better
performance. However, m.EstimationInfo.FPE values might be unreliable when
the model contains a large number of parameters relative to the estimation data size.
Use these indicators in combination with other validation techniques to draw reliable
conclusions.

5-73
5 Nonlinear Black-Box Model Identification

Simulate and Predict Model Response

Use sim(idnlhw) and predict to simulate and predict model response, respectively. To
compute the step response of the model, use step. See the corresponding reference page
for more information.

Analyze Residuals
Residuals are differences between the model output and the measured output. Thus,
residuals represent the portion of the output not explained by the model. Use resid to
compute and plot the residuals.

Plot Nonlinearity
Access the object representing the nonlinearity estimator and its parameters using
M.InputNonlinearity (or M.unl) and M.OutputNonlinearity (or M.ynl), where M
is the estimated model.

Use plot to view the shape of the nonlinearity and properties of the linear block. For
example:

plot(M)

You can use additional plot arguments to specify the following information:

• Include several Hammerstein-Wiener models on the plot.

• Configure how to evaluate the nonlinearity at each input and output channel.
• Specify the time or frequency values for computing transient and frequency response
plots of the linear block.

Check Iterative Search Termination Conditions

Use M.EstimationInfo to display the estimation termination conditions, where M is the
estimated idnlhw model. For example, check the WhyStop field of the EstimationInfo

5-74
 Check Iterative Search Termination Conditions

property, which describes why the estimation was stopped. For example, the algorithm
might have reached the maximum number of iterations or the required tolerance value.

Plotting Nonlinear Block Characteristics

The Hammerstein-Wiener model can contain up to two nonlinear blocks. The

nonlinearity at the input to the Linear Block is labeled uNL and is called the input
nonlinearity. The nonlinearity at the output of the Linear Block is labeled yNL and is
called the output nonlinearity.

To configure the plot, perform the following steps:

1
If the top pane is not visible, click to expand the Hammerstein-Wiener Model
Plot window.
2 Select the nonlinear block you want to plot:

• To plot uNL as a command of the input data, click the uNL block.
• To plot yNL as a command of its inputs, click the yNL block.

The selected block is highlighted in green.

Note: The input to the output nonlinearity block yNL is the output from the Linear
Block and not the measured input data.
3 If your model contains multiple inputs or outputs, select the channel in the Select
nonlinearity at channel list. Selecting the channel updates the plot and displays
the nonlinearity values versus the corresponding input to this nonlinear block.
4 Click Apply to update the plot.

1 To change the range of the horizontal axis, select Options > Set input range to
open the Range for Input to Nonlinearity dialog box. Enter the range using the
format [MinValue MaxValue]. Click Apply and then Close to update the plot.

Plotting Linear Block Characteristics

The Hammerstein-Wiener model contains one Linear Block that represents the
embedded linear model.

To configure the plot:

5-75
5 Nonlinear Black-Box Model Identification

1
If the top pane is not visible, click to expand the Hammerstein-Wiener Model
Plot window.
2 Click the Linear Block to select it. The Linear Block is highlighted in green.
3 In the Select I/O pair list, select the input and output data pair for which to view
the response.
4 In the Choose plot type list, select the linear plot from the following options:

• Step
• Impulse
• Bode
• Pole-Zero Map

1 If you selected to plot step or impulse response, you can set the time span. Select
Options > Time span and enter a new time span in units of time you specified for
the model.

For a time span T, the resulting response is plotted from -T/4 to T. The default time
span is 10.

Click Apply and then Close.

2 If you selected to plot a Bode plot, you can set the frequency range.

The default frequency vector is 128 linearly distributed values, greater than zero and
less than or equal to the Nyquist frequency. To change the range, select Options
> Frequency range, and specify a new frequency vector in units of rad per model
time units.

Enter the frequency vector using any one of following methods:

• MATLAB expression, such as [1:100]*pi/100 or logspace(-3,-1,200).

Cannot contain variables in the MATLAB workspace.
• Row vector of values, such as [1:.1:100].

Click Apply and then Close.

5-76
 Simulation and Prediction

Using Hammerstein-Wiener Models

Simulation and Prediction

Use sim(idnlhw) to simulate the model output, and predict to predict the model
output. To compare models to measured output and to each other, use compare.

This toolbox provides a number of options to facilitate how you specify initial states. For
example, you can use findstates(idnlhw) to automatically search for state values in
simulation and prediction applications. You can also specify the states manually.

If you need to specify the states manually, see the idnlhw reference page for a definition
of the Hammerstein-Wiener model states.

To learn more about how sim and predict compute the model output, see “How the
Software Computes Hammerstein-Wiener Model Output” on page 5-78.

Linearization
Compute linear approximation of Hammerstein-Wiener models using linearize or
linapp.

linearize provides a first-order Taylor series approximation of the system about

an operation point (also called tangent linearization). linapp computes a linear
approximation of a nonlinear model for a given input data. For more information, see the
“Linear Approximation of Nonlinear Black-Box Models” on page 5-81.

You can compute the operating point for linearization using findop(idnlhw).

After computing a linear approximation of a nonlinear model, you can perform linear
analysis and control design on your model using Control System Toolbox commands. For
more information, see “Using Identified Models for Control Design Applications” and
“Create and Plot Identified Models Using Control System Toolbox Software”.

Simulation and Code Generation Using Simulink

You can import the estimated Hammerstein-Wiener Model into the Simulink software
using the Hammerstein-Wiener block (IDNLHW Model) from the System Identification

5-77
5 Nonlinear Black-Box Model Identification

Toolbox block library. After you bring the idnlhw object from the workspace into
Simulink, you can simulate the model output.

The IDNLHW Model block supports code generation with the Simulink Coder software,
using both generic and embedded targets. Code generation does not work when the model
contains customnet as the input or output nonlinearity.

How the Software Computes Hammerstein-Wiener Model Output

In most applications, sim(idnlhw) and predict are sufficient for computing the
simulated and predicted model response, respectively. This advanced topic describes
how the software evaluates the output of nonlinearity estimators and uses this output to
compute the model response.

Evaluating Nonlinearities (SISO)

Evaluating the predicted output of a nonlinearity for a input u requires that you first
extract the input or output nonlinearity F from the model:
F = M.InputNonlinearity % equivalent to F = M.unl
H = M.OutputNonlinearity % equivalent to F = M.ynl

Evaluate F(u):
w = evaluate(F,u)

where u is a scalar representing the value of the input signal at a given time.

You can evaluate predicted output values at multiple time instants by evaluating F for
several time values simultaneously using a column vector of input values:
w = evaluate(F,[u1;u2;u3])

Similarly, you can evaluate the value of the nonlinearity H using the output of the linear
block x(t) as its input:
y = evaluate(H,x)

Evaluating Nonlinearities (MIMO)

For MIMO models, F and H are vectors of length nu and ny, respectively. nu is the
number of inputs and ny is the number of outputs. In this case, you must evaluate the
predicted output of each nonlinearity separately.

5-78
 Simulation and Code Generation Using Simulink

For example, suppose that you estimate a two-input model:

M = nlhw(data,[nb nf nk],[wavenet;poly1d],'saturation')

In the input nonlinearity:

F = M.InputNonlinearity
F1 = F(1);
F2 = F(2);

F is a vector function containing two elements: F=[F1(u1_value); F2(u2_value)],

where F1 is a wavenet object and F2 is a poly1d object. u1_value is the first input
signal and u2_value is the second input signal.

Evaluate F by evaluating F1 and F2 separately:

w1 = evaluate(F(1), u1_value);
w2 = evaluate(F(2), u2_value);

The total input to the linear block, w, is a vector of w1 and w2 (w = [w1 w2]).

Similarly, you can evaluate the value of the nonlinearity H:

H = M.OutputNonlinearity %equivalent to H = M.ynl

Low-level Simulation of Hammerstein-Wiener Model

This example shows how the software evaluates the simulated output by first computing
the output of the input and output nonlinearity estimators. For same initial conditions,
the prediction results match the simulation results.

Estimate a Hammerstein-Wiener model.

load twotankdata
estData = iddata(y,u,0.2);
M = nlhw(estData,[1 5 3],'pwlinear','poly1d');

Extract the input nonlinearity, linear model, and output nonlinearity as separate
variables.
uNL = M.InputNonlinearity;
linModel = M.LinearModel;
yNL = M.OutputNonlinearity;

Simulate the output of the input nonlinearity estimator.

5-79
5 Nonlinear Black-Box Model Identification

% Input data for simulation

u = estData.u;
% Compute output of input nonlinearity
w = evaluate(uNL, u);
% Response of linear model to input w and zero initial conditions.
x = sim(linModel, w);
% Compute the output of the Hammerstein-Wiener model M
% as the output of the output nonlinearity estimator to input x.
y = evaluate(yNL, x);
% Previous commands are equivalent to.
ysim = sim(M, u);
% Compare low-level and direct simulation results.
time = estData.SamplingInstants;
plot(time, y, time, ysim, '.')

5-80
 Linear Approximation of Nonlinear Black-Box Models

Linear Approximation of Nonlinear Black-Box Models

In this section...
“Why Compute a Linear Approximation of a Nonlinear Model?” on page 5-81
“Choosing Your Linear Approximation Approach” on page 5-81
“Linear Approximation of Nonlinear Black-Box Models for a Given Input” on page
5-81
“Tangent Linearization of Nonlinear Black-Box Models” on page 5-82
“Computing Operating Points for Nonlinear Black-Box Models” on page 5-83

Why Compute a Linear Approximation of a Nonlinear Model?

Control design and linear analysis techniques using Control System Toolbox software
require linear models. You can use your estimated nonlinear model in these applications
after you linear the model. After you linearize your model, you can use the model for
control design and linear analysis.

Choosing Your Linear Approximation Approach

System Identification Toolbox software provides two approaches for computing a linear
approximation of nonlinear ARX and Hammerstein-Wiener models.

To compute a linear approximation of a nonlinear model for a given input signal, use the
linapp command. The resulting model is only valid for the same input that you use to
compute the linear approximation. For more information, see “Linear Approximation of
Nonlinear Black-Box Models for a Given Input” on page 5-81.

If you want a tangent approximation of the nonlinear dynamics that is accurate near the
system operating point, use the linearize command. The resulting model is a first-
order Taylor series approximation for the system about the operating point, which is
defined by a constant input and model state values. For more information, see “Tangent
Linearization of Nonlinear Black-Box Models” on page 5-82.

Linear Approximation of Nonlinear Black-Box Models for a Given Input

linapp computes the best linear approximation, in a mean-square-error sense, of a
nonlinear ARX or Hammerstein-Wiener model for a given input or a randomly generated

5-81
5 Nonlinear Black-Box Model Identification

input. The resulting linear model might only be valid for the same input signal as you the
one you used to generate the linear approximation.

linapp estimates the best linear model that is structurally similar to the original
nonlinear model and provides the best fit between a given input and the corresponding
simulated response of the nonlinear model.

To compute a linear approximation of a nonlinear black-box model for a given input, you
must have these variables:

• Nonlinear ARX model (idnlarx object) or Hammerstein-Wiener model (idnlhw

object)
• Input signal for which you want to obtain a linear approximation, specified as a real
matrix or an iddata object

linapp uses the specified input signal to compute a linear approximation:

• For nonlinear ARX models, linapp estimates a linear ARX model using the same
model orders na, nb, and nk as the original model.
• For Hammerstein-Wiener models, linapp estimates a linear Output-Error (OE)
model using the same model orders nb, nf, and nk.

To compute a linear approximation of a nonlinear black-box model for a randomly

generated input, you must specify the minimum and maximum input values for
generating white-noise input with a magnitude in this rectangular range, umin and
umax.

For more information, see the linapp reference page.

Tangent Linearization of Nonlinear Black-Box Models

linearize computes a first-order Taylor series approximation for nonlinear system
dynamics about an operating point, which is defined by a constant input and model state
values. The resulting linear model is accurate in the local neighborhood of this operating
point.

To compute a tangent linear approximation of a nonlinear black-box model, you must

have these variables:

• Nonlinear ARX model (idnlarx object) or Hammerstein-Wiener model (idnlhw

object)

5-82
 Linear Approximation of Nonlinear Black-Box Models

• Operating point

To specify the operating point of your system, you must specify the constant input and
the states. For more information about state definitions for each type of parametric
model, see these reference pages:

• idnlarx — Nonlinear ARX model

• idnlhw — Nonlinear Hammerstein-Wiener model

If you do not know the operating point values for your system, see “Computing Operating
Points for Nonlinear Black-Box Models” on page 5-83.

For more information, see the idnlarx/linearize or idnlhw/linearize reference

page.

Computing Operating Points for Nonlinear Black-Box Models

An operating point is defined by a constant input and model state values.

If you do not know the operating conditions of your system for linearization, you can use
findop to compute the operating point from specifications:

• “Computing Operating Point from Steady-State Specifications” on page 5-83

• “Computing Operating Points at a Simulation Snapshot” on page 5-84

Computing Operating Point from Steady-State Specifications

Use findop to compute an operating point from steady-state specifications:

• Values of input and output signals.

If either the steady-state input or output value is unknown, you can specify it as NaN
to estimate this value. This is especially useful when modeling MIMO systems, where
only a subset of the input and output steady-state values are known.
• More complex steady-state specifications.

Construct an object that stores specifications for computing the operating point,
including input and output bounds, known values, and initial guesses. For more
information, see operspec(idnlarx) or operspec(idnlhw).

For more information, see the findop(idnlarx) or findop(idnlhw) reference page.

5-83
5 Nonlinear Black-Box Model Identification

Computing Operating Points at a Simulation Snapshot

Compute an operating point at a specific time during model simulation (snapshot) by

specifying the snapshot time and the input value. To use this method for computing the
equilibrium operating point, choose an input that leads to a steady-state output value.
Use that input and the time value at which the output reaches steady state (snapshot
time) to compute the operating point.

It is optional to specify the initial conditions for simulation when using this method
because initial conditions often do not affect the steady-state values. By default, the
initial conditions are zero.

However, for nonlinear ARX models, the steady-state output value might depend
on initial conditions. For these models, you should investigate the effect of initial
conditions on model response and use the values that produce the desired output. You
can use data2state(idnlarx) to map the input-output signal values from before the
simulation starts to the model's initial states. Because the initial states are a function
of the past history of the model's input and output values, data2state generates the
initial states by transforming the data.

5-84
 6

ODE Parameter Estimation (Grey-Box

Modeling)

• “Supported Grey-Box Models” on page 6-2

• “Data Supported by Grey-Box Models” on page 6-3
• “Choosing idgrey or idnlgrey Model Object” on page 6-4
• “Estimating Linear Grey-Box Models” on page 6-6
• “Estimating Nonlinear Grey-Box Models” on page 6-15
• “After Estimating Grey-Box Models” on page 6-38
• “Estimating Coefficients of ODEs to Fit Given Solution” on page 6-39
• “Estimate Model Using Zero/Pole/Gain Parameters” on page 6-46
• “Creating IDNLGREY Model Files” on page 6-52
6 ODE Parameter Estimation (Grey-Box Modeling)

Supported Grey-Box Models

If you understand the physics of your system and can represent the system using
ordinary differential or difference equations (ODEs) with unknown parameters, then you
can use System Identification Toolbox commands to perform linear or nonlinear grey-
box modeling. Grey-box model ODEs specify the mathematical structure of the model
explicitly, including couplings between parameters. Grey-box modeling is useful when
you know the relationships between variables, constraints on model behavior, or explicit
equations representing system dynamics.

The toolbox supports both continuous-time and discrete-time linear and nonlinear
models. However, because most laws of physics are expressed in continuous time, it
is easier to construct models with physical insight in continuous time, rather than in
discrete time.

In addition to dynamic input-output models, you can also create time-series models that
have no inputs and static models that have no states.

If it is too difficult to describe your system using known physical laws, you can use the
black-box modeling approach. For more information, see “Linear Model Identification”
and “Nonlinear Model Identification”.

You can also use the idss model object to perform structured model estimation by using
its Structure property to fix or free specific parameters. However, you cannot use
this approach to estimate arbitrary structures (arbitrary parameterization). For more
information about structure matrices, see “Estimate State-Space Models with Structured
Parameterization”.

6-2
 Data Supported by Grey-Box Models

Data Supported by Grey-Box Models

You can estimate both continuous-time or discrete-time grey-box models for data with the
following characteristics:

• Time-domain or frequency-domain data, including time-series data with no inputs.

Note: Nonlinear grey-box models support only time-domain data.

• Single-output or multiple-output data

You must first import your data into the MATLAB workspace. You must represent your
data as an iddata or idfrd object. For more information about preparing data for
identification, see “Data Preparation”.

6-3
6 ODE Parameter Estimation (Grey-Box Modeling)

Choosing idgrey or idnlgrey Model Object

Grey-box models require that you specify the structure of the ODE model in a file.
You use this file to create the idgrey or idnlgrey model object. You can use both
the idgrey and the idnlgrey objects to model linear systems. However, you can only
represent nonlinear dynamics using the idnlgrey model object.

The idgrey object requires that you write a function to describe the linear dynamics in
the state-space form, such that this file returns the state-space matrices as a function
of your parameters. For more information, see “Specifying the Linear Grey-Box Model
Structure” on page 6-6.

The idnlgrey object requires that you write a function or MEX-file to describe the
dynamics as a set of first-order differential equations, such that this file returns the
output and state derivatives as a function of time, input, state, and parameter values.
For more information, see “Specifying the Nonlinear Grey-Box Model Structure” on page
6-15.

The following table compares idgrey and idnlgrey model objects.

Comparison of idgrey and idnlgrey Objects

Settings and Operations Supported by idgrey? Supported by idnlgrey?

Set bounds on parameter Yes Yes
values.
Handle initial states Yes Yes
individually.
Perform linear analysis. Yes No
For example, use the bode
command.
Honor stability constraints. Yes No
Specify constraints using the
Advanced.StabilityThreshold
estimation option. For Note: You can use parameter
more information, see bounds to ensure stability of
greyestOptions. an idnlgrey model, if these
bounds are known.
Estimate a disturbance model. Yes No

6-4
 Choosing idgrey or idnlgrey Model Object

Settings and Operations Supported by idgrey? Supported by idnlgrey?

The disturbance model is
represented by K in state-space
equations.
Optimize estimation results for Yes No
simulation or prediction. Set the Focus estimation Because idnlgrey models are
option to 'Simulation' Output-Error models, there is no
or 'Prediction'. For difference between simulation
more information, see and prediction results.
greyestOptions.

6-5
6 ODE Parameter Estimation (Grey-Box Modeling)

Estimating Linear Grey-Box Models

In this section...
“Specifying the Linear Grey-Box Model Structure” on page 6-6
“Create Function to Represent a Grey-Box Model” on page 6-7
“Estimate Continuous-Time Grey-Box Model for Heat Diffusion” on page 6-9
“Estimate Discrete-Time Grey-Box Model with Parameterized Disturbance” on page
6-12

Specifying the Linear Grey-Box Model Structure

You can estimate linear discrete-time and continuous-time grey-box models for arbitrary
ordinary differential or difference equations using single-output and multiple-output
time-domain data, or time-series data (output-only).

You must represent your system equations in state-space form. State-space models use
state variables x(t) to describe a system as a set of first-order differential equations,
rather than by one or more nth-order differential equations.

The first step in grey-box modeling is to write a function that returns state-space
matrices as a function of user-defined parameters and information about the model.

Use the following format to implement the linear grey-box model in the file:
[A,B,C,D] = myfunc(par1,par2,...,parN,Ts,aux1,aux2,...)
where the output arguments are the state-space matrices and myfunc is the name of
the file. par1,par2,...,parN are the N parameters of the model. Each entry may
be a scalar, vector or matrix.Ts is the sample time. aux1,aux2,... are the optional
input arguments that myfunc uses to compute the state-space matrices in addition to the
parameters and sample time.

aux contains auxiliary variables in your system. You use auxiliary variables to vary
system parameters at the input to the function, and avoid editing the file.

You can write the contents of myfunc to parameterize either a continuous-time, or a

discrete-time state-space model, or both. When you create the idgrey model using
myfunc, you can specify the nature of the output arguments of myfunc. The continuous-
time state-space model has the form:

In continuous-time, the state-space description has the following form:

6-6
 Estimating Linear Grey-Box Models

x& (t) = Ax( t) + Bu(t) + Ke(t)

y( t) = Cx( t) + Du(t) + e( t)
x(0) = x0

Where, A,B,C and D are matrices that are parameterized by the parameters
par1,par2,...,parN. The noise matrix K and initial state vector, x0, are not
parameterized by myfunc. They can still be estimated as additional quantities along with
the model parameters. To configure the handling of initial states, x0, and the disturbance
component, K, during estimation, use the greyestOptions option set.

In discrete-time, the state-space description has a similar form:

x(k + 1) = Ax( k) + Bu( k) + Ke(t)

y( k) = Cx(k) + Du(k) + e(t)
x(0) = x0

Where, A,B,C and D are now the discrete-time matrices that are parameterized by the
parameters par1,par2,...,parN. K and x0 are not directly parameterized, but can be
estimated if required by configuring the corresponding estimation options.

Parameterizing Disturbance Model and Initial States

In some applications, you may want to express K and x0 as quantities that are
parameterized by chosen parameters, just as the A, B, C and D matrices. To handle
such cases, you can write the ODE file, myfunc, to return K and x0 as additional output
arguments:
[A,B,C,D,K,x0] = myfunc(par1,par2,...,parN,Ts,aux1,aux2,...)

K and x0 are thus treated in the same way as the A, B, C and D matrices. They are all
functions of the parameters par1,par2,...,parN.

Constructing the idgrey Object

After creating the function or MEX-file with your model structure, you must define as
idgrey object. For information regarding creating this, see idgrey.

Create Function to Represent a Grey-Box Model

This example shows how to represent the structure of the following continuous-time
model:

6-7
6 ODE Parameter Estimation (Grey-Box Modeling)

È0 1 ˘ È0˘
x& (t) = Í ˙ x(t) + Í ˙ u(t)
Î 0 q1 ˚ Îq2 ˚
È1 0 ˘
y( t) = Í ˙ x( t) + e(t)
Î0 1 ˚
Èq ˘
x(0) = Í 3 ˙
Î0 ˚

This equation represents an electrical motor, where y1 ( t) = x1 (t) is the angular position
of the motor shaft, and y2 (t) = x2 (t) is the angular velocity. The parameter -q1 is the

inverse time constant of the motor, and - q 2 q is the static gain from the input to the
1

angular velocity.

The motor is at rest at t=0, but its angular position q3 is unknown. Suppose that the
approximate nominal values of the unknown parameters are q1 = -1 , q2 = 0 .25 and
q3 = 0 . For more information about this example, see the section on state-space models
in System Identification: Theory for the User, Second Edition, by Lennart Ljung, Prentice
Hall PTR, 1999.

The continuous-time state-space model structure is defined by the following equation:

% (t)
x& (t) = Fx(t) + Gu(t) + Kw
y( t) = Hx(t) + Du( t) + w(t)
x(0) = x0

To prepare this model for identification:

1 Create the following file to represent the model structure in this example:

function [A,B,C,D,K,x0] = myfunc(par,T)

A = [0 1; 0 par(1)];
B = [0;par(2)];
C = eye(2);
D = zeros(2,1);
K = zeros(2,2);
x0 =[par(3);0];

6-8
 Estimating Linear Grey-Box Models

2 Use the following syntax to define an idgrey model object based on the myfunc file:
par = [-1; 0.25; 0];
aux = {};
T = 0;
m = idgrey('myfunc',par,'c',aux,T);

where par represents a vector of all the user-defined parameters and contains
their nominal (initial) values. In this example, all the scalar-valued parameters
are grouped in the par vector. The scalar-valued parameters could also have been
treated as independent input arguments to the ODE function myfunc.'c' specifies
that the underlying parameterization is in continuous time. aux represents optional
arguments. As myfunc does not have any optional arguments, use aux = {}. T
specifies the sample interval; T = 0 indicates a continuous-time model.

Use greyest to estimate the grey-box parameter values:

m_est = greyest(data,m)

where data is the estimation data and m is an estimation initialization idgrey model.
m_est is the estimated idgrey model.

Note: Compare this example to “Estimate Structured Continuous-Time State-Space

Models” on page 4-34, where the same problem is solved using a structured state-space
representation.

Estimate Continuous-Time Grey-Box Model for Heat Diffusion

This example shows how to estimate the heat conductivity and the heat-transfer
coefficient of a continuous-time grey-box model for a heated-rod system.

This system consists of a well-insulated metal rod of length L and a heat-diffusion

coefficient k . The input to the system is the heating power u(t) and the measured output
y(t) is the temperature at the other end.

Under ideal conditions, this system is described by the heat-diffusion equation—which is

a partial differential equation in space and time.

∂ x( t, x ) ∂ 2 x( t, x )
=k
∂t ∂x 2

6-9
6 ODE Parameter Estimation (Grey-Box Modeling)

To get a continuous-time state-space model, you can represent the second-derivative

using the following difference approximation:

∂ 2 x( t, x ) x ( t, x + D L) - 2 x(t, x) + x ( t, x - DL )
=
2 2
∂x ( DL )
where x = k ◊ DL

This transformation produces a state-space model of order n = DLL , where the state
variables x(t, k ◊ DL) are lumped representations for x(t, x) for the following range of
values:

k ◊ DL £ x < ( k + 1) DL

The dimension of x depends on the spatial grid size DL in the approximation.

The heat-diffusion equation is mapped to the following continuous-time state-space

model structure to identify the state-space matrices:

% (t)
x& (t) = Fx(t) + Gu(t) + Kw
y( t) = Hx(t) + Du( t) + w(t)
x(0) = x0

The state-space matrices are parameterized by the heat diffusion coefficient κ and the
heat transfer coefficient at the far end of the rod htf. The expressions also depend upon
the grid size, Ngrid, and the length of the rod L. The initial conditions x0 are a function
of the initial room temperature, treated as a known quantity in this example.

1 Create a MATLAB file.

The following code describes the state-space equation for this model. The parameters
are κ and htf while the auxiliary variables are Ngrid, L and initial room temperature
temp. The grid size is supplied as an auxiliary variable so that the ODE function can
be easily adapted for various grid sizes.
function [A,B,C,D,K,x0] = heatd(kappa, htf, T, Ngrid, L, temp)
% ODE file parameterizing the heat diffusion model

6-10
 Estimating Linear Grey-Box Models

% kappa (first parameter) - heat diffusion coefficient

% htf (second parameter) - heat transfer coefficient
% at the far end of rod

% Auxiliary variables for computing state-space matrices:

% Ngrid: Number of points in the space-discretization
% L: Length of the rod
% temp: Initial room temperature (uniform)

% Compute space interval

deltaL = L/Ngrid;

% A matrix
A = zeros(Ngrid,Ngrid);
for kk = 2:Ngrid-1
A(kk,kk-1) = 1;
A(kk,kk) = -2;
A(kk,kk+1) = 1;
end

% Boundary condition on insulated end

A(1,1) = -1; A(1,2) = 1;
A(Ngrid,Ngrid-1) = 1;
A(Ngrid,Ngrid) = -1;
A = A*kappa/deltaL/deltaL;

% B matrix
B = zeros(Ngrid,1);
B(Ngrid,1) = htf/deltaL;

% C matrix
C = zeros(1,Ngrid);
C(1,1) = 1;

% D matrix (fixed to zero)

D = 0;

% K matrix: fixed to zero

K = zeros(Ngrid,1);

% Initial states: fixed to room temperature

x0 = temp*ones(Ngrid,1);
2 Use the following syntax to define an idgrey model object based on the heatd code
file:

6-11
6 ODE Parameter Estimation (Grey-Box Modeling)

m = idgrey('heatd', {0.27 1}, 'c', {10, 1, 22});

This command specifies the auxiliary parameters as inputs to the function, include
the model order (grid size) 10, the rod length of 1 meter, and an initial temperature
of 22 degrees Celsius. The command also specifies the initial values for heat
conductivity as 0.27, and for the heat transfer coefficient as 1.
3 For given data, you can use greyest to estimate the grey-box parameter values:
me = greyest(data,m)

The following command shows how you can specify to estimate a new model with
different auxiliary variables:
m.Structure.ExtraArgs = {20,1,22};
me = greyest(data,m);

This syntax uses the ExtraArgs model structure attribute to specify a finer grid using a
larger value for Ngrid. For more information about linear grey-box model properties, see
the idgrey reference page.

Estimate Discrete-Time Grey-Box Model with Parameterized Disturbance

This example shows how to create a single-input and single-output grey-box model
structure when you know the variance of the measurement noise. The code in this
example uses the Control System Toolbox command kalman for computing the Kalman
gain from the known and estimated noise variance.

Description of the SISO System

This example is based on a discrete, single-input and single-output (SISO) system

represented by the following state-space equations:

È par1 par 2 ˘ È1 ˘
x(kT + T ) = Í ˙ x( kT ) + Í ˙ u(kT ) + w(kT )
Î 1 0 ˚ Î0 ˚
y( kT ) = [ par3 par 4 ] x( kT) + e(kT )
x(0) = x0

where w and e are independent white-noise terms with covariance matrices R1 and R2,
respectively. R1=E{ww'} is a 2–by-2 matrix and R2=E{ee'} is a scalar. par1, par2, par3,
and par4 represent the unknown parameter values to be estimated.

6-12
 Estimating Linear Grey-Box Models

Assume that you know the variance of the measurement noise R2 to be 1. R1(1,1) is
unknown and is treated as an additional parameter par5. The remaining elements of R1
are known to be zero.

Estimating the Parameters of an idgrey Model

You can represent the system described in “Description of the SISO System” on page
6-12 as an idgrey (grey-box) model using a function. Then, you can use this file and
the greyest command to estimate the model parameters based on initial parameter
guesses.

To run this example, you must load an input-output data set and represent it as an
iddata or idfrd object called data. For more information about this operation, see
“Representing Time- and Frequency-Domain Data Using iddata Objects” on page 2-50
or “Representing Frequency-Response Data Using idfrd Objects” on page 2-83.

To estimate the parameters of a grey-box model:

1 Create the file mynoise that computes the state-space matrices as a function of
the five unknown parameters and the auxiliary variable that represents the known
variance R2. The initial conditions are not parameterized; they are assumed to be
zero during this estimation.

Note: R2 is treated as an auxiliary variable rather than assigned a value in the file
to let you change this value directly at the command line and avoid editing the file.

function [A,B,C,D,K] = mynoise(par,T,aux)

R2 = aux(1); % Known measurement noise variance
A = [par(1) par(2);1 0];
B = [1;0];
C = [par(3) par(4)];
D = 0;
R1 = [par(5) 0;0 0];
[~,K] = kalman(ss(A,eye(2),C,0,T),R1,R2);
% Uses Control System Toolbox product
% u=[]
2 Specify initial guesses for the unknown parameter values and the auxiliary
parameter value R2:
par1 = 0.1; % Initial guess for A(1,1)
par2 = -2; % Initial guess for A(1,2)

6-13
6 ODE Parameter Estimation (Grey-Box Modeling)

par3 = 1; % Initial guess for C(1,1)

par4 = 3; % Initial guess for C(1,2)
par5 = 0.2; % Initial guess for R1(1,1)
Pvec = [par1; par2; par3; par4; par5]
auxVal = 1; % R2=1
3 Construct an idgrey model using the mynoise file:

Minit = idgrey('mynoise',Pvec,'d',auxVal);

The third input argument 'd' specifies a discrete-time system.

4 Estimate the model parameter values from data:
opt = greyestOptions;
opt.InitialState = 'zero';
opt.Display = 'full';
Model = greyest(data,Minit,opt)

6-14
 Estimating Nonlinear Grey-Box Models

Estimating Nonlinear Grey-Box Models

In this section...
“Specifying the Nonlinear Grey-Box Model Structure” on page 6-15
“Constructing the idnlgrey Object” on page 6-17
“Using pem to Estimate Nonlinear Grey-Box Models” on page 6-17
“Nonlinear Grey-Box Model Estimation Algorithm Options” on page 6-18
“Represent Nonlinear Dynamics Using MATLAB File for Grey-Box Estimation” on page
6-20

Specifying the Nonlinear Grey-Box Model Structure

You must represent your system as a set of first-order nonlinear difference or differential
equations:

x† (t) = F(t, x(t), u(t), par1, par 2,..., parN )

y( t) = H ( t, x(t), u(t), par1, par2,..., parN ) + e( t)
x(0) = x0

where x† (t) = dx( t) dt for continuous-time representation and x† (t) = x( t + Ts) for discrete-
time representation with Ts as the sampling interval. F and H are arbitrary linear or
nonlinear functions with Nx and Ny components, respectively. Nx is the number of states
and Ny is the number of outputs.

After you establish the equations for your system, create a function or MEX-file. MEX-
files, which can be created in C or Fortran, are dynamically linked subroutines that can
be loaded and executed by the MATLAB interpreter. For more information about MEX-
files, see “MEX-File Creation API”.

The purpose of the model file is to return the state derivatives and model outputs as a
function of time, states, inputs, and model parameters, as follows:
[dx,y] = MODFILENAME(t,x,u,p1,p2, ...,pN,FileArgument)

Tip The template file for writing the C MEX-file, IDNLGREY_MODEL_TEMPLATE.c, is

located in matlab/toolbox/ident/nlident.

6-15
6 ODE Parameter Estimation (Grey-Box Modeling)

The output variables are:

• dx — Represents the right side(s) of the state-space equation(s). A column vector with
Nx entries. For static models, dx=[].

For discrete-time models. dx is the value of the states at the next time step x(t
+Ts).

For continuous-time models. dx is the state derivatives at time t, or dx

dt
.

• y — Represents the right side(s) of the output equation(s). A column vector with Ny
entries.

The file inputs are:

• t — Current time.
• x — State vector at time t. For static models, equals [].
• u — Input vector at time t. For time-series models, equals [].
• p1,p2, ...,pN — Parameters, which can be real scalars, column vectors or two-
dimensional matrices. N is the number of parameter objects. For scalar parameters, N
is the total number of parameter elements.
• FileArgument — Contains auxiliary variables that might be required for updating
the constants in the state equations.

Tip After creating a model file, call it directly from the MATLAB software with
reasonable inputs and verify the output values.

For an example of creating grey-box model files and idnlgrey model object, see Creating
idnlgrey Model Files.

For examples of code files and MEX-files that specify model structure, see the toolbox/
ident/iddemos/examples folder. For example, the model of a DC motor is described in
files dcmotor_m and dcmotor_c.

6-16
 Estimating Nonlinear Grey-Box Models

Constructing the idnlgrey Object

After you create the function or MEX-file with your model structure, you must define an
idnlgrey object. This object shares many of the properties of the linear idgrey model
object.

Use the following syntax to define the idnlgrey model object:

m = idnlgrey('filename',Order,Parameters,InitialStates)

The idnlgrey arguments are defined as follows:

• 'filename' — Name of the function or MEX-file storing the model structure.

This file must be on the MATLAB path when you use this model object for model
estimation, prediction, or simulation.
• Order — Vector with three entries [Ny Nu Nx], specifying the number of model
outputs Ny, the number of inputs Nu, and the number of states Nx.
• Parameters — Parameters, specified as struct arrays, cell arrays, or double arrays.
• InitialStates — Specified in the same way as parameters. Must be the fourth
input to the idnlgrey constructor.

For detailed information about this object and its properties, see the idnlgrey reference
page.

Use pem to estimate your grey-box model.

Using pem to Estimate Nonlinear Grey-Box Models

You can use the pem command to estimate the unknown idnlgrey model parameters
and initial states using measured data.

The input-output dimensions of the data must be compatible with the input and output
orders you specified for the idnlgrey model.

Use the following general estimation syntax:

m = pem(data,m)

where data is the estimation data and m is the idnlgrey model object you constructed.

6-17
6 ODE Parameter Estimation (Grey-Box Modeling)

You can pass additional property-value pairs to pem to specify the properties of the
model or the estimation algorithm. Assignable properties include the ones returned
by the get(idnlgrey) command and the algorithm properties returned by the
get(idnlgrey, 'Algorithm'), such as MaxIter and Tolerance. For detailed
information about these model properties, see the idnlgrey reference page.

For more information about validating your models, see “Model Validation”.

Nonlinear Grey-Box Model Estimation Algorithm Options

The Algorithm property of the model specifies the estimation algorithm, which
simulates the model several times by trying various parameter values to reduce the
prediction error.

The following algorithm properties can affect the quality of the results:

• “Simulation Method” on page 6-18

• “Search Method” on page 6-19
• “Gradient Options” on page 6-19
• “Example – Specifying Algorithm Properties” on page 6-19

For detailed information about these and other model properties, see the idnlgrey
reference page.

Simulation Method

You can specify the simulation method using the SimulationOptions (struct) fields
of the model Algorithm property.

System Identification Toolbox software provides several variable-step and fixed-step

solvers for simulating idnlgrey models. To view a list of available solvers and their
properties, type the following command at the prompt:
idprops idnlgrey algorithm.simulationoptions

For discrete-time systems, the default solver is 'FixedStepDiscrete'. For continuous-

time systems, the default solver is 'ode45'.

By default, SimulationOptions.Solver is set to 'Auto', which automatically

selects either 'ode45' or 'FixedStepDiscrete' during estimation and simulation—
depending on whether the system is continuous or discrete in time.

6-18
 Estimating Nonlinear Grey-Box Models

Search Method

You can specify the search method for estimating model parameters using the
SearchMethod field of the Algorithm property. Two categories of methods are available
for nonlinear grey-box modeling.

One category of methods consists of the minimization schemes that are based on line-
search methods, including Gauss-Newton type methods, steepest-descent methods, and
Levenberg-Marquardt methods.

The Trust-Region Reflective Newton method of nonlinear least-squares (lsqnonlin),

where the cost is the sum of squares of errors between the measured and simulated
outputs, requires Optimization Toolbox™ software. When the parameter bounds differ
from the default +/- Inf, this search method handles the bounds better than the schemes
based on a line search. However, unlike the line-search-based methods, lsqnonlin only
works with Criterion='Trace'.

By default, SearchMethod is set to Auto, which automatically selects a method from the
available minimizers. If the Optimization Toolbox product is installed, SearchMethod
is set to 'lsqnonlin'. Otherwise, SearchMethod is a combination of line-search based
schemes.

Gradient Options

You can specify the method for calculating gradients using the GradientOptions
field of the Algorithm property. Gradients are the derivatives of errors with respect to
unknown parameters and initial states.

Gradients are calculated by numerically perturbing unknown quantities and measuring

their effects on the simulation error.

Option for gradient computation include the choice of the differencing scheme (forward,
backward or central), the size of minimum perturbation of the unknown quantities, and
whether the gradients are calculated simultaneously or individually.

Example – Specifying Algorithm Properties

You can specify the Algorithm fields directly in the estimation syntax, as property-value
pairs.

For example, you can specify the following properties as part of the pem syntax:
m = pem(data,init_model,'Search','gn',...

6-19
6 ODE Parameter Estimation (Grey-Box Modeling)

'MaxIter',5,...
'Display','On')

Represent Nonlinear Dynamics Using MATLAB File for Grey-Box

Estimation
This example shows how to construct, estimate and analyze nonlinear grey-box models.

Nonlinear grey-box (idnlgrey) models are suitable for estimating parameters of systems
that are described by nonlinear state-space structures in continuous or discrete time.
You can use both idgrey (linear grey-box model) and idnlgrey objects to model linear
systems. However, you can only use idnlgrey to represent nonlinear dynamics. To
learn about linear grey-box modeling using idgrey, see “"Building Structured and User-
Defined Models Using System Identification Toolbox™"”.

About the Model

In this example, you model the dynamics of a linear DC motor using the idnlgrey
object.

Figure 1: Schematic diagram of a DC-motor.

6-20
 Estimating Nonlinear Grey-Box Models

If you ignore the disturbances and choose y(1) as the angular position [rad] and y(2)
as the angular velocity [rad/s] of the motor, you can set up a linear state-space structure
of the following form (see Ljung, L. System Identification: Theory for the User, Upper
Saddle River, NJ, Prentice-Hall PTR, 1999, 2nd ed., p. 95-97 for the derivation):

d | 0 1 | | 0 |
-- x(t) = | | x(t) + | | u(t)
dt | 0 -1/tau | | k/tau |

| 1 0 |
y(t) = | | x(t)
| 0 1 |

tau is the time-constant of the motor in [s] and k is the static gain from the input to the
angular velocity in [rad/(V*s)] . See Ljung (1999) for how tau and k relate to the physical
parameters of the motor.

About the Input-Output Data

1. Load the DC motor data.

load(fullfile(matlabroot, 'toolbox', 'ident', 'iddemos', 'data', 'dcmotordata'));

2. Represent the estimation data as an iddata object.

z = iddata(y, u, 0.1, 'Name', 'DC-motor');

3. Specify input and output signal names, start time and time units.

z.InputName = 'Voltage';
z.InputUnit = 'V';
z.OutputName = {'Angular position', 'Angular velocity'};
z.OutputUnit = {'rad', 'rad/s'};
z.Tstart = 0;
z.TimeUnit = 's';

4. Plot the data.

The data is shown in two plot windows.

figure('Name', [z.Name ': Voltage input -> Angular position output']);

plot(z(:, 1, 1)); % Plot first input-output pair (Voltage -> Angular position).
figure('Name', [z.Name ': Voltage input -> Angular velocity output']);
plot(z(:, 2, 1)); % Plot second input-output pair (Voltage -> Angular velocity).

6-21
6 ODE Parameter Estimation (Grey-Box Modeling)

6-22
 Estimating Nonlinear Grey-Box Models

Figure 2: Input-output data from a DC-motor.

Linear Modeling of the DC-Motor

1. Represent the DC motor structure in a function.

In this example, you use a MATLAB® file, but you can also use C MEX-files (to gain
computational speed), P-files or function handles. For more information, see “"Creating
IDNLGREY Model Files"”.

The DC-motor function is called dcmotor_m.m and is shown below.

function [dx, y] = dcmotor_m(t, x, u, tau, k, varargin)

% Output equations.

6-23
6 ODE Parameter Estimation (Grey-Box Modeling)

y = [x(1); ... % Angular position.

x(2) ... % Angular velocity.
];

% State equations.
dx = [x(2); ... % Angular velocity.
-(1/tau)*x(2)+(k/tau)*u(1) ... % Angular acceleration.
];

The file must always be structured to return the following:

Output arguments:

• dx is the vector of state derivatives in continuous-time case, and state update values
in the discrete-time case.

• y is the output equation

Input arguments:

• The first three input arguments must be: t (time), x (state vector, [] for static
systems), u (input vector, [] for time-series).

• Ordered list of parameters follow. The parameters can be scalars, column vectors, or
2-dimensional matrices.

• varargin for the auxiliary input arguments

2. Represent the DC motor dynamics using an idnlgrey object.

The model describes how the inputs generate the outputs using the state equation(s).

FileName = 'dcmotor_m'; % File describing the model structure.

Order = [2 1 2]; % Model orders [ny nu nx].
Parameters = [1; 0.28]; % Initial parameters. Np = 2.
InitialStates = [0; 0]; % Initial initial states.
Ts = 0; % Time-continuous system.
nlgr = idnlgrey(FileName, Order, Parameters, InitialStates, Ts, ...
'Name', 'DC-motor');

In practice, there are disturbances that affect the outputs. An idnlgrey model does not
explicitly model the disturbances, but assumes that these are just added to the output(s).
Thus, idnlgrey models are equivalent to Output-Error (OE) models. Without a noise

6-24
 Estimating Nonlinear Grey-Box Models

model, past outputs do not influence prediction of future outputs, which means that
predicted output for any prediction horizon k coincide with simulated outputs.

3. Specify input and output names, and units.

set(nlgr, 'InputName', 'Voltage', 'InputUnit', 'V', ...

'OutputName', {'Angular position', 'Angular velocity'}, ...
'OutputUnit', {'rad', 'rad/s'}, ...
'TimeUnit', 's');

4. Specify names and units of the initial states and parameters.

nlgr = setinit(nlgr, 'Name', {'Angular position' 'Angular velocity'});

nlgr = setinit(nlgr, 'Unit', {'rad' 'rad/s'});
nlgr = setpar(nlgr, 'Name', {'Time-constant' 'Static gain'});
nlgr = setpar(nlgr, 'Unit', {'s' 'rad/(V*s)'});

You can also use setinit and setpar to assign values, minima, maxima, and
estimation status for all initial states or parameters simultaneously.

5. View the initial model.

a. Get basic information about the model.

The DC-motor has 2 (initial) states and 2 model parameters.

size(nlgr)

Nolinear grey-box model with 2 outputs, 1 inputs, 2 states and 2 parameters (2 free).

b. View the initial states and parameters.

Both the initial states and parameters are structure arrays. The fields specify the
properties of an individual initial state or parameter. Type idprops idnlgrey
InitialStates and idprops idnlgrey Parameters for more information.

nlgr.InitialStates(1)
nlgr.Parameters(2)

ans =

Name: 'Angular position'

6-25
6 ODE Parameter Estimation (Grey-Box Modeling)

Unit: 'rad'
Value: 0
Minimum: -Inf
Maximum: Inf
Fixed: 1

ans =

Name: 'Static gain'

Unit: 'rad/(V*s)'
Value: 0.2800
Minimum: -Inf
Maximum: Inf
Fixed: 0

c. Retrieve information for all initial states or model parameters in one call.

For example, obtain information on initial states that are fixed (not estimated) and the
minima of all model parameters.

getinit(nlgr, 'Fixed')
getpar(nlgr, 'Min')

ans =

[1]
[1]

ans =

[-Inf]
[-Inf]

d. Obtain basic information about the object:

nlgr

nlgr =
Continuous-time nonlinear grey-box model defined by 'dcmotor_m' (MATLAB file):

6-26
 Estimating Nonlinear Grey-Box Models

dx/dt = F(t, u(t), x(t), p1, p2)

y(t) = H(t, u(t), x(t), p1, p2) + e(t)

with 1 input, 2 states, 2 outputs, and 2 free parameters (out of 2).

Use get to obtain more information about the model properties. The idnlgrey object
shares many properties of parametric linear model objects.

get(nlgr)

FileName: 'dcmotor_m'
Order: [1x1 struct]
Parameters: [2x1 struct]
InitialStates: [2x1 struct]
FileArgument: {}
CovarianceMatrix: 'estimate'
EstimationInfo: [1x1 struct]
TimeVariable: 't'
NoiseVariance: [2x2 double]
Algorithm: [1x1 struct]
Ts: 0
TimeUnit: 'seconds'
InputName: {'Voltage'}
InputUnit: {'V'}
InputGroup: [1x1 struct]
OutputName: {2x1 cell}
OutputUnit: {2x1 cell}
OutputGroup: [1x1 struct]
Name: 'DC-motor'
Notes: {}
UserData: []

Performance Evaluation of the Initial DC-Motor Model

Before estimating the parameters tau and k, simulate the output of the system with
the parameter guesses using the default differential equation solver (a Runge-Kutta 45
solver with adaptive step length adjustment).

1. Set the absolute and relative error tolerances to small values (1e-6 and 1e-5,
respectively).

nlgr.Algorithm.SimulationOptions.AbsTol = 1e-6;

6-27
6 ODE Parameter Estimation (Grey-Box Modeling)

nlgr.Algorithm.SimulationOptions.RelTol = 1e-5;

2. Compare the simulated output with the measured data.

compare displays both measured and simulated outputs of one or more models, whereas
predict, called with the same input arguments, displays the simulated outputs.

The simulated and measured outputs are shown in a plot window.

compare(z, nlgr);

Figure 3: Comparison between measured outputs and the simulated outputs of the
initial DC-motor model.

Parameter Estimation

6-28
 Estimating Nonlinear Grey-Box Models

Estimate the parameters and initial states using pem (Prediction-Error identification
Method).

nlgr = setinit(nlgr, 'Fixed', {false false}); % Estimate the initial state.

nlgr = pem(z, nlgr, 'Display', 'Full');

Performance Evaluation of the Estimated DC-Motor Model

1. Review the information about the estimation process.

This information is stored in the EstimationInfo property of the idnlgrey object. The
property also contains information about how the model was estimated, such as solver
and search method, data set, and why the estimation was terminated.

nlgr.EstimationInfo

ans =

Status: 'Estimated model (PEM)'

Method: 'Solver: ode45; Search: lsqnonlin'
LossFcn: 0.0011
FPE: 0.0011
DataName: 'DC-motor'
DataLength: 400
DataTs: {[0.1000]}
DataDomain: 'Time'
DataInterSample: {'zoh'}
WhyStop: 'Change in cost was less than the specified tolerance'
UpdateNorm: []
LastImprovement: []
Iterations: 5
InitialGuess: [1x1 struct]
Warning: ''
EstimationTime: 10.9981

2. Evaluate the model quality by comparing simulated and measured outputs.

The fits are 98% and 84%, which indicate that the estimated model captures the
dynamics of the DC motor well.

compare(z, nlgr);

6-29
6 ODE Parameter Estimation (Grey-Box Modeling)

Figure 4: Comparison between measured outputs and the simulated outputs of the
estimated IDNLGREY DC-motor model.

3. Compare the performance of the idnlgrey model with a second-order ARX model.

na = [2 2; 2 2];
nb = [2; 2];
nk = [1; 1];
dcarx = arx(z, [na nb nk]);
compare(z, nlgr, dcarx);

6-30
 Estimating Nonlinear Grey-Box Models

Figure 5: Comparison between measured outputs and the simulated outputs of the
estimated IDNLGREY and ARX DC-motor models.

4. Check the prediction errors.

The prediction errors obtained are small and are centered around zero (non-biased).

pe(z, nlgr);

6-31
6 ODE Parameter Estimation (Grey-Box Modeling)

Figure 6: Prediction errors obtained with the estimated IDNLGREY DC-motor model.

5. Check the residuals ("leftovers").

Residuals indicate what is left unexplained by the model and are small for good model
quality. Execute the following two lines of code to generate the residual plot. Press any
key to advance from one plot to another.

figure('Name', [nlgr.Name ': residuals of estimated model']);

resid(z, nlgr);

6-32
Estimating Nonlinear Grey-Box Models

6-33
6 ODE Parameter Estimation (Grey-Box Modeling)

Figure 7: Residuals obtained with the estimated IDNLGREY DC-motor model.

6. Plot the step response.

A unit input step results in an angular position showing a ramp-type behavior and to an
angular velocity that stabilizes at a constant level.

figure('Name', [nlgr.Name ': step response of estimated model']);

step(nlgr);

6-34
 Estimating Nonlinear Grey-Box Models

Figure 8: Step response with the estimated IDNLGREY DC-motor model.

7. Examine the model covariance.

You can assess the quality of the estimated model to some extent by looking at the
estimated covariance matrix and the estimated noise variance. A "small" value of the (i,
i) diagonal element of the covariance matrix indicates that the i:th model parameter is
important for explaining the system dynamics when using the chosen model structure.
Small noise variance (covariance for multi-output systems) elements are also a good
indication that the model captures the estimation data in a good way.
nlgr.CovarianceMatrix
nlgr.NoiseVariance

6-35
6 ODE Parameter Estimation (Grey-Box Modeling)

ans =

1.0e-04 *

0.1521 0.0015
0.0015 0.0007

ans =

0.0099 -0.0004
-0.0004 0.1094

For more information about the estimated model, use present to display the initial
states and estimated parameter values, and estimated uncertainty (standard deviation)
for the parameters.
present(nlgr);

nlgr =
Continuous-time nonlinear grey-box model defined by 'dcmotor_m' (MATLAB file):

dx/dt = F(t, u(t), x(t), p1, p2)

y(t) = H(t, u(t), x(t), p1, p2) + e(t)

with 1 input, 2 states, 2 outputs, and 2 free parameters (out of 2).

Input:
u(1) Voltage(t) [V]
States: initial value
x(1) Angular position(t) [rad] xinit@exp1 0.0302675 (est) in [-Inf, Inf]
x(2) Angular velocity(t) [rad/s] xinit@exp1 -0.133777 (est) in [-Inf, Inf]
Outputs:
y(1) Angular position(t) [rad]
y(2) Angular velocity(t) [rad/s]
Parameters: value standard dev
p1 Time-constant [s] 0.243649 0.00390034 (est) in [-Inf, Inf]
p2 Static gain [rad/(V*s)] 0.249644 0.000272168 (est) in [-Inf, Inf]

The model was estimated from the data set 'DC-motor', which
contains 400 data samples.
Loss function 0.00107459 and Akaike's FPE 0.00108534
Created: 04-Sep-2014 10:06:38

6-36
 Estimating Nonlinear Grey-Box Models

Last modified: 04-Sep-2014 10:06:53

Conclusions

This example illustrates the basic tools for performing nonlinear grey-box modeling. See
the other nonlinear grey-box examples to learn about:

• Using nonlinear grey-box models in more advanced modeling situations, such as

building nonlinear continuous- and discrete-time, time-series and static models.

• Writing and using C MEX model-files.

• Handling nonscalar parameters.

• Impact of certain algorithm choices.

For more information on identification of dynamic systems with System Identification

Toolbox, visit the System Identification Toolbox product information page.

6-37
6 ODE Parameter Estimation (Grey-Box Modeling)

After Estimating Grey-Box Models

After estimating linear and nonlinear grey-box models, you can simulate the model
output using the sim command. For more information, see “Validating Models After
Estimation”.

The toolbox represents linear grey-box models using the idgrey model object. To
convert grey-box models to state-space form, use the idss command, as described in
“Transforming Between Linear Model Representations”. You must convert your model to
an idss object to perform input-output concatenation or to use sample time conversion
functions (c2d, d2c, d2d).

Note: Sample-time conversion functions require that you convert idgrey models with
FcnType ='cd' to idss models.

The toolbox represents nonlinear grey-box models as idnlgrey model objects. These
model objects store the parameter values resulting from the estimation. You can access
these parameters from the model objects to use these variables in computation in the
MATLAB workspace.

Note: Linearization of nonlinear grey-box models is not supported.

You can import nonlinear and linear grey box models into a Simulink model using the
System Identification Toolbox Block Library. For more information, see “Simulating
Identified Model Output in Simulink”.

6-38
 Estimating Coefficients of ODEs to Fit Given Solution

Estimating Coefficients of ODEs to Fit Given Solution

This example shows how to estimate model parameters using linear and nonlinear grey-
box modeling.

Use grey-box identification to estimate coefficients of ODEs that describe the model
dynamics to fit a given response trajectory.

• For linear dynamics, represent the model using a linear grey-box model (idgrey).
Estimate the model coefficients using greyest.
• For nonlinear dynamics, represent the model using a nonlinear grey-box model
(idnlgrey). Estimate the model coefficients using pem.

In this example, you estimate the value of the friction coefficient of a simple pendulum
using its oscillation data. The motion of a simple pendulum is described by:

ml 2q&& + bq& + mglsinq = 0

θ is the angular displacement of the pendulum relative to its state of rest. g is the
gravitational acceleration constant. m is the mass of the pendulum and l is the length of
the pendulum. b is the viscous friction coefficient whose value will be estimated to fit the
given angular displacement data. There is no external driving force that is contributing
to the pendulum’s motion.

Load measured data.

load(fullfile(matlabroot, 'toolbox', 'ident', ...
'iddemos', 'data', 'pendulumdata'));
data = iddata(y, [], 0.1, 'Name', 'Pendulum');
data.OutputName = 'Pendulum position';
data.OutputUnit = 'rad';
data.Tstart = 0;
data.TimeUnit = 's';

The measured angular displacement data is loaded and saved as data, an iddata object
with a sampling time of 0.1 seconds. The set command is used to specify data attributes
such as the output name, output unit and the start time and units of the time vector.

Perform linear grey-box estimation.

Assuming that the pendulum undergoes only small angular displacements, the equation
describing the pendulum motion can be simplified:

6-39
6 ODE Parameter Estimation (Grey-Box Modeling)

ml 2q&& + bq& + mglq = 0

Using the angular displacement ( q ) and the angular velocity ( q& ) as state variables, the
simplified equation can be rewritten in the form:

.
X (t) = AX ( t) + Bu(t)
y( t) = CX (t) + Du(t)

Here,

Èq ˘
X (t) = Í & ˙
Îq ˚
È 0 1 ˘
A = Í-g -b ˙
˙
Í
ÍÎ l ml 2 ˙˚
B=0
C = [1 0 ]
D=0

The B and D matrices are zero because there is no external driving force for the simple
pendulum.

1 Create an ODE file that relates the model coefficients to its state space
representation.
function [A,B,C,D] = LinearPendulum(m,g,l,b,Ts)
% Function mapping ODE coefficients to state-space matrices.
% Save this function as a file on your computer.
A = [0 1; -g/l, -b/m/l^2];
B = zeros(2,0);
C = [1 0];
D = zeros(1,0);
end

The function, LinearPendulum, returns the state space representation of the linear
motion model of the simple pendulum using the model coefficients m, g, l and b. The
sample time used is specified by Ts.

6-40
 Estimating Coefficients of ODEs to Fit Given Solution

Save this function as LinearPendulum.m.

2 Create a linear grey-box model associated with the LinearPendulum function.

m = 1;
g = 9.81;
l = 1;
b = 0.2;
linear_model = idgrey('LinearPendulum', {m,g,l,b},'c');

m, g and l specify the values of the known model coefficients. b specifies the initial
guess for the viscous friction coefficient.

The function LinearPendulum must be on the MATLAB path. Alternatively, you

can specify the full path name for this function.

The 'c' input argument in the call to idgrey specifies linear_model as a

continuous-time system.
3 Specify m, g and l as known parameters.

linear_model.Structure.Parameters(1).Free = false;
linear_model.Structure.Parameters(2).Free = false;
linear_model.Structure.Parameters(3).Free = false;

As defined in the previous step, m, g, and l are the first three parameters of
linear_model. Using the Structure.Parameters.Free field for each of the
parameters, m, g, and l are specified as fixed values.
4 Create an estimation options set that specifies the initial state to be estimated and
turns on the estimation progress display.

opt = greyestOptions('InitialState','estimate','Display','on');
opt.Focus = 'stability';
5 Estimate the viscous friction coefficient.

linear_model = greyest(data,linear_model,opt);

The greyest command updates the parameter of linear_model.

b_est = linear_model.Structure.Parameters(4).Value;
[linear_b_est, dlinear_b_est] = getpvec(linear_model, 'free')

6-41
6 ODE Parameter Estimation (Grey-Box Modeling)

getpvec returns, as dlinear_b_est, the 1 standard deviation uncertainty

associated with b, the free estimation parameter of linear_model.

The estimated value of b, the viscous friction coefficient, using linear grey-box
estimation is 0.1178 with a standard deviation of 0.0088.
6 Compare the response of the linear grey-box model to the measured data.

compare(data,linear_model)

The linear grey-box estimation model provides a 49.9% fit to measured data. This is
not surprising given that the underlying assumption of the linear pendulum motion
model is that the pendulum undergoes small angular displacements, whereas the
measured data shows large oscillations.

Perform nonlinear grey-box estimation.

Nonlinear grey-box estimation requires that you express the differential equation as a set
of first order equations.

6-42
 Estimating Coefficients of ODEs to Fit Given Solution

Using the angular displacement ( q ) and the angular velocity ( q& ) as state variables,
the equation of motion can be rewritten as a set of first order nonlinear differential
equations:

x1 (t) = q ( t)
x (t) = q&( t)
2
x& 1 (t) = x2 ( t)
-g b
x& 2 (t) = sin( x1 ( t)) - x2 (t)
l ml 2
y( t) = x1 (t)

1 Create an ODE file that relates the model coefficients to its nonlinear
representation.

function [dx,y] = NonlinearPendulum(t,x,u,m,g,l,b,varargin)

% Function that maps the ODE coefficients to state
% variable derivatives and output.
% Save this function as a file on your computer.

% Output equation.
y = x(1); % Angular position.

% State equations.
dx = [x(2); ... % Angular position
-(g/l)*sin(x(1))-b/(m*l^2)*x(2) ... % Angular velocity
];
end

The function, NonlinearPendulum, returns the state derivatives and output of the
nonlinear motion model of the pendulum using the model coefficients m, g, l and b.

Save this function as NonlinearPendulum.m.

2 Create a nonlinear grey-box model associated with the NonlinearPendulum
function.

m = 1;
g = 9.81;
l = 1;
b = 0.2;

6-43
6 ODE Parameter Estimation (Grey-Box Modeling)

order = [1 0 2];
parameters = {m,g,l,b};
initial_states = [1; 0];
Ts = 0;
nonlinear_model = idnlgrey('NonlinearPendulum', order, ...
parameters, initial_states, Ts);
3 Specify m, g and l as known parameters.

setpar(nonlinear_model, 'Fixed', {true true true false});

As defined in the previous step, m, g, and l are the first three parameters of
nonlinear_model. Using the setpar command, m, g, and l are specified as fixed
values and b is specified as a free estimation parameter.
4 Estimate the viscous friction coefficient.

nonlinear_model = pem(data,nonlinear_model,'Display','Full');

The pem command updates the parameter of nonlinear_model.

b_est = nonlinear_model.Parameters(4).Value;
[nonlinear_b_est, dnonlinear_b_est] = ...
getpvec(nonlinear_model, 'free')

getpvec returns, as dnonlinear_b_est, the 1 standard deviation uncertainty

associated with b, the free estimation parameter of nonlinear_model.

The estimated value of b, the viscous friction coefficient, using nonlinear grey-box
estimation is 0.1 with a standard deviation of 0.0149.
5 Compare the response of the linear and nonlinear grey-box models to the measured
data.

compare(data,linear_model,nonlinear_model)

6-44
 Estimating Coefficients of ODEs to Fit Given Solution

The nonlinear grey-box model estimation provides a closer fit (95%) to the measured
data.

6-45
6 ODE Parameter Estimation (Grey-Box Modeling)

Estimate Model Using Zero/Pole/Gain Parameters

This example shows how to estimate a model that is parameterized by poles, zeros, and
gains. The example requires Control System Toolbox™ software.

You parameterize the model using complex-conjugate pole/zero pairs. When you
parameterize a real, grey-box model using complex-conjugate pairs of parameters, the
software updates parameter values such that the estimated values are also complex
conjugate pairs.

Load the measured data.

load zpkestdata zd;

The variable zd, which contains measured data, is loaded into the MATLAB® workspace.

plot(zd);

6-46
 Estimate Model Using Zero/Pole/Gain Parameters

The output shows an input delay of approximately 3.14 seconds.

Estimate the model using the zero-pole-gain (zpk) form using the zpkestODE function.
To view this function, enter

type zpkestODE

function [a,b,c,d] = zpkestODE(z,p,k,Ts,varargin)

%zpkestODE ODE file that parameterizes a state-space model using poles and
%zeros as its parameters.
%
% Requires Control System Toolbox.

6-47
6 ODE Parameter Estimation (Grey-Box Modeling)

% Copyright 2011 The MathWorks, Inc.

sysc = zpk(z,p,k);
if Ts==0
[a,b,c,d] = ssdata(sysc);
else
[a,b,c,d] = ssdata(c2d(sysc,Ts,'foh'));
end

Create a linear grey-box model associated with the ODE function.

Assume that the model has five poles and four zeros. Assume that two of the poles and
two of the zeros are complex conjugate pairs.

z = [-0.5+1i, -0.5-1i, -0.5, -1];

p = [-1.11+2i, -1.11-2i, -3.01, -4.01, -0.02];
k = 10.1;
parameters = {z,p,k};
Ts = 0;
odefun = @zpkestODE;
init_sys = idgrey(odefun,parameters,'cd',{},Ts,'InputDelay',3.14);

z, p, and k are the initial guesses for the model parameters.

init_sys is an idgrey model that is associated with the zpkestODE.m function. The
'cd' flag indicates that the ODE function, zpkestODE, returns continuous or discrete
models, depending on the sampling period.

Evaluate the quality of the fit provided by the initial model.

compare(zd,init_sys);

6-48
 Estimate Model Using Zero/Pole/Gain Parameters

The initial model provides a poor fit (18%).

Specify estimation options.

opt = greyestOptions('InitialState','zero','DisturbanceModel','none','SearchMethod','gn

Estimate the model.

sys = greyest(zd,init_sys,opt);

sys, an idgrey model, contains the estimated zero-pole-gain model parameters.

Compare the estimated and initial parameter values.

[getpvec(init_sys) getpvec(sys)]

6-49
6 ODE Parameter Estimation (Grey-Box Modeling)

ans =

-0.5000 + 1.0000i -1.6158 + 1.6173i

-0.5000 - 1.0000i -1.6158 - 1.6173i
-0.5000 + 0.0000i -0.9417 + 0.0000i
-1.0000 + 0.0000i -1.4099 + 0.0000i
-1.1100 + 2.0000i -2.4050 + 1.4340i
-1.1100 - 2.0000i -2.4050 - 1.4340i
-3.0100 + 0.0000i -2.3387 + 0.0000i
-4.0100 + 0.0000i -2.3392 + 0.0000i
-0.0200 + 0.0000i -0.0082 + 0.0000i
10.1000 + 0.0000i 9.7881 + 0.0000i

The getpvec command returns the parameter values for a model. In the output above,
each row displays corresponding initial and estimated parameter values. All parameters
that were initially specified as complex conjugate pairs remain so after estimation.

Evaluate the quality of the fit provided by the estimated model.

compare(zd,init_sys,sys);

6-50
 Estimate Model Using Zero/Pole/Gain Parameters

sys provides a closer fit (98.35%) to the measured data.

See Also
c2d | getpvec | greyest | idgrey | ssdata

Related Examples
• “Estimating Coefficients of ODEs to Fit Given Solution” on page 6-39
• “Creating IDNLGREY Model Files” on page 6-52

6-51
6 ODE Parameter Estimation (Grey-Box Modeling)

Creating IDNLGREY Model Files

This example shows how to write ODE files for nonlinear grey-box models as MATLAB
and C MEX files.

Grey box modeling is conceptually different to black box modeling in that it involves
a more comprehensive modeling step. For IDNLGREY (the nonlinear grey-box model
object; the nonlinear counterpart of IDGREY), this step consists of creating an ODE file,
also called a "model file". The ODE file specifes the right-hand sides of the state and the
output equations typically arrived at through physical first principle modeling. In this
example we will concentrate on general aspects of implementing it as a MATLAB file or a
C MEX file.

IDNLGREY Model Files

IDNLGREY supports estimation of parameters and initial states in nonlinear model

structures written on the following explicit state-space form (so-called output-error,
OE, form, named so as the noise e(t) only affects the output of the model structure in an
additive manner):
xn(t) = F(t, x(t), u(t), p1, ..., pNpo); x(0) = X0;
y(t) = H(t, x(t), u(t), p1, ..., pNpo) + e(t)

For discrete-time structures, xn(t) = x(T+Ts) with Ts being the sampling time, and for
continuous-time structures xn(t) = d/dt x(t). In addition, F(.) and H(.) are arbitrary
linear or nonlinear functions with Nx (number of states) and Ny (number of outputs)
components, respectively. Any of the model parameters p1, ..., pNpo as well as the initial
state vector X(0) can be estimated. Worth stressing is that
1. time-series modeling, i.e., modeling without an exogenous input
signal u(t), and
2. static modeling, i.e., modeling without any states x(t)

are two special cases that are supported by IDNLGREY. (See the tutorials idnlgreydemo3
and idnlgreydemo5 for examples of these two modeling categories).

The first IDNLGREY modeling step to perform is always to implement a MATLAB® or

C MEX model file specifying how to update the states and compute the outputs. More to
the point, the user must write a model file, MODFILENAME.m or MODFILENAME.c,
defined with the following input and output arguments (notice that this form is required
for both MATLAB and C MEX type of model files)
[dx, y] = MODFILENAME(t, x, u, p1, p2, ..., pNpo, FileArgument)

6-52
 Creating IDNLGREY Model Files

MODFILENAME can here be any user chosen file name of a MATLAB or C MEX-file,
e.g., see twotanks_m.m, pendulum_c.c etc. This file should be defined to return two
outputs
dx: the right-hand side(s) of the state-space equation(s) (a column
vector with Nx real entries; [] for static models)
y: the right-hand side(s) of the output equation(s) (a column vector
with Ny real entries)

and it should take 3+Npo(+1) input arguments specified as follows:

t: the current time
x: the state vector at time t ([] for static models)
u: the input vector at time t ([] for time-series models)
p1, p2, ..., pNpo: the individual parameters (which can be real
scalars, column vectors or 2-dimensional matrices); Npo is here
the number of parameter objects, which for models with scalar
parameters coincide with the number of parameters Np
FileArgument: optional inputs to the model file

In the onward discussion we will focus on writing model using either MATLAB language
or using C-MEX files. However, IDNLGREY also supports P-files (protected MATLAB
files obtained using the MATLAB command "pcode") and function handles. In fact, it
is not only possible to use C MEX model files but also Fortran MEX files. Consult the
MATLAB documentation on External Interfaces for more information about the latter.

What kind of model file should be implemented? The answer to this question really
depends on the use of the model.

Implementation using MATLAB language (resulting in a *.m file) has some distinct
advantages. Firstly, one can avoid time-consuming, low-level programming and
concentrate more on the modeling aspects. Secondly, any function available within
MATLAB and its toolboxes can be used directly in the model files. Thirdly, such files
will be smaller and, without any modifications, all built-in MATLAB error checking will
automatically be enforced. In addition, this is obtained without any code compilation.

C MEX modeling is much more involved and requires basic knowledge about the C
programming language. The main advantage with C MEX model files is the improved
execution speed. Our general advice is to pursue C MEX modeling when the model is
going to be used many times, when large data sets are employed, and/or when the model
structure contains a lot of computations. It is often worthwhile to start with using a
MATLAB file and later on turn to the C MEX counterpart.

IDNLGREY Model Files Written Using MATLAB Language

6-53
6 ODE Parameter Estimation (Grey-Box Modeling)

With this said, let us next move on to MATLAB file modeling and use a nonlinear second
order model structure, describing a two tank system, as an example. See idnlgreydemo2
for the modeling details. The contents of twotanks_m.m are as follows.
type twotanks_m.m

function [dx, y] = twotanks_m(t, x, u, A1, k, a1, g, A2, a2, varargin)

%TWOTANKS_M A two tank system.

% Copyright 2005-2006 The MathWorks, Inc.

% Output equation.
y = x(2); % Water level, lower tank.

% State equations.
dx = [1/A1*(k*u(1)-a1*sqrt(2*g*x(1))); ... % Water level, upper tank.
1/A2*(a1*sqrt(2*g*x(1))-a2*sqrt(2*g*x(2))) ... % Water level, lower tank.
];

In the function header, we here find the required t, x, and u input arguments followed
by the six scalar model parameters, A1, k, a1, g, A2 and a2. In the MATLAB file
case, the last input argument should always be varargin to support the passing of an
optional model file input argument, FileArgument. In an IDNLGREY model object,
FileArgument is stored as a cell array that might hold any kind of data. The first element
of FileArgument is here accessed through varargin{1}{1}.

The variables and parameters are referred in the standard MATLAB way. The first state
is x(1) and the second x(2), the input is u(1) (or just u in case it is scalar), and the scalar
parameters are simply accessed through their names (A1, k, a1, g, A2 and a2). Individual
elements of vector and matrix parameters are accessed as P(i) (element i of a vector
parameter named P) and as P(i, j) (element at row i and column j of a matrix parameter
named P), respectively.

IDNLGREY C MEX Model Files

Writing a C MEX model file is more involved than writing a MATLAB model file. To
simplify this step, it is recommended that the available IDNLGREY C MEX model
template is copied to MODFILENAME.c. This template contains skeleton source code as
well as detailed instructions on how to customize the code for a particular application.
The location of the template file is found by typing the following at the MATLAB
command prompt.

6-54
 Creating IDNLGREY Model Files

fullfile(matlabroot, 'toolbox', 'ident', 'nlident', 'IDNLGREY_MODEL_TEMPLATE.c')

For the two tank example, this template was copied to twotanks_c.c. After some initial
modifications and configurations (described below) the state and output equations were
entered, thereby resulting in the following C MEX source code.
type twotanks_c.c

/* Copyright 2005-2008 The MathWorks, Inc. */

/* Written by Peter Lindskog. */

/* Include libraries. */
#include "mex.h"
#include <math.h>

/* Specify the number of outputs here. */

#define NY 1

/* State equations. */
void compute_dx(double *dx, double t, double *x, double *u, double **p,
const mxArray *auxvar)
{
/* Retrieve model parameters. */
double *A1, *k, *a1, *g, *A2, *a2;
A1 = p[0]; /* Upper tank area. */
k = p[1]; /* Pump constant. */
a1 = p[2]; /* Upper tank outlet area. */
g = p[3]; /* Gravity constant. */
A2 = p[4]; /* Lower tank area. */
a2 = p[5]; /* Lower tank outlet area. */

/* x[0]: Water level, upper tank. */

/* x[1]: Water level, lower tank. */
dx[0] = 1/A1[0]*(k[0]*u[0]-a1[0]*sqrt(2*g[0]*x[0]));
dx[1] = 1/A2[0]*(a1[0]*sqrt(2*g[0]*x[0])-a2[0]*sqrt(2*g[0]*x[1]));
}

/* Output equation. */
void compute_y(double *y, double t, double *x, double *u, double **p,
const mxArray *auxvar)
{
/* y[0]: Water level, lower tank. */
y[0] = x[1];
}

6-55
6 ODE Parameter Estimation (Grey-Box Modeling)

/*----------------------------------------------------------------------- *
DO NOT MODIFY THE CODE BELOW UNLESS YOU NEED TO PASS ADDITIONAL
INFORMATION TO COMPUTE_DX AND COMPUTE_Y

To add extra arguments to compute_dx and compute_y (e.g., size

information), modify the definitions above and calls below.
*-----------------------------------------------------------------------*/

void mexFunction(int nlhs, mxArray *plhs[],

int nrhs, const mxArray *prhs[])
{
/* Declaration of input and output arguments. */
double *x, *u, **p, *dx, *y, *t;
int i, np, nu, nx;
const mxArray *auxvar = NULL; /* Cell array of additional data. */

if (nrhs < 3) {
mexErrMsgIdAndTxt("IDNLGREY:ODE_FILE:InvalidSyntax",
"At least 3 inputs expected (t, u, x).");
}

/* Determine if auxiliary variables were passed as last input. */

if ((nrhs > 3) && (mxIsCell(prhs[nrhs-1]))) {
/* Auxiliary variables were passed as input. */
auxvar = prhs[nrhs-1];
np = nrhs - 4; /* Number of parameters (could be 0). */
} else {
/* Auxiliary variables were not passed. */
np = nrhs - 3; /* Number of parameters. */
}

/* Determine number of inputs and states. */

nx = mxGetNumberOfElements(prhs[1]); /* Number of states. */
nu = mxGetNumberOfElements(prhs[2]); /* Number of inputs. */

/* Obtain double data pointers from mxArrays. */

t = mxGetPr(prhs[0]); /* Current time value (scalar). */
x = mxGetPr(prhs[1]); /* States at time t. */
u = mxGetPr(prhs[2]); /* Inputs at time t. */

p = mxCalloc(np, sizeof(double*));

6-56
 Creating IDNLGREY Model Files

for (i = 0; i < np; i++) {

p[i] = mxGetPr(prhs[3+i]); /* Parameter arrays. */
}

/* Create matrix for the return arguments. */

plhs[0] = mxCreateDoubleMatrix(nx, 1, mxREAL);
plhs[1] = mxCreateDoubleMatrix(NY, 1, mxREAL);
dx = mxGetPr(plhs[0]); /* State derivative values. */
y = mxGetPr(plhs[1]); /* Output values. */

/*
Call the state and output update functions.

Note: You may also pass other inputs that you might need,
such as number of states (nx) and number of parameters (np).
You may also omit unused inputs (such as auxvar).

For example, you may want to use orders nx and nu, but not time (t)
or auxiliary data (auxvar). You may write these functions as:
compute_dx(dx, nx, nu, x, u, p);
compute_y(y, nx, nu, x, u, p);
*/

/* Call function for state derivative update. */

compute_dx(dx, t[0], x, u, p, auxvar);

/* Call function for output update. */

compute_y(y, t[0], x, u, p, auxvar);

/* Clean up. */
mxFree(p);
}

Let us go through the contents of this file. As a first observation, we can divide the work
of writing a C MEX model file into four separate sub-steps, the last one being optional:

1. Inclusion of C-libraries and definitions of the number of outputs.

2. Writing the function computing the right-hand side(s) of the state
equation(s), compute_dx.
3. Writing the function computing the right-hand side(s) of the output
equation(s), compute_y.
4. Optionally updating the main interface function which includes
basic error checking functionality, code for creating and handling
input and output arguments, and calls to compute_dx and compute_y.

6-57
6 ODE Parameter Estimation (Grey-Box Modeling)

Before we address these sub-steps in more detail, let us briefly comment upon a couple of
general features of the C programming language.
A. High-precision variables (all inputs, states, outputs and
parameters of an IDNLGREY object) should be defined to be of the
data type "double".
B. The unary * operator placed just in front of the variable or
parameter names is a so-called dereferencing operator. The
C-declaration "double *A1;" specifies that A1 is a pointer to a
double variable. The pointer construct is a concept within C that
is not always that easy to comprehend. Fortunately, if the
declarations of the output/input variables of compute_y and
compute_x are not changed and all unpacked model parameters are
internally declared with a *, then there is no need to know more
about pointers from an IDNLGREY modeling point of view.
C. Both compute_y and compute_dx are first declared and implemented,
where after they are called in the main interface function. In the
declaration, the keyword "void" states explicitly that no value is
to be returned.

For further details of the C programming language we refer to the book

B.W. Kernighan and D. Ritchie. The C Programming Language, 2nd
edition, Prentice Hall, 1988.

1. In the first sub-step we first include the C-libraries "mex.h" (required) and
"math.h" (required for more advanced mathematics). The number of outputs is also
declared per modeling file using a standard C-define:
/* Include libraries. */
#include "mex.h"
#include "math.h"

/* Specify the number of outputs here. */

#define NY 1

If desired, one may also include more C-libraries than the ones above.

The "math.h" library must be included whenever any state or output equation contains
more advanced mathematics, like trigonometric and square root functions. Below
is a selected list of functions included in "math.h" and the counterpart found within
MATLAB:
C-function MATLAB function
========================================

6-58
 Creating IDNLGREY Model Files

sin, cos, tan sin, cos, tan

asin, acos, atan asin, acos, atan
sinh, cosh, tanh sinh, cosh, tanh
exp, log, log10 exp, log, log10
pow(x, y) x^y
sqrt sqrt
fabs abs

Notice that the MATLAB functions are more versatile than the corresponding C-
functions, e.g., the former handle complex numbers, while the latter do not.

2-3. Next in the file we find the functions for updating the states, compute_dx, and
the output, compute_y. Both these functions hold argument lists, with the output to
be computed (dx or y) at position 1, after which follows all variables and parameters
required to compute the right-hand side(s) of the state and the output equations,
respectively.

All parameters are contained in the parameter array p. The first step in compute_dx
and compute_y is to unpack and name the parameters to be used in the subsequent
equations. In twotanks_c.c, compute_dx declares six parameter variables whose values
are determined accordingly:
/* Retrieve model parameters. */
double *A1, *k, *a1, *g, *A2, *a2;
A1 = p[0]; /* Upper tank area. */
k = p[1]; /* Pump constant. */
a1 = p[2]; /* Upper tank outlet area. */
g = p[3]; /* Gravity constant. */
A2 = p[4]; /* Lower tank area. */
a2 = p[5]; /* Lower tank outlet area. */

compute_y on the other hand does not require any parameter for computing the output,
and hence no model parameter is retrieved.

As is the case in C, the first element of an array is stored at position 0. Hence, dx[0]
in C corresponds to dx(1) in MATLAB (or just dx in case it is a scalar), the input u[0]
corresponds to u (or u(1)), the parameter A1[0] corresponds to A1, and so on.

In the example above, we are only using scalar parameters, in which case the overall
number of parameters Np equals the number of parameter objects Npo. If any vector or
matrix parameter is included in the model, then Npo < Np.

The scalar parameters are referenced as P[0] (P(1) or just P in a MATLAB file) and the
i:th vector element as P[i-1] (P(i) in a MATLAB file). The matrices passed to a C MEX

6-59
6 ODE Parameter Estimation (Grey-Box Modeling)

model file are different in the sense that the columns are stacked upon each other in
the obvious order. Hence, if P is a 2-by-2 matrix, then P(1, 1) is referred as P[0], P(2,
1) as P[1], P(1, 2) as P[2] and P(2, 2) as P[3]. See "Tutorials on Nonlinear Grey Box
Identification: An Industrial Three Degrees of Freedom Robot : C MEX-File Modeling of
MIMO System Using Vector/Matrix Parameters", idnlgreydemo8, for an example where
scalar, vector and matrix parameters are used.

The state and output update functions may also include other computations than just
retrieving parameters and computing right-hand side expressions. For execution speed,
one might, e.g., declare and use intermediate variables, whose values are used several
times in the coming expressions. The robot tutorial mentioned above, idnlgreydemo8, is a
good example in this respect.

compute_dx and compute_y are also able to handle an optional FileArgument. The
FileArgument data is passed to these functions in the auxvar variable, so that the first
component of FileArgument (a cell array) can be obtained through

mxArray* auxvar1 = mxGetCell(auxvar, 0);

Here, mxArray is a MATLAB-defined data type that enables interchange of data

between the C MEX-file and MATLAB. In turn, auxvar1 may contain any data. The
parsing, checking and use of auxvar1 must be handled solely within these functions,
where it is up to the model file designer to implement this functionality. Let us here
just refer to the MATLAB documentation on External Interfaces for more information
about functions that operate on mxArrays. An example of how to use optional C MEX
model file arguments is provided in idnlgreydemo6, "Tutorials on Nonlinear Grey Box
Identification: A Signal Transmission System : C MEX-File Modeling Using Optional
Input Arguments".

4. The main interface function should almost always have the same content and for most
applications no modification whatsoever is needed. In principle, the only part that might
be considered for changes is where the calls to compute_dx and compute_y are made.
For static systems, one can leave out the call to compute_dx. In other situations, it might
be desired to only pass the variables and parameters referred in the state and output
equations. For example, in the output equation of the two tank system, where only one
state is used, one could very well shorten the input argument list to

void compute_y(double *y, double *x)

and call compute_y in the main interface function as

compute_y(y, x);

6-60
 Creating IDNLGREY Model Files

The input argument lists of compute_dx and compute_y might also be extended to
include further variables inferred in the interface function. The following integer
variables are computed and might therefore be passed on: nu (the number of inputs), nx
(the number of states), and np (here the number of parameter objects). As an example, nx
is passed to compute_y in the model investigated in the tutorial idnlgreydemo6.

The completed C MEX model file must be compiled before it can be used for IDNLGREY
modeling. The compilation can readily be done from the MATLAB command line as

mex MODFILENAME.c

Notice that the mex-command must be configured before it is used for the very first time.
This is also achieved from the MATLAB command line via

mex -setup

IDNLGREY Model Object

With an execution ready model file, it is straightforward to create IDNLGREY model

objects for which simulations, parameter estimations, and so forth can be carried out. We
exemplify this by creating two different IDNLGREY model objects for describing the two
tank system, one using the model file written in MATLAB and one using the C MEX file
detailed above (notice here that the C MEX model file has already been compiled).

Order = [1 1 2]; % Model orders [ny nu nx].

Parameters = [0.5; 0.003; 0.019; ...
9.81; 0.25; 0.016]; % Initial parameter vector.
InitialStates = [0; 0.1]; % Initial initial states.
nlgr_m = idnlgrey('twotanks_m', Order, Parameters, InitialStates, 0)
nlgr_cmex = idnlgrey('twotanks_c', Order, Parameters, InitialStates, 0)

nlgr_m =
Continuous-time nonlinear grey-box model defined by 'twotanks_m' (MATLAB file):

dx/dt = F(t, u(t), x(t), p1, ..., p6)

y(t) = H(t, u(t), x(t), p1, ..., p6) + e(t)

with 1 input, 2 states, 1 output, and 6 free parameters (out of 6).

nlgr_cmex =
Continuous-time nonlinear grey-box model defined by 'twotanks_c' (MEX-file):

dx/dt = F(t, u(t), x(t), p1, ..., p6)

6-61
6 ODE Parameter Estimation (Grey-Box Modeling)

y(t) = H(t, u(t), x(t), p1, ..., p6) + e(t)

with 1 input, 2 states, 1 output, and 6 free parameters (out of 6).

Conclusions

In this tutorial we have discussed how to write IDNLGREY MATLAB and C MEX model
files. We finally conclude the presentation by listing the currently available IDNLGREY
model files and the tutorial/case study where they are being used. To simplify further
comparisons, we list both the MATLAB (naming convention FILENAME_m.m) and the
C MEX model files (naming convention FILENAME_c.c), and indicate in the tutorial
column which type of modeling approach that is being employed in the tutorial or case
study.
Tutorial/Case study MATLAB file C MEX-file
======================================================================
idnlgreydemo1 (MATLAB) dcmotor_m.m dcmotor_c.c
idnlgreydemo2 (C MEX) twotanks_m.m twotanks_c.c
idnlgreydemo3 (MATLAB) preys_m.m preys_c.c
(C MEX) predprey1_m.m predprey1_c.c
(C MEX) predprey2_m.m predprey2_c.c
idnlgreydemo4 (MATLAB) narendrali_m.m narendrali_c.c
idnlgreydemo5 (MATLAB) friction_m.m friction_c.c
idnlgreydemo6 (C MEX) signaltransmission_m.m signaltransmission_c.c
idnlgreydemo7 (C MEX) twobodies_m.m twobodies_c.c
idnlgreydemo8 (C MEX) robot_m.m robot_c.c
idnlgreydemo9 (MATLAB) cstr_m.m cstr_c.c
idnlgreydemo10 (MATLAB) pendulum_m.m pendulum_c.c
idnlgreydemo11 (C MEX) vehicle_m.m vehicle_c.c
idnlgreydemo12 (C MEX) aero_m.m aero_c.c
idnlgreydemo13 (C MEX) robotarm_m.m robotarm_c.c

The contents of these model files can be displayed in the MATLAB command window
through the command "type FILENAME_m.m" or "type FILENAME_c.c". All model files
are found in the directory returned by the following MATLAB command.
fullfile(matlabroot, 'toolbox', 'ident', 'iddemos', 'examples')

Additional Information

For more information on identification of dynamic systems with System Identification

Toolbox™ visit the System Identification Toolbox product information page.

6-62
 7

Time Series Identification

• “What Are Time-Series Models?” on page 7-2

• “Preparing Time-Series Data” on page 7-3
• “Estimate Time-Series Power Spectra” on page 7-4
• “Estimate AR and ARMA Models” on page 7-7
• “Estimate State-Space Time-Series Models” on page 7-11
• “Identify Time-Series Models at Command Line” on page 7-12
• “Estimate Nonlinear Models for Time-Series Data” on page 7-14
• “Estimate ARIMA Models” on page 7-15
• “Spectrum Estimation Using Complex Data - Marple's Test Case” on page 7-18
• “Analyzing Time-Series Models” on page 7-28
7 Time Series Identification

What Are Time-Series Models?

A time series is one or more measured output channels with no measured input. A time-
series model, also called a signal model, is a dynamic system that is identified to fit a
given signal or time series data. The time series can be multivariate, which leads to
multivariate models.

A time series is modeled by assuming it to be the output of a system that takes a white
noise signal e(t) of variance NV as its virtual input. The true measured input size of such
models is zero, and their governing equation takes the form y(t) = He(t), where y(t) is
the signal being modeled and H is the transfer function that represents the relationship
between y(t) and e(t). The power spectrum of the time series is given by H*(NV*Ts)*H',
where NV is the noise variance and Ts is the model sample time.

System Identification Toolbox software provides tools for modeling and forecasting
time-series data. You can estimate both linear and nonlinear black-box and grey-box
models for time-series data. A linear time-series model can be a polynomial (idpoly)
or state-space (idss, idgrey) model. Some particular types of models are parametric
autoregressive (AR), autoregressive and moving average (ARMA), and autoregressive
models with integrated moving average (ARIMA).

You can estimate time-series spectra using both time- and frequency-domain data.
Time-series spectra describe time-series variations using cyclic components at different
frequencies.

The following example illustrates a 4th order autoregressive model estimation for time
series data:
load iddata9
sys = ar(z9,4);

Because the model has no measured inputs, size(sys,2) returns zero. The governing
equation of sys is A(q)y(t) = e(t). You can access the A polynomial using sys.a and the
estimated variance of the noise e(t) using sys.NoiseVariance.

7-2
 Preparing Time-Series Data

Preparing Time-Series Data

Before you can estimate models for time-series data, you must import your data into the
MATLAB software. You can only use time domain data. For information about which
variables you need to represent time-series data, see “Time-Series Data Representation”
on page 2-10.

For more information about preparing data for modeling, see “Ways to Prepare Data for
System Identification” on page 2-6.

If your data is already in the MATLAB workspace, you can import it directly into the
System Identification app. If you prefer to work at the command line, you must represent
the data as a System Identification Toolbox data object instead.

In the System Identification app. When you import scalar or multiple-output time
series data into the app, leave the Input field empty. For more information about
importing data, see “Represent Data”.

At the command line. To represent a time series vector or a matrix s as an iddata

object, use the following syntax:
y = iddata(s,[],Ts);

s contains as many columns as there are measured outputs and Ts is the sample time.

7-3
7 Time Series Identification

Estimate Time-Series Power Spectra

In this section...
“How to Estimate Time-Series Power Spectra Using the App” on page 7-4
“How to Estimate Time-Series Power Spectra at the Command Line” on page 7-5

How to Estimate Time-Series Power Spectra Using the App

You must have already imported your data into the app, as described in “Preparing
Time-Series Data” on page 7-3.

To estimate time-series spectral models in the System Identification app:

1 In the System Identification app, select Estimate > Spectral Models to open the
Spectral Model dialog box.
2 In the Method list, select the spectral analysis method you want to use. For
information about each method, see “Selecting the Method for Computing Spectral
Models” on page 3-11.
3 Specify the frequencies at which to compute the spectral model in either of the
following ways:

• In the Frequencies field, enter either a vector of values, a MATLAB expression

that evaluates to a vector, or a variable name of a vector in the MATLAB
workspace. For example, logspace(-1,2,500).
• Use the combination of Frequency Spacing and Frequencies to construct the
frequency vector of values:

• In the Frequency Spacing list, select Linear or Logarithmic frequency

spacing.

Note: For etfe, only the Linear option is available.

• In the Frequencies field, enter the number of frequency points.

For time-domain data, the frequency ranges from 0 to the Nyquist frequency. For
frequency-domain data, the frequency ranges from the smallest to the largest
frequency in the data set.

7-4
 Estimate Time-Series Power Spectra

4 In the Frequency Resolution field, enter the frequency resolution, as described

in “Controlling Frequency Resolution of Spectral Models ” on page 3-12. To use the
default value, enter default or leave the field empty.
5 In the Model Name field, enter the name of the correlation analysis model. The
model name should be unique in the Model Board.
6 Click Estimate to add this model to the Model Board in the System Identification
app.
7 In the Spectral Model dialog box, click Close.
8 To view the estimated disturbance spectrum, select the Noise spectrum check box
in the System Identification app. For more information about working with this plot,
see “Noise Spectrum Plots” on page 10-46.

To export the model to the MATLAB workspace, drag it to the To Workspace rectangle
in the System Identification app. You can view the power spectrum and the confidence
intervals of the resulting idfrd model object using the bode command.

How to Estimate Time-Series Power Spectra at the Command Line

You can use the etfe, spa, and spafdr commands to estimate power spectra of time
series for both time-domain and frequency-domain data. The following table provides a
brief description of each command.

You must have already prepared your data, as described in “Preparing Time-Series Data”
on page 7-3.

The resulting models are stored as an idfrd model object, which contains
SpectrumData and its variance. For multiple-output data, SpectrumData contains
power spectra of each output and the cross-spectra between each output pair.

Estimating Frequency Response of Time Series

Command Description
etfe Estimates a periodogram using Fourier analysis.
spa Estimates the power spectrum with its standard
deviation using spectral analysis.
spafdr Estimates the power spectrum with its standard
deviation using a variable frequency resolution.

7-5
7 Time Series Identification

For example, suppose y is time-series data. The following commands estimate the power
spectrum g and the periodogram p, and plot both models with three standard deviation
confidence intervals:
g = spa(y);
p = etfe(y);
spectrum(g,p);

For detailed information about these commands, see the corresponding reference pages.

7-6
 Estimate AR and ARMA Models

Estimate AR and ARMA Models

In this section...
“Definition of AR and ARMA Models” on page 7-7
“Estimating Polynomial Time-Series Models in the App” on page 7-7
“Estimating AR and ARMA Models at the Command Line” on page 7-10

Definition of AR and ARMA Models

For a single-output signal y(t), the AR model is given by the following equation:
A( q) y( t) = e( t)

The AR model is a special case of the ARX model with no input.

The ARMA model for a single-output time-series is given by the following equation:
A( q) y( t) = C( q) e( t)

The ARMA structure reduces to the AR structure for C(q)=1. The ARMA model is a
special case of the ARMAX model with no input.

For more information about polynomial models, see “What Are Polynomial Models?” on
page 3-40.

For information on models containing noise integration see “Estimate ARIMA Models” on
page 7-15

Estimating Polynomial Time-Series Models in the App

Before you begin, you must have accomplished the following:

• Prepared the data, as described in “Preparing Time-Series Data” on page 7-3

• Estimated model order, as described in “Preliminary Step – Estimating Model Orders
and Input Delays” on page 3-46
• (Multiple-output AR models only) Specified the model-order matrix in the MATLAB
workspace before estimation, as described in “Polynomial Sizes and Orders of Multi-
Output Polynomial Models” on page 3-61

To estimate AR and ARMA models using the System Identification app:

7-7
7 Time Series Identification

1 In the System Identification app, select Estimate > Polynomial Models to open
the Polynomial Models dialog box.
2 In the Structure list, select the polynomial model structure you want to estimate
from the following options:

• AR:[na]
• ARMA:[na nc]

This action updates the options in the Polynomial Models dialog box to correspond
with this model structure. For information about each model structure, see
“Definition of AR and ARMA Models” on page 7-7.

Note: OE and BJ structures are not available for time-series models.

3 In the Orders field, specify the model orders, as follows:

• For single-output models. Enter the model orders according to the sequence
displayed in the Structure field.
• For multiple-output ARX models. Enter the model orders directly, as
described in “Polynomial Sizes and Orders of Multi-Output Polynomial Models”.
Alternatively, enter the name of the matrix NA in the MATLAB Workspace
browser that stores model orders, which is Ny-by-Ny.

Tip To enter model orders and delays using the Order Editor dialog box, click Order
Editor.
4 (AR models only) Select the estimation Method as ARX or IV (instrumental
variable method). For more information about these methods, see “Polynomial Model
Estimation Algorithms”.

Note: IV is not available for multiple-output data.

5 Select the Add noise integration check box if you want to include an integrator
in noise source e(t). This selection changes an AR model into an ARI model
e C
( Ay = ) and an ARMA model into an ARIMA model ( Ay = e(t) ).
1- q -1
1 - q-1
6 In the Name field, edit the name of the model or keep the default. The name of the
model should be unique in the Model Board.

7-8
 Estimate AR and ARMA Models

7 In the Initial state list, specify how you want the algorithm to treat initial states.
For more information about the available options, see “Specifying Initial States for
Iterative Estimation Algorithms”.

Tip If you get an inaccurate fit, try setting a specific method for handling initial
states rather than choosing it automatically.
8 In the Covariance list, select Estimate if you want the algorithm to compute
parameter uncertainties. Effects of such uncertainties are displayed on plots as
model confidence regions.

To omit estimating uncertainty, select None. Skipping uncertainty computation

might reduce computation time for complex models and large data sets.
9 Click Regularization to obtain regularized estimates of model parameters.
Specify regularization constants in the Regularization Options dialog box. For more
information, see “Regularized Estimates of Model Parameters”.
10 To view the estimation progress at the command line, select the Display progress
check box. During estimation, the following information is displayed for each
iteration:

• Loss function — Equals the determinant of the estimated covariance matrix of

the input noise.
• Parameter values — Values of the model structure coefficients you specified.
• Search direction — Changes in parameter values from the previous iteration.
• Fit improvements — Shows the actual versus expected improvements in the fit.
11 Click Estimate to add this model to the Model Board in the System Identification
app.
12 (Prediction-error method only) To stop the search and save the results after the
current iteration has been completed, click Stop Iterations. To continue iterations
from the current model, click the Continue iter button to assign current parameter
values as initial guesses for the next search and start a new search. For the multi-
output case, you can stop iterations for each output separately. Note that the
software runs independent searches for each output.
13 To plot the model, select the appropriate check box in the Model Views area of the
System Identification app.

You can export the model to the MATLAB workspace for further analysis by dragging it
to the To Workspace rectangle in the System Identification app.

7-9
7 Time Series Identification

Estimating AR and ARMA Models at the Command Line

You can estimate AR and ARMA models at the command line. The estimated models are
represented by idpoly model objects. For more information about models objects, see
“What Are Model Objects?” on page 1-3.

The following table summarizes the commands and specifies whether single-output or
multiple-output models are supported.

Commands for Estimating Polynomial Time-Series Models

Method Name Description

ar Noniterative, least-squares method to estimate linear, discrete-time
single-output AR models.
armax Iterative prediction-error method to estimate linear ARMAX
models.
arx Noniterative, least-squares method for estimating linear AR
models.
ivar Noniterative, instrumental variable method for estimating single-
output AR models.

The following code shows usage examples for estimating AR models:

% For scalar signals
m = ar(y,na)
% For multiple-output vector signals
m = arx(y,na)
% Instrumental variable method
m = ivar(y,na)
% For ARMA, do not need to specify nb and nk
th = armax(y,[na nc])

The ar command provides additional options to let you choose the algorithm for
computing the least-squares from a group of several popular techniques from the
following methods:

• Burg (geometric lattice)

• Yule-Walker
• Covariance

7-10
 Estimate State-Space Time-Series Models

Estimate State-Space Time-Series Models

In this section...
“Definition of State-Space Time-Series Model” on page 7-11
“Estimating State-Space Models at the Command Line” on page 7-11

Definition of State-Space Time-Series Model

The discrete-time state-space model for a time series is given by the following equations:

x(kT + T ) = Ax( kT ) + Ke(kT )

y( kT ) = Cx( kT ) + e(kT)

where T is the sampling interval and y(kT) is the output at time instant kT.

The time-series structure corresponds to the general structure with empty B and D
matrices.

For information about general discrete-time and continuous-time structures for state-
space models, see “What Are State-Space Models?” on page 4-2.

Estimating State-Space Models at the Command Line

You can estimate single-output and multiple-output state-space models at the command
line for time-domain data (iddata object).

The following table provides a brief description of each command. The resulting models
are idss model objects. You can estimate either continuous-time, or discrete-time models
using these commands.

Commands for Estimating State-Space Time-Series Models

Command Description
n4sid Noniterative subspace method for estimating linear state-
space models.
ssest Estimates linear time-series models using an iterative
estimation method that minimizes the prediction error.

7-11
7 Time Series Identification

Identify Time-Series Models at Command Line

This example shows how to simulate a time-series model, compare the spectral
estimates, estimate covariance, and predict output of the model.

Generate time-series data.

ts0 = idpoly([1 -1.5 0.7],[]);
e = idinput(200,'rgs');
% Define y vector
y = sim(ts0,e);
% iddata object with sampling time 1
y = iddata(y);
plot(y);

Compare the spectral estimates.

% Estimate periodogram and spectrum
per = etfe(y);
speh = spa(y);
spectrum(per,speh,ts0);
% Estimate a second-order AR model
ts2 = ar(y,2);
spectrum(speh,ts2,ts0);

Define the true covariance function.

ir = sim(ts0,[1;zeros(24,1)]);
Ry0 = conv(ir,ir(25:-1:1));
ir2 = sim(ts2,[1;zeros(24,1)]);
Ry2 = conv(ir2,ir2(25:-1:1));

Estimate covariance.
z = [y.y ; zeros(25,1)];
j=1:200;
Ryh = zeros(25,1);
for k=1:25,
a = z(j,:)'*z(j+k-1,:);
Ryh(k) = Ryh(k)+conj(a(:));
end
Ryh = Ryh/200; % biased estimate
Ryh = [Ryh(end:-1:2); Ryh];

Alternatively, you can use the Signal Processing Toolbox command xcorr.

7-12
 Identify Time-Series Models at Command Line

Ryh = xcorr(y.y, 24 ,'biased');

Plot and compare the covariance.

plot([-24:24]'*ones(1,3),[Ryh,Ry2,Ry0]);

Predict model output.

compare(y,ts2,5);

7-13
7 Time Series Identification

Estimate Nonlinear Models for Time-Series Data

When a linear model provides an insufficient description of the dynamics, you can try
estimating a nonlinear models. To learn more about when to estimate nonlinear models,
see “Building Models from Data” in the Getting Started Guide.

Before you can estimate models for time-series data, you must have already prepared the
data as described in “Preparing Time-Series Data” on page 7-3.

For black-box modeling of time-series data, the toolbox supports nonlinear ARX models.
To learn how to estimate this type of model, see “Identifying Nonlinear ARX Models” on
page 5-7.

If you understand the underlying physics of the system, you can specify an ordinary
differential or difference equation and estimate the coefficients. To learn how to estimate
this type of model, see “Estimating Nonlinear Grey-Box Models” on page 6-15. See also
“Estimating Coefficients of ODEs to Fit Given Solution” on page 6-39 for an example of
time series modeling using the grey-box approach.

For more information about validating models, see “Validating Models After Estimation”
on page 10-2.

7-14
 Estimate ARIMA Models

Estimate ARIMA Models

This example shows how to estimate Autoregressive Integrated Moving Average or
ARIMA models.

Models of time series containing non-stationary trends (seasonality) are sometimes

required. One category of such models are the ARIMA models. These models contain a
fixed integrator in the noise source. Thus, if the governing equation of an ARMA model is
expressed as A(q)y(t)=Ce(t), where A(q) represents the auto-regressive term and C(q) the
moving average term, the corresponding model of an ARIMA model is expressed as

where the term represents the discrete-time integrator. Similarly, you can
formulate the equations for ARI and ARIX models.

Using time-series model estimation commands ar, arx and armax you can introduce
integrators into the noise source e(t). You do this by using the IntegrateNoise
parameter in the estimation command.

The estimation approach does not account any constant offsets in the time-series data.
The ability to introduce noise integrator is not limited to time-series data alone. You
can do so also for input-output models where the disturbances might be subject to
seasonality. One example is the polynomial models of ARIMAX structure:

See the armax reference page for examples.

Estimate an ARI model for a scalar time-series with linear trend.

load iddata9 z9
Ts = z9.Ts;
y = cumsum(z9.y);
model = ar(y,4,'ls','Ts',Ts,'IntegrateNoise', true);
% 5 step ahead prediction

7-15
7 Time Series Identification

compare(y,model,5)

Estimate a multivariate time-series model such that the noise integration is present in
only one of the two time series.

load iddata9 z9
Ts = z9.Ts;
y = z9.y;
y2 = cumsum(y);
% artificially construct a bivariate time series
data = iddata([y, y2],[],Ts); na = [4 0; 0 4];
nc = [2;1];
model1 = armax(data, [na nc], 'IntegrateNoise',[false; true]);
% Forecast the time series 100 steps into future

7-16
 Estimate ARIMA Models

yf = forecast(model1,data(1:100), 100);
plot(data(1:100),yf)

If the outputs were coupled ( na was not a diagonal matrix), the situation will be more
complex and simply adding an integrator to the second noise channel will not work.

7-17
7 Time Series Identification

Spectrum Estimation Using Complex Data - Marple's Test Case

This example shows how to perform spectral estimation on time series data. We use
Marple's test case (The complex data in L. Marple: S.L. Marple, Jr, Digital Spectral
Analysis with Applications, Prentice-Hall, Englewood Cliffs, NJ 1987.)

Test Data

Let us begin by loading the test data:

load marple

Most of the routines in System Identification Toolbox™ support complex data. For
plotting we examine the real and imaginary parts of the data separately, however.

First, take a look at the data:

subplot(211),plot(real(marple)),title('Real part of data.')

subplot(212),plot(imag(marple)),title('Imaginary part of data.')

7-18
 Spectrum Estimation Using Complex Data - Marple's Test Case

As a preliminary analysis step, let us check the periodogram of the data:

per = etfe(marple);
w = per.Frequency;
clf
h = spectrumplot(per,w);
opt = getoptions(h);
opt.FreqScale = 'linear';
opt.FreqUnits = 'Hz';
setoptions(h,opt)

7-19
7 Time Series Identification

Since the data record is only 64 samples, and the periodogram is computed for 128
frequencies, we clearly see the oscillations from the narrow frequency window. We
therefore apply some smoothing to the periodogram (corresponding to a frequency
resolution of 1/32 Hz):

sp = etfe(marple,32);
spectrumplot(per,sp,w);

7-20
 Spectrum Estimation Using Complex Data - Marple's Test Case

Let us now try the Blackman-Tukey approach to spectrum estimation:

ssm = spa(marple); % Function spa performs spectral estimation

spectrumplot(sp,'b',ssm,'g',w,opt);
legend({'Smoothed periodogram','Blackman-Tukey estimate'});

7-21
7 Time Series Identification

The default window length gives a very narrow lag window for this small amount of data.
We can choose a larger lag window by:

ss20 = spa(marple,20);
spectrumplot(sp,'b',ss20,'g',w,opt);
legend({'Smoothed periodogram','Blackman-Tukey estimate'});

7-22
 Spectrum Estimation Using Complex Data - Marple's Test Case

Estimating an Autoregressive (AR) Model

A parametric 5-order AR-model is computed by:

t5 = ar(marple,5);

Compare with the periodogram estimate:

spectrumplot(sp,'b',t5,'g',w,opt);
legend({'Smoothed periodogram','5th order AR estimate'});

7-23
7 Time Series Identification

The AR-command in fact covers 20 different methods for spectrum estimation. The above
one was what is known as 'the modified covariance estimate' in Marple's book.

Some other well known ones are obtained with:

tb5 = ar(marple,5,'burg'); % Burg's method

ty5 = ar(marple,5,'yw'); % The Yule-Walker method
spectrumplot(t5,tb5,ty5,w,opt);
legend({'Modified covariance','Burg','Yule-Walker'})

7-24
 Spectrum Estimation Using Complex Data - Marple's Test Case

Estimating AR Model using Instrumental Variable Approach

AR-modeling can also be done using the Instrumental Variable approach. For this, we
use the function ivar:

ti = ivar(marple,4);
spectrumplot(t5,ti,w,opt);
legend({'Modified covariance','Instrumental Variable'})

7-25
7 Time Series Identification

Autoregressive-Moving Average (ARMA) Model of the Spectra

Furthermore, System Identification Toolbox covers ARMA-modeling of spectra:

ta44 = armax(marple,[4 4]); % 4 AR-parameters and 4 MA-parameters

spectrumplot(t5,ta44,w,opt);
legend({'Modified covariance','ARMA'})

7-26
 Spectrum Estimation Using Complex Data - Marple's Test Case

Additional Information

For more information on identification of dynamic systems with System Identification

Toolbox visit the System Identification Toolbox product information page.

7-27
7 Time Series Identification

Analyzing Time-Series Models

A time-series model has no inputs. However, you can use many response computation
commands on such models. The software treats (implicitly) the noise source e(t) as a
measured input. Thus, step(sys) plots the step response assuming that the step input
was applied to the noise channel e(t).

To avoid ambiguity in how the software treats a time-series model, you can transform it
explicitly into an input-output model using noise2meas. This command causes the noise
input e(t) to be treated as a measured input and transforms the linear time series model
with Ny outputs into an input-output model with Ny outputs and Ny inputs. You can use
the resulting model with commands, such as, bode, nyquist, and iopzmap to study the
characteristics of the H transfer function. For example:
iosys = noise2meas(sys);
% step response of H if the step command was applied
% to the noise source e(t)
step(iosys)
% poles and zeros of H
iopzmap(iosys)

You can calculate and plot the time-series spectrum directly (without conversion using
noise2meas) using spectrum. For example:

spectrum(sys)

plots the time-series spectrum amplitude:

2
F(w ) = H (w)

7-28
 8

Recursive Model Identification

• “What Is Recursive Estimation?” on page 8-2

• “Data Supported for Recursive Estimation” on page 8-3
• “Algorithms for Recursive Estimation” on page 8-4
• “Data Segmentation” on page 8-11
• “Recursive Estimation and Data Segmentation Techniques in System Identification
Toolbox™” on page 8-12
8 Recursive Model Identification

What Is Recursive Estimation?

Many real-world applications, such as adaptive control, adaptive filtering, and adaptive
prediction, require a model of the system to be available online while the system is in
operation. Estimating models for batches of input-output data is useful for addressing
the following types of questions regarding system operation:

• Which input should be applied at the next sampling instant?

• How should the parameters of a matched filter be tuned?
• What are the predictions of the next few outputs?
• Has a failure occurred? If so, what type of failure?

You might also use online models to investigate time variations in system and signal
properties.

The methods for computing online models are called recursive identification methods.
Recursive algorithms are also called recursive parameter estimation, adaptive parameter
estimation, sequential estimation, and online algorithms.

For more information, see “Recursive Estimation and Data Segmentation Techniques
in System Identification Toolbox™” on page 8-12. For detailed information about
recursive parameter estimation algorithms, see the corresponding chapter in System
Identification: Theory for the User by Lennart Ljung (Prentice Hall PTR, Upper Saddle
River, NJ, 1999).

At the command line, you can recursively estimate linear polynomial models, such as
ARX, ARMAX, Box-Jenkins, and Output-Error models. For time-series data containing
no inputs and a single output, you can estimate AR (autoregressive) and ARMA
(autoregressive and moving average) single-output models.

For information about performing online estimation in Simulink, see “Online

Estimation”.

8-2
 Data Supported for Recursive Estimation

Data Supported for Recursive Estimation

To recursively estimate linear models, data must be in one of the following formats:

• Matrix of the form [y u], where y represents the output data using one or more
column vectors. u represents the input data using one or more column vectors.
• iddata object.

Related Examples
• “Recursive Estimation and Data Segmentation Techniques in System Identification
Toolbox™”

More About
• “What Is Recursive Estimation?” on page 8-2
• “Algorithms for Recursive Estimation”
• “Preliminary Step – Estimating Model Orders and Input Delays”
• “Polynomial Sizes and Orders of Multi-Output Polynomial Models”

8-3
8 Recursive Model Identification

Algorithms for Recursive Estimation

In this section...
“Types of Recursive Estimation Algorithms” on page 8-4
“General Form of Recursive Estimation Algorithm” on page 8-4
“Kalman Filter Algorithm” on page 8-5
“Forgetting Factor Algorithm” on page 8-7
“Unnormalized and Normalized Gradient Algorithms” on page 8-9

Types of Recursive Estimation Algorithms

You can choose from the following four recursive estimation algorithms:

• “General Form of Recursive Estimation Algorithm” on page 8-4

• “Kalman Filter Algorithm” on page 8-5
• “Forgetting Factor Algorithm” on page 8-7
• “Unnormalized and Normalized Gradient Algorithms” on page 8-9

You specify the type of recursive estimation algorithms as arguments in the recursive
estimation commands.

For detailed information about these algorithms, see the corresponding chapter in
System Identification: Theory for the User by Lennart Ljung (Prentice Hall PTR, Upper
Saddle River, NJ, 1999).

General Form of Recursive Estimation Algorithm

The general recursive identification algorithm is given by the following equation:

qˆ ( t ) = qˆ ( t - 1) + K ( t ) ( y ( t ) - yˆ ( t ) )

q̂ ( t) is the parameter estimate at time t. y(t) is the observed output at time t and ŷ ( t )
is the prediction of y(t) based on observations up to time t-1. The gain, K(t), determines
how much the current prediction error y ( t ) - yˆ ( t ) affects the update of the parameter
estimate. The estimation algorithms minimize the prediction-error term y ( t ) - yˆ ( t ) .

8-4
 Algorithms for Recursive Estimation

The gain has the following general form:

K ( t ) = Q ( t) y ( t )

The recursive algorithms supported by the System Identification Toolbox product differ
based on different approaches for choosing the form of Q(t) and computing y ( t ) , where
y ( t ) represents the gradient of the predicted model output yˆ ( t| q) with respect to the
parameters q .

The simplest way to visualize the role of the gradient y ( t ) of the parameters, is to
consider models with a linear-regression form:

y ( t ) = yT ( t) q0 ( t ) + e ( t )

In this equation, y ( t ) is the regression vector that is computed based on previous values
of measured inputs and outputs. q0 ( t ) represents the true parameters. e(t) is the noise
source (innovations), which is assumed to be white noise. The specific form of y ( t )
depends on the structure of the polynomial model.

For linear regression equations, the predicted output is given by the following equation:

yˆ ( t ) = yT ( t) qˆ ( t - 1 )

For models that do not have the linear regression form, it is not possible to compute
exactly the predicted output and the gradient y ( t ) for the current parameter estimate

q̂ ( t - 1 ) . To learn how you can compute approximation for y ( t ) and q̂ ( t - 1 ) for general
model structures, see the section on recursive prediction-error methods in System
Identification: Theory for the User by Lennart Ljung (Prentice Hall PTR, Upper Saddle
River, NJ, 1999).

Kalman Filter Algorithm

• “Mathematics of the Kalman Filter Algorithm” on page 8-6

8-5
8 Recursive Model Identification

• “Using the Kalman Filter Algorithm” on page 8-7

Mathematics of the Kalman Filter Algorithm

The following set of equations summarizes the Kalman filter adaptation algorithm:

qˆ ( t ) = qˆ ( t - 1) + K ( t ) ( y ( t ) - yˆ ( t ) )

yˆ ( t ) = yT ( t) qˆ ( t - 1 )

K ( t ) = Q ( t) y ( t )

P ( t - 1) y ( t )
Q (t ) =
R2 + y T ( t ) P ( t - 1) y ( t )

T
P ( t - 1 ) y ( t) y ( t ) P ( t - 1 )
P ( t ) = P ( t - 1) + R1 -
T
R2 + y ( t ) P ( t - 1 ) y ( t )

This formulation assumes the linear-regression form of the model:

y ( t ) = yT ( t) q0 ( t ) + e ( t )

The Kalman filter is used to obtain Q(t).

This formulation also assumes that the true parameters q0 ( t ) are described by a random
walk:

q0 ( t ) = q0 ( t - 1) + w ( t )

w(t) is Gaussian white noise with the following covariance matrix, or drift matrix R1:

Ew ( t ) wT ( t ) = R1

R2 is the variance of the innovations e(t) in the following equation:

8-6
 Algorithms for Recursive Estimation

y ( t ) = yT ( t) q0 ( t ) + e ( t )

The Kalman filter algorithm is entirely specified by the sequence of data y(t), the
gradient y ( t ) , R1, R2, and the initial conditions q ( t = 0 ) (initial guess of the parameters)
and P ( t = 0 ) (covariance matrix that indicates parameters errors).

Note: To simplify the inputs, you can scale R1, R2, and P ( t = 0 ) of the original problem
by the same value such that R2 is equal to 1. This scaling does not affect the parameters
estimates.

Using the Kalman Filter Algorithm

The general syntax for the command described in “Algorithms for Recursive Estimation”
on page 8-4 is the following:
[params,y_hat]=command(data,nn,adm,adg)

To specify the Kalman filter algorithm, set adm to 'kf' and adg to the value of the drift
matrix R1 (described in “Mathematics of the Kalman Filter Algorithm” on page 8-6).

Forgetting Factor Algorithm

• “Mathematics of the Forgetting Factor Algorithm” on page 8-7
• “Using the Forgetting Factor Algorithm” on page 8-8

Mathematics of the Forgetting Factor Algorithm

The following set of equations summarizes the forgetting factor adaptation algorithm:

qˆ ( t ) = qˆ ( t - 1) + K ( t ) ( y ( t ) - yˆ ( t ) )

yˆ ( t ) = yT ( t) qˆ ( t - 1 )

K ( t ) = Q ( t) y ( t )

8-7
8 Recursive Model Identification

P ( t - 1)
Q (t ) =
T
l+y ( t ) P ( t - 1 ) y ( t)

Ê P ( t - 1) y ( t ) y ( t ) P ( t - 1 ) ˆ
T
P (t ) = (t - 1) -
1 ÁP ˜
lÁ
( ) ( ) ( ) ˜¯
T
Ë l + y t P t - 1 y t

To obtain Q(t), the following function is minimized at time t:

t
 k=1 l t-k ( y(k) - y(k))2
This approach discounts old measurements exponentially such that an observation that
is t samples old carries a weight that is equal to l t times the weight of the most recent
observation. t = 1 1-l represents the memory horizon of this algorithm. Measurements
older than t = 1 1-l typically carry a weight that is less than about 0.3.

l is called the forgetting factor and typically has a positive value between 0.97 and
0.995.

Note: In the linear regression case, the forgetting factor algorithm is known as the
recursive least-squares (RLS) algorithm. The forgetting factor algorithm for l = 1 is
equivalent to the Kalman filter algorithm with R1=0 and R2=1. For more information
about the Kalman filter algorithm, see “Kalman Filter Algorithm” on page 8-5.

Using the Forgetting Factor Algorithm

The general syntax for the command described in “Algorithms for Recursive Estimation”
on page 8-4 is the following:

[params,y_hat]=command(data,nn,adm,adg)

To specify the forgetting factor algorithm, set adm to 'ff' and adg to the value of the
forgetting factor l (described in “Mathematics of the Forgetting Factor Algorithm” on
page 8-7).

8-8
 Algorithms for Recursive Estimation

Tip l typically has a positive value from 0.97 to 0.995.

Unnormalized and Normalized Gradient Algorithms

• “Mathematics of the Unnormalized and Normalized Gradient Algorithm” on page
8-9
• “Using the Unnormalized and Normalized Gradient Algorithms” on page 8-10

Mathematics of the Unnormalized and Normalized Gradient Algorithm

In the linear regression case, the gradient methods are also known as the least mean
squares (LMS) methods.

The following set of equations summarizes the unnormalized gradient and normalized
gradient adaptation algorithm:

qˆ ( t ) = qˆ ( t - 1) + K ( t ) ( y ( t ) - yˆ ( t ) )

yˆ ( t ) = yT ( t) qˆ ( t - 1 )

K ( t ) = Q ( t) y ( t )

In the unnormalized gradient approach, Q(t) is the product of the gain g and the identity
matrix:

Q( t) = g * y (t)

In the normalized gradient approach, Q(t) is the product of the gain g , and the identity
matrix is normalized by the magnitude of the gradient y ( t ) :

g * y(t)
Q (t ) =
2
y (t )

These choices of Q(t) update the parameters in the negative gradient direction, where the
gradient is computed with respect to the parameters.

8-9
8 Recursive Model Identification

Using the Unnormalized and Normalized Gradient Algorithms

The general syntax for the command described in “Algorithms for Recursive Estimation”
on page 8-4 is the following:
[params,y_hat]=command(data,nn,adm,adg)

To specify the unnormalized gain algorithm, set adm to 'ug' and adg to the value of
the gain g (described in “Mathematics of the Unnormalized and Normalized Gradient
Algorithm” on page 8-9).

To specify the normalized gain algorithm, set adm to 'ng' and adg to the value of the
gain g .

8-10
 Data Segmentation

Data Segmentation
For systems that exhibit abrupt changes while the data is being collected, you might
want to develop models for separate data segments such that the system does not change
during a particular data segment. Such modeling requires identification of the time
instants when the changes occur in the system, breaking up the data into segments
according to these time instants, and identification of models for the different data
segments.

The following cases are typical applications for data segmentation:

• Segmentation of speech signals, where each data segment corresponds to a phonem.

• Detection of trend breaks in time series.
• Failure detection, where the data segments correspond to operation with and without
failure.
• Estimating different working modes of a system.

Use segment to build polynomial models, such as ARX, ARMAX, AR, and ARMA, so that
the model parameters are piece-wise constant over time. For detailed information about
this command, see the corresponding reference page.

To see an example of using data segmentation, run the Recursive Estimation and Data
Segmentation demonstration by typing to the following command at the prompt:
iddemo5

8-11
8 Recursive Model Identification

Recursive Estimation and Data Segmentation Techniques in System

Identification Toolbox™
This example shows the use of recursive ("online") algorithms available in the System
Identification Toolbox™. It also describes the data segmentation scheme SEGMENT
which is an alternative to recursive or adaptive estimation schemes for capturing time
varying behavior. This utility is useful for capturing abrupt changes in the system
because of a failure or change of operating conditions.

Introduction

The recursive estimation functions include RPEM, RPLR, RARMAX, RARX, ROE, and
RBJ. These algorithms implement all the recursive algorithms described in Chapter 11 of
Ljung(1987).

RPEM is the general Recursive Prediction Error algorithm for arbitrary multiple-input-
single-output models (the same models as PEM works for).

PRLR is the general Recursive PseudoLinear Regression method for the same family of
models.

RARX is a more efficient version of RPEM (and RPLR) for the ARX-case.

ROE, RARMAX and RBJ are more efficient versions of RPEM for the OE, ARMAX, and
BJ cases (compare these functions to the off-line methods).

Adaptation Schemes

Each one of the algorithms implement the four most common adaptation principles:

KALMAN FILTER approach: The true parameters are supposed to vary like a random
walk with incremental covariance matrix R1.

FORGETTING FACTOR approach: Old measurements are discounted exponentially.

The base of the decay is the forgetting factor lambda.

GRADIENT method: The update step is taken as a gradient step of length gamma
(th_new = th_old + gamma*psi*epsilon).

NORMALIZED GRADIENT method: As above, but gamma is replaced by gamma/

(psi'*psi). The Gradient methods are also known as LMS (least mean squares) for the
ARX case.

8-12
 Recursive Estimation and Data Segmentation Techniques in System Identification Toolbox™

Analysis Data

In order to illustrate some of these schemes, let us pick a model and generate some input-
output data:

u = sign(randn(50,1)); % input
e = 0.2*randn(50,1); % noise
th0 = idpoly([1 -1.5 0.7],[0 1 0.5],[1 -1 0.2]); % a low order idpoly model
opt = simOptions('AddNoise',true,'NoiseData',e);
y = sim(th0,u,opt);
z = iddata(y,u);
plot(z) % analysis data object

Output Error Model Estimation Using ROE

8-13
8 Recursive Model Identification

First we build an Output-Error model of the data we just plotted. Use a second order
model with one delay, and apply the forgetting factor algorithm with lambda = 0.98:

thm1 = roe(z,[2 2 1],'ff',0.98);

The four parameters can now be plotted as functions of time.

plot(thm1), title('Estimated parameters')

legend('par1','par2','par3','par4','location','southwest')

The true values are as follows:

hold on, plot(ones(50,1)*[1 0.5 -1.5 0.7],'--','linewidth',2),

8-14
 Recursive Estimation and Data Segmentation Techniques in System Identification Toolbox™

title('Estimated parameters (solid) and true values (dashed)')

hold off

ARMAX Model Estimation USING RPLR

Now let us try a second order ARMAX model, using the RPLR approach (i.e. ELS) with
Kalman filter adaptation, assuming a parameter variance of 0.001:

thm2 = rplr(z,[2 2 2 0 0 1],'kf',0.001*eye(6));

plot(thm2), title('Estimated parameters')
legend('par1','par2','par3','par4','par5','par6','location','bestoutside')
axis([0 50 -2 2])

8-15
8 Recursive Model Identification

The true values are as follows:

hold on, plot(ones(50,1)*[-1.5 0.7 1 0.5 -1 0.2],'--','linewidth',2)

title('Estimated parameters and true values')
title('Estimated parameters (solid) and true values (dashed)')
hold off

8-16
 Recursive Estimation and Data Segmentation Techniques in System Identification Toolbox™

So far we have assumed that all data are available at once. We are thus studying the
variability of the system rather than doing real on-line calculations. The algorithms are
also prepared for such applications, but they must then store more update information.
The conceptual update then becomes:

1. Wait for measurements y and u. 2. Update: [th,yh,p,phi] = rarx([y u],[na nb

nk],'ff',0.98,th',p,phi); 3. Use th for whatever on-line application required. 4.
Go to 1.

Thus the previous estimate th is fed back into the algorithm along with the previous
value of the "P-matrix" and the data vector phi.

We now do an example of this where we plot just the current value of th.

8-17
8 Recursive Model Identification

[th,yh,p,phi] = rarx(z(1,:),[2 2 1],'ff',0.98);

plot(1,th(1),'*',1,th(2),'+',1,th(3),'o',1,th(4),'*'),
axis([1 50 -2 2]),title('Estimated Parameters'),drawnow
hold on;
for kkk = 2:50
[th,yh,p,phi] = rarx(z(kkk,:),[2 2 1],'ff',0.98,th',p,phi);
plot(kkk,th(1),'*',kkk,th(2),'+',kkk,th(3),'o',kkk,th(4),'*')
end
hold off

Data Segmentation as an Alternative To Recursive Estimation Schemes

The command SEGMENT segments data that are generated from systems that may
undergo abrupt changes. Typical applications for data segmentation are segmentation of

8-18
 Recursive Estimation and Data Segmentation Techniques in System Identification Toolbox™

speech signals (each segment corresponds to a phonem), failure detection (the segments
correspond to operation with and without failures) and estimating different working
modes of a system. We shall study a system whose time delay changes from two to one.

load iddemo6m.mat
z = iddata(z(:,1),z(:,2));

First, take a look at the data:

plot(z)

The change takes place at sample number 20, but this is not so easy to see. We would
like to estimate the system as an ARX-structure model with one a-parameter, two b-
parameters and one delay:

8-19
8 Recursive Model Identification

y(t) + a*y(t-1) = b1*u(t-1) + b2*u(t-2)

The three pieces of information to be given are: the data, the model orders, and a guess of
the variance (r2) of the noise that affects the system. If the variance is entirely unknown,
it can be estimated automatically. Here we set it to 0.1:
nn = [1 2 1];
[seg,v,tvmod] = segment(z,nn,0.1);

Let's take a look at the segmented model. On light-colored axes, the lines for the
parameters a, b1 and b2 appear in blue, green, and red colors, respectively. On dark
colored axes, these lines appear in yellow, magenta, and cyan colors, respectively.
plot(seg)
hold on

8-20
 Recursive Estimation and Data Segmentation Techniques in System Identification Toolbox™

We see clearly the jump around sample number 19. b1 goes from 0 to 1 and b2 vice versa,
which shows the change of the delay. The true values can also be shown:

plot(pars)
hold off

The method for segmentation is based on AFMM (adaptive forgetting through multiple
models), Andersson, Int. J. Control Nov 1985. A multi-model approach is used in a first
step to track the time varying system. The resulting tracking model could be of interest
in its own right, and are given by the third output argument of SEGMENT (tvmod in our
case). They look as follows:

plot(tvmod)

8-21
8 Recursive Model Identification

The SEGMENT alternative is thus an alternative to the recursive algorithms RPEM,

RARX etc for tracking time varying systems. It is particularly suited for systems that
may change rapidly.

From the tracking model, SEGMENT estimates the time points when jumps have
occurred, and constructs the segmented model by a smoothing procedure over the
tracking model.

The two most important "knobs" for the algorithm are r2, as mentioned before, and the
guessed probability of jumps, q, the fourth input argument to SEGMENT. The smaller
r2 and the larger q, the more willing SEGMENT will be to indicate segmentation points.
In an off line situation, the user will have to try a couple of choices (r2 is usually more
sensitive than q). The second output argument to SEGMENT, v, is the loss function for

8-22
 Recursive Estimation and Data Segmentation Techniques in System Identification Toolbox™

the segmented model (i.e. the estimated prediction error variance for the segmented
model). A goal will be to minimize this value.

Application of SEGMENT: Object Detection in Laser Range Data

The reflected signal from a laser (or radar) beam contains information about the distance
to the reflecting object. The signals can be quite noisy. The presence of objects affects
both the distance information and the correlation between neighboring points. (A smooth
object increases the correlation between nearby points.)

In the following we study some quite noisy laser range data. They are obtained by one
horizontal sweep, like one line on a TV-screen. The value is the distance to the reflecting
object. We happen to know that an object of interest hides between sample numbers 17
and 48.

plot(hline)

8-23
8 Recursive Model Identification

The eye is not good at detecting the object. We shall use "segment". First we detrend and
normalize the data to a variance about one. (This is not necessary, but it means that the
default choices in the algorithm are better tuned).

hline = detrend(hline)/200;

We shall now build a model of the kind:

y(t) + a y(t-1) = b

The coefficient 'a' will pick up correlation information. The value 'b' takes up the possible
changes in level. We thus introduce a fake input of all ones:

[m,n] = size(hline);

8-24
 Recursive Estimation and Data Segmentation Techniques in System Identification Toolbox™

zline = [hline ones(m,n)];

s = segment(zline,[1 1 1],0.2);
subplot(211), plot(hline),title('LASER RANGE DATA')
subplot(212), plot(s)
title('SEGMENTED MODELS, blue: correlation, green: distance')

The segmentation has thus been quite successful. SEGMENT is capable of handling
multi-input systems, and of using ARMAX models for the added noise. We can try this on
the test data iddata1.mat (which contains no jumps):

load iddata1.mat
s = segment(z1(1:100),[2 2 2 1],1);
clf
plot(s),hold on

8-25
8 Recursive Model Identification

Compare this with the true values:

plot([ones(100,1)*[-1.5 0.7],ones(100,1)*[1 0.5],ones(100,1)*[-1 0.2]],...

'--','linewidth',2)
axis([0 100 -1.6 1.1])
title('Estimated (solid) and true (dashed) parameters')
hold off

8-26
 Recursive Estimation and Data Segmentation Techniques in System Identification Toolbox™

SEGMENT thus correctly finds that no jumps have occurred, and also gives good
estimates of the parameters.

Additional Information

For more information on identification of dynamic systems with System Identification

Toolbox visit the System Identification Toolbox product information page.

8-27
8-28
 9

Online Estimation

• “What Is Online Estimation?” on page 9-2

• “How Online Estimation Differs from Offline Estimation” on page 9-4
• “Preprocess Online Estimation Data” on page 9-6
• “Validate Online Estimation Results” on page 9-7
• “Troubleshooting Online Estimation” on page 9-9
• “Generate Online Estimation Code” on page 9-12
• “Recursive Algorithms for Online Estimation” on page 9-14
• “Online Recursive Least Squares Estimation” on page 9-19
• “Online ARMAX Polynomial Model Estimation” on page 9-32
• “State Estimation Using Time-Varying Kalman Filter” on page 9-48
9 Online Estimation

What Is Online Estimation?

Online estimation algorithms estimate the parameters of a model when new data is
available during the operation of the model. Consider a heating/cooling system that does
not have prior information about the environment in which it operates. Suppose that
this system must heat/cool a room to achieve a desired temperature in a given amount
of time. To fulfil its objective, the system must obtain knowledge of the temperature and
insulation characteristics of the room. You can estimate the insulation characteristics of
the room while the heating/cooling system is online (operational). For this estimation, use
the system effort as the input and the room temperature as the output. You can feed the
estimated model to the heating/cooling system to govern its behavior so that it achieves
its objective. Online estimation is ideal for estimating small deviations in the parameter
values of a system at a known operating point.

Online estimation is typically performed using a recursive algorithm. To estimate the

parameter values at a time step, recursive algorithms use the current measurements
and previous parameter estimates. Therefore, recursive algorithms are efficient in terms
of memory storage. Additionally, recursive algorithms have smaller computational
demands. This efficiency makes them well suited to online and embedded applications.

Common applications of online estimation include:

• Adaptive control — Estimate a plant model to modify the controller based on changes
in the plant model.
• Fault detection — Compare the online plant model with the idealized/reference plant
model to detect a fault (anomaly) in the plant.
• Soft sensing — Generate a “measurement” based on the estimated plant model, and
use this measurement for feedback control or fault detection.
• Verify the experiment-data quality before starting offline estimation — Before using
the measured data for offline estimation, perform online estimation for a small
number of iterations. The online estimation provides a quick check of whether the
experiment used excitation signals that captured the relevant system dynamics. Such
a check is useful because offline estimation can be time intensive.

The System Identification Toolbox software provides the Recursive Least Squares
Estimator and Recursive Polynomial Model Estimator blocks to perform online
estimation using Simulink. You can also estimate a state-space model online from these
models by using the Recursive Polynomial Model Estimator and Model Type Converter
blocks together. You can generate C/C++ code and Structured Text for these blocks using
products such as Simulink Coder and Simulink PLC Coder™. Use the generated code

9-2
 What Is Online Estimation?

to deploy online estimation to an embedded target. For example, you can estimate the
insulation characteristics of a room from measured input-output data and feed them to
the controller for the heating/cooling system. After validating the online estimation in
simulation, you can generate code for your Simulink model and deploy the same to the
target hardware.

Requirements
When you perform online estimation using the Recursive Least Squares Estimator and
Recursive Polynomial Model Estimator blocks, the following requirements apply:

• Model must be discrete-time linear or nearly linear with parameters that vary slowly
with respect to time.
• Structure of the estimated model must be fixed while the Simulink model is running.

References
[1] Ljung, L. System Identification: Theory for the User. Upper Saddle River, NJ:
Prentice-Hall PTR, 1999, pp. 428–440.

See Also
Kalman Filter | Recursive Least Squares Estimator | Recursive Polynomial Model
Estimator

Related Examples
• “Online Recursive Least Squares Estimation” on page 9-19
• “Online ARMAX Polynomial Model Estimation” on page 9-32
• “Generate Online Estimation Code” on page 9-12

More About
• “How Online Estimation Differs from Offline Estimation” on page 9-4
• “Recursive Algorithms for Online Estimation” on page 9-14

9-3
9 Online Estimation

How Online Estimation Differs from Offline Estimation

Online estimation algorithms estimate the parameters of a model when new data is
available during the operation of the model. In contrast, if you first collect all the input/
output data and then estimate the model parameters, you perform offline estimation.
Parameter values estimated using online estimation can vary with time, but parameters
estimated using offline estimation do not.

Use the Recursive Least Squares Estimator and Recursive Polynomial Model Estimator
blocks to perform online estimation. Use tools such as arx, pem, ssest, tfest, nlarx,
and the System Identification app to perform offline estimation.

Online estimation differs from offline estimation in the following ways:

• Model delays — You can estimate model delays in offline estimation using tools
such as delayest (see “Determining Model Order and Delay”). Online estimation,
however, provides limited ability to handle delays. For polynomial model estimation
using the Recursive Polynomial Model Estimation block, you can specify as
known value of the input delay (nk) in the block dialog. If nk is unknown, choose a
sufficiently large value for the number of coefficients of B (nb). The input delay is
indicated by the number of leading coefficients of the B polynomial that are close to
zero.
• Data preprocessing — For offline estimation data preprocessing, you can use
functions such as detrend, retrend, idfilt, and the System Identification app.
For online estimation, however, you must use the tools available in the Simulink
environment. (See “Preprocess Online Estimation Data” on page 9-6.)
• Reset estimation — Online estimation allows you to reset the estimation at a specific
time step during estimation. For example, reset the estimation when the system
changes modes. In contrast, you cannot reset an offline estimation.

To reset estimation, in the online estimation block’s dialog, select the Algorithm and
Block Options tab. Select the appropriate External reset option.
• Enable/disable estimation — Online estimation allows you to enable/disable
estimation for chosen time spans. For example, suppose the measured data is
especially noisy or faulty (contains many outliers) for a specific time interval. Disable
online estimation for this interval. You cannot selectively enable/disable offline
estimation.

To enable/disable estimation, in the online estimation block’s dialog, select the

Algorithm and Block Options tab. Select the Add enable port check box.

9-4
 How Online Estimation Differs from Offline Estimation

More About
• “What Is Online Estimation?” on page 9-2

9-5
9 Online Estimation

Preprocess Online Estimation Data

Estimation data that contains deficiencies, such as drift, offset, missing samples,
seasonalities, equilibrium behavior, and outliers, can adversely affect the quality of the
estimation. Therefore, it is recommended that you preprocess your estimation data as
needed.

Use the tools in the Simulink software to preprocess data for online estimation. Common
tools to perform data preprocessing in Simulink are:

• Blocks in the “Math Operations” library. Use these blocks, for example, to subtract or
add an offset or normalize a signal.
• Blocks in the “Continuous” and “Discrete” library. Use these blocks, for example, to
filter a signal.
• Rate Transition block, which allows you to handle the transfer of data between blocks
operating at different rates. Use this block, for example, to resample your data from a
source that is operating at a different sampling rate than the online estimation block.
• MATLAB Function block, which allows you to include MATLAB code in your model.
Use this block, for example, to implement a custom preprocessing algorithm.

See “Online ARMAX Polynomial Model Estimation” on page 9-32 for an example of
how you can preprocess an estimation input signal by removing its mean.

See Also
Kalman Filter | Recursive Least Squares Estimator | Recursive Polynomial Model
Estimator

More About
• “What Is Online Estimation?” on page 9-2

9-6
 Validate Online Estimation Results

Validate Online Estimation Results

Use the following approaches to validate an online estimation performed using the
Recursive Least Squares Estimator or Recursive Polynomial Model Estimator block:

• Examine the estimation error (residuals), which is the difference between the
measured and estimated outputs. For systems whose parameters are constant or vary
slowly with respect to time, the estimation generally takes some time to converge
(settle). During this initial period, the estimation error can be high. However, after
the estimation converges, a low estimation error value gives confidence in the
estimated values.

You can also analyze the residuals using techniques such as the whiteness test and
the independence test. For such analysis, use the measured data and estimation error
collected after the parameter values have settled to approximately constant values.
For more information regarding these tests, see “What Is Residual Analysis?”

To obtain the estimation error, in the Algorithm and Block Options tab of the
online estimation block’s dialog, select the Output estimation error check box. The
software adds an Error outport to the block, which you can monitor using a Scope
block. This outport provides the one-step-ahead estimation error, e(t) = y(t)–yest(t). For
the time step, t, y and yest are the measured and estimated outputs, respectively.
• Examine the parameter covariance matrix, which measures the estimation
uncertainty. A smaller covariance value gives confidence in the estimated values.

To obtain the parameter covariance, in the Algorithm and Block Options tab
of the online estimation block’s dialog, select the Output parameter covariance
matrix check box. The software adds a Covariance outport to the block, which you
can monitor using a Display block. This outport provides the parameter covariance
matrix.
• Simulate the estimated model and compare the simulated and measured outputs.
That is, feed the measured input into a model that uses the estimated parameter
values. Then, compare this system’s output with the measured output. The simulated
output closely matching the measured output gives confidence in the estimated
values.

For examples of such validation, see “Online Recursive Least Squares Estimation” on
page 9-19 and “Online ARMAX Polynomial Model Estimation” on page 9-32.

9-7
9 Online Estimation

If the validation indicates low confidence in the estimation, then see “Troubleshooting
Online Estimation” on page 9-9 for ideas on improving the quality of the fit.

See Also
Kalman Filter | Recursive Least Squares Estimator | Recursive Polynomial Model
Estimator

9-8
 Troubleshooting Online Estimation

Troubleshooting Online Estimation

In this section...
“Check Model Structure” on page 9-9
“Check Model Order” on page 9-10
“Preprocess Estimation Data” on page 9-10
“Check Initial Guess for Parameter Values” on page 9-10
“Check Estimation Settings” on page 9-10

Validating the results of online estimation can indicate low confidence in the estimation
results. Try the following tips to improve the quality of the fit.

Check Model Structure

Check that you have chosen a model structure that is appropriate for the system to be
estimated. Ideally, you want the simplest model structure that adequately captures the
system dynamics.

• Recursive least squares (RLS) estimation — Use the Recursive Least Squares
Estimator block to estimate a system that is linear in the parameters to be estimated.

Suppose the inputs to the RLS block are simply the time-shifted versions of some
fundamental input/output variables. You can estimate this system using an ARX
model structure instead. ARX models can express the time-shifted regressors using
the A and B parameters. The autoregressive term, A(q), allows representation of
dynamics using fewer coefficients than an RLS model. Also, configuring an ARX
structure is simpler, because you provide fewer inputs.

For example, the a and b parameters of the system y(t) = b1u(t)+b2u(t-1)-a1y(t-1) can
be estimated using either recursive least squares (RLS) or ARX models. The RLS
estimation requires you to provide u(t), u(t-1) and y(t-1) as regressors. An ARX model
eliminates this requirement because you can express these time-shifted parameters
using the A and B parameters. Therefore, you provide the Recursive Polynomial
Model Estimator block only u and y. For more information regarding ARX models, see
“What Are Polynomial Models?”.
• ARX and ARMAX models — Use the Recursive Polynomial Model Estimator block to
estimate ARX models (SISO and MISO) and ARMAX models (SISO). For information
about these models, see “What Are Polynomial Models?”

9-9
9 Online Estimation

The ARMAX model has more dynamic elements (C parameters) than the ARX model
to express noise. However, ARMAX models are more sensitive to initial guess values
than ARX models, and therefore require more careful initialization.

Check Model Order

You can underfit (model order is too low) or overfit (model order is too high) data
by choosing an incorrect model order. For example, suppose you estimate the model
parameters using the Recursive Polynomial Model Estimator block and the estimated
parameters underfit the data. Increasing the number of parameters to be estimated can
improve the goodness of the fit. Ideally, you want the lowest-order model that adequately
captures the system dynamics.

Preprocess Estimation Data

Estimation data that contains deficiencies can lead to poor estimation results. Data
deficiencies include drift, offset, missing samples, equilibrium behavior, seasonalities,
and outliers. It is recommended that you preprocess the estimation data as needed. For
information on how to preprocess estimation data, see “Preprocess Online Estimation
Data” on page 9-6.

Check Initial Guess for Parameter Values

ARMAX models are especially sensitive to the initial guess of the parameter values. Poor
guesses can result in the algorithm finding a local minima of the objective function in the
parameter space, which can lead to a poor fit. You can also change the initial parameter
covariance matrix values. For uncertain initial guesses, use large values in the initial
parameter covariance matrix.

Check Estimation Settings

Check that you have specified appropriate settings for the estimation algorithm. For
example, for the Forgetting Factor algorithm, you must choose the forgetting factor
(λ) carefully. If λ is too small, the estimation algorithm assumes that the parameter
value is varying quickly with respect to time. Conversely, if λ is too large, the estimation
algorithm assumes that the parameter value does not vary much with respect to time.
For more information regarding the estimation algorithms, see “Recursive Algorithms for
Online Estimation” on page 9-14.

9-10
 Troubleshooting Online Estimation

Related Examples
• “Validate Online Estimation Results” on page 9-7

More About
• “What Is Online Estimation?” on page 9-2

9-11
9 Online Estimation

Generate Online Estimation Code

You can generate C/C++ code and Structured Text for Recursive Least Squares Estimator
and other online estimation blocks using products such as Simulink Coder and Simulink
PLC Coder. The Model Type Converter block, which you can use with the Recursive
Polynomial Model Estimator block, also supports code generation. Use the generated
code to deploy online model estimation to an embedded target. For example, you can
estimate the coefficients of a time-varying plant from measured input-output data and
feed the coefficients to an adaptive controller. After validating the online estimation in
simulation, you can generate code for your Simulink model and deploy that code to the
target.

To generate code for online estimation, use the following workflow:

1 Develop a Simulink model that simulates the online model estimation. For example,
create a model that simulates the input/output data, performs online estimation for
this data, and uses the estimated parameter values.
2 After validating the online estimation performance in simulation, create a subsystem
for the online estimation block. If you preprocess the inputs or postprocess the
parameter estimates, include the relevant blocks in the subsystem.
3 Convert the subsystem to a referenced model. You generate code for this referenced
model, so ensure that it uses only the blocks that support code generation. For a list
of blocks that support code generation, see “Simulink Built-In Blocks That Support
Code Generation”.

The original model, which now contains a model reference, is now referred to as the
top model.
4 In the top model, replace the model source and sink blocks with their counterpart
hardware blocks. For example, replace the simulated inputs/output blocks with the
relevant hardware source block. You generate code for this model, which includes the
online estimation. So, ensure that it uses only blocks that support code generation.
5 Generate code for the top model.

For details on configuring the subsystem and converting it to a referenced model, see the
“Generate Code for Referenced Models” example in the Simulink Coder documentation.

See Also
Kalman Filter | Model Type Converter | Recursive Least Squares Estimator | Recursive
Polynomial Model Estimator

9-12
 Generate Online Estimation Code

Related Examples
• “Generate Code for Referenced Models”

More About
• “Code Generation for Referenced Models”
• “Simulink Built-In Blocks That Support Code Generation”

9-13
9 Online Estimation

Recursive Algorithms for Online Estimation

In this section...
“General Form of Recursive Estimation” on page 9-14
“Types of Recursive Estimation Algorithms” on page 9-15

General Form of Recursive Estimation

The general form of the recursive estimation algorithm is as follows:

qˆ ( t ) = qˆ ( t - 1) + K ( t ) ( y ( t ) - yˆ ( t ) )

q̂ ( t) is the parameter estimate at time t. y(t) is the observed output at time t and ŷ ( t )
is the prediction of y(t) based on observations up to time t-1. The gain, K(t), determines
how much the current prediction error y ( t ) - yˆ ( t ) affects the update of the parameter
estimate. The estimation algorithms minimize the prediction-error term y ( t ) - yˆ ( t ) .

The gain has the following form:

K ( t ) = Q ( t) y ( t )

The recursive algorithms supported by the System Identification Toolbox product differ
based on different approaches for choosing the form of Q(t) and computing y ( t ) , where
y ( t ) represents the gradient of the predicted model output yˆ ( t| q) with respect to the
parameters q .

The simplest way to visualize the role of the gradient y ( t ) of the parameters, is to
consider models with a linear-regression form:

y ( t ) = yT ( t) q0 ( t ) + e ( t )

In this equation, y ( t ) is the regression vector that is computed based on previous values
of measured inputs and outputs. q0 ( t ) represents the true parameters. e(t) is the noise

9-14
 Recursive Algorithms for Online Estimation

source (innovations), which is assumed to be white noise. The specific form of y ( t )

depends on the structure of the polynomial model.

For linear regression equations, the predicted output is given by the following equation:

yˆ ( t ) = yT ( t) qˆ ( t - 1 )

For models that do not have the linear regression form, it is not possible to compute
exactly the predicted output and the gradient y ( t ) for the current parameter estimate

q̂ ( t - 1 ) . To learn how you can compute approximation for y ( t ) and q̂ ( t - 1 ) for general
model structures, see the section on recursive prediction-error methods in [1].

Types of Recursive Estimation Algorithms

The System Identification Toolbox software provides the following recursive estimation
algorithms for online estimation:

• “Forgetting Factor” on page 9-15

• “Kalman Filter” on page 9-16
• “Normalized and Unnormalized Gradient” on page 9-18

The forgetting factor and Kalman Filter formulations are more computationally intensive
than gradient and unnormalized gradient methods. However, they have much better
convergence properties.

Forgetting Factor

The following set of equations summarizes the forgetting factor adaptation algorithm:

qˆ ( t ) = qˆ ( t - 1) + K ( t ) ( y ( t ) - yˆ ( t ) )

yˆ ( t ) = yT ( t) qˆ ( t - 1 )

K ( t ) = Q ( t) y ( t )

9-15
9 Online Estimation

P ( t - 1)
Q (t ) =
T
l+y ( t ) P ( t - 1 ) y ( t)

Ê P ( t - 1) y ( t ) y ( t ) P ( t - 1 ) ˆ
T
P (t ) = 1 ÁP
(t - 1) - ˜
lÁ
( ) ( ) ( ) ˜¯
T
Ë l + y t P t - 1 y t

Q(t) is obtained by minimizing the following function at time t:

t
 k=1 l t-k ( y(k) - y(k))2
See section 11.2 in [1] for details.

This approach discounts old measurements exponentially such that an observation that
is t samples old carries a weight that is equal to l t times the weight of the most recent
observation. t = 1 1-l represents the memory horizon of this algorithm. Measurements
older than t = 1 1-l typically carry a weight that is less than about 0.3.

l is called the forgetting factor and typically has a positive value between 0.97 and
0.995.

Note: The forgetting factor algorithm for l = 1 is equivalent to the Kalman filter
algorithm with R1=0 and R2=1. For more information about the Kalman filter algorithm,
see “Kalman Filter” on page 9-16.

Kalman Filter

The following set of equations summarizes the Kalman filter adaptation algorithm:

qˆ ( t ) = qˆ ( t - 1) + K ( t ) ( y ( t ) - yˆ ( t ) )

yˆ ( t ) = yT ( t) qˆ ( t - 1 )

9-16
 Recursive Algorithms for Online Estimation

K ( t ) = Q ( t) y ( t )

P ( t - 1) y ( t )
Q (t ) =
R2 + y T ( t ) P ( t - 1) y ( t )

T
P ( t - 1 ) y ( t) y ( t ) P ( t - 1 )
P ( t ) = P ( t - 1) + R1 -
T
R2 + y ( t ) P ( t - 1 ) y ( t )

This formulation assumes the linear-regression form of the model:

y ( t ) = yT ( t) q0 ( t ) + e ( t )

Q(t) is computed using a Kalman filter.

This formulation also assumes that the true parameters q0 ( t ) are described by a random
walk:

q0 ( t ) = q0 ( t - 1) + w ( t )

w(t) is Gaussian white noise with the following covariance matrix, or drift matrix R1:

Ew ( t ) wT ( t ) = R1

R2 is the variance of the innovations e(t) in the following equation:

y ( t ) = yT ( t) q0 ( t ) + e ( t )

The Kalman filter algorithm is entirely specified by the sequence of data y(t), the
gradient y ( t ) , R1, R2, and the initial conditions q ( t = 0 ) (initial guess of the parameters)
and P ( t = 0 ) (covariance matrix that indicates parameters errors).

Note: It is assumed that R1 and P(t = 0) matrices are scaled such that R2 = 1. This
scaling does not affect the parameter estimates.

9-17
9 Online Estimation

Normalized and Unnormalized Gradient

In the linear regression case, the gradient methods are also known as the least mean
squares (LMS) methods.

The following set of equations summarizes the unnormalized gradient and normalized
gradient adaptation algorithm:

qˆ ( t ) = qˆ ( t - 1) + K ( t ) ( y ( t ) - yˆ ( t ) )

yˆ ( t ) = yT ( t) qˆ ( t - 1 )

K ( t ) = Q ( t) y ( t )

In the unnormalized gradient approach, Q(t) is given by:

Q( t) = g * y (t)

In the normalized gradient approach, Q(t) is given by:

g * y(t)
Q (t ) =
2
y (t )

These choices of Q(t) update the parameters in the negative gradient direction, where the
gradient is computed with respect to the parameters. See pg. 372 in [1] for details.

References
[1] Ljung, L. System Identification: Theory for the User. Upper Saddle River, NJ:
Prentice-Hall PTR, 1999.

See Also
Recursive Least Squares Estimator | Recursive Polynomial Model Estimator

More About
• “What Is Online Estimation?” on page 9-2

9-18
 Online Recursive Least Squares Estimation

Online Recursive Least Squares Estimation

This example shows how to implement an online recursive least squares estimator. You
estimate a nonlinear model of an internal combustion engine and use recursive least
squares to detect changes in engine inertia.

Engine Model

The engine model includes nonlinear elements for the throttle and manifold system, and
the combustion system. The model input is the throttle angle and the model output is the
engine speed in rpm.

open_system('iddemo_engine');
sim('iddemo_engine')

9-19
9 Online Estimation

The engine model is set up with a pulse train driving the throttle angle from open to
closed. The engine response is nonlinear, specifically the engine rpm response time when
the throttle is open and closed are different.

At 100 seconds into the simulation an engine fault occurs causing the engine inertia to
increase (the engine inertia, J, is modeled in the iddemo_engine/Vehicle Dynamics
block). The inertia change causes engine response times at open and closed throttle
positions to increase. You use online recursive least squares to detect the inertia change.

open_system('iddemo_engine/rpm')

9-20
 Online Recursive Least Squares Estimation

Estimation Model

The engine model is a damped second order system with input and output nonlinearities
to account for different response times at different throttle positions. Use the recursive
least squares block to identify the following discrete system that models the engine:

Since the estimation model does not explicitly include inertia we expect the values to
change as the inertia changes. We use the changing values to detect the inertia change.

The engine has significant bandwidth up to 16Hz. Set the estimator sampling frequency
to 2*160Hz or a sample time of seconds.

Recursive Least Squares Estimator Block Setup

The terms in the estimated model are the model regressors

and inputs to the recursive least squares block that estimates the values. You can
implement the regressors as shown in the iddemo_engine/Regressors block.

9-21
9 Online Estimation

open_system('iddemo_engine/Regressors');

Configure the Recursive Least Squares Estimator block:

• Initial Estimate: None. By default, the software uses a value of 1.

• Number of parameters: 5, one for each regressor coefficient.

• Parameter Covariance Matrix: 1e4, set high as we expect the parameter values to
vary from their initial guess.

• Sample Time: .

9-22
Online Recursive Least Squares Estimation

9-23
9 Online Estimation

Click Algorithm and Block Options to set the estimation options:

• Estimation Method: Forgetting Factor

• Forgetting Factor: 1-2e-4. Since the estimated values are expected to change
with the inertia, set the forgetting factor to a value less than 1. Choose = 1-2e-4
which corresponds to a memory time constant of or 15 seconds. A 15 second
memory time ensures that significant data from both the open and closed throttle
position are used for estimation as the position is changed every 10 seconds.

• Select the Output estimation error check box. You use this block output to validate
the estimation.

• Select the Output parameter covariance matrix check box. You use this block
output to validate the estimation.

• Clear the Add enable port check box.

• External reset: None.

9-24
Online Recursive Least Squares Estimation

9-25
9 Online Estimation

Validating the Estimated Model

The Error output of the Recursive Least Squares Estimator block gives the one-
step-ahead error for the estimated model. This error is less than 5% indicating that for
one-step-ahead prediction the estimated model is accurate.
open_system('iddemo_engine/Error (%)')

9-26
 Online Recursive Least Squares Estimation

The diagonal of the parameter covariances matrix gives the variances for the
parameters. The and variances are small relative to the parameter values
indicating good confidence in the estimated values. In contrast, the covariance is large
relative to the parameter values indicating a low confidence in these values.

9-27
9 Online Estimation

While the small estimation error and covariances give confidence that the model is being
estimated correctly, it is limited in that the error is a one-step-ahead predictor. A more
rigorous check is to use the estimated model in a simulation model and compare with the
actual model output. The Estimated Model section of the simulink model implements
this.

The Regressors1 block is identical to the Regressors block use in the recursive
estimator. The only difference is that the y signal is not measured from the plant but
fed back from the output of the estimated model. The Output of the regressors block
is multiplied by estimated values to give an estimate of the engine speed. The
values were taken from a snapshot of the estimated parameter values at 60 seconds.

open_system('iddemo_engine/rpm Est')

9-28
 Online Recursive Least Squares Estimation

The estimated model output matches the model output fairly well. The steady-state
values are close and the transient behavior is slightly different but not significantly
so. Note that after 100 seconds when the engine inertia changes the estimated model
output differs more significantly from the model output. This difference is expected as
the estimated model parameters are fixed and the estimated model does not include any
information about the inertia change (the only model input is the throttle position).

The estimated model output combined with the low one-step-ahead error and parameter
covariances gives us confidence in the recursive estimator.

Detecting Changes in Engine Inertia

The engine model is setup to introduce an inertia change 100 seconds into the simulation.
The recursive estimator can be used to detect the change in inertia.

The recursive estimator takes around 50 seconds to converge to an initial set of

parameter values. To detect the inertia change we examine the model coefficient that
influences the term of the estimated model.

open_system('iddemo_engine/Detect Inertia Change')

9-29
9 Online Estimation

The covariance for , 0.0001273, is large relative to the parameter value 0.9805
indicating low confidence in the estimated value. The time plot of shows why the
covariance is large. Specifically is varying as the throttle position varies indicating
that the estimated model is not rich enough to fully capture different rise times at
different throttle positions and needs to adjust . However, we can use this to identify
the inertia changes as the average value of changes as the inertia changes. You can

9-30
 Online Recursive Least Squares Estimation

use a threshold detector on the moving average of the parameter to detect changes in
the engine inertia.
bdclose('iddemo_engine')

9-31
9 Online Estimation

Online ARMAX Polynomial Model Estimation

This example shows how to implement an online polynomial model estimator. You
estimate two ARMAX models for a nonlinear chemical reaction process. These models
capture the behavior of the process at two operating conditions. The model behavior is
identified online and used to adjust the gains of an adaptive PI controller during system
operation.

Continuously Stirred Tank Reactor

A Continuously Stirred Tank Reactor (CSTR) is a common chemical system in the

process industry. A schematic of the CSTR system is:

This is a jacketed diabatic (i.e., nondiabatic) tank reactor described extensively in

Bequette's book "Process Dynamics: Modeling, Analysis and Simulation", published by
Prentice-Hall, 1998. The vessel is assumed to be perfectly mixed, and a single first-order

9-32
 Online ARMAX Polynomial Model Estimation

exothermic and irreversible reaction, A --> B, takes place. The inlet stream of reagent
A is fed to the tank at a constant rate. After stirring, the end product streams out of the
vessel at the same rate as reagent A is fed into the tank (the volume in the reactor tank
is constant). Details of the operation of the CSTR and its 2-state nonlinear model used
in this example are explained in the example “"Non-Adiabatic Continuous Stirred Tank
Reactor: MATLAB File Modeling with Simulations in Simulink"”

The inputs of the CSTR model are:

and the outputs (y(t)), which are also the states of the model (x(t)), are:

The control objective is to maintain the concentration of reagent A, at the

desired level , which changes over time. The jacket temperature is
manipulated by a PI controller in order to reject disturbances arising from the inlet feed
stream temperature . The input of the PI controller is the tracking error signal,
. The inlet feed stream concentration, , is assumed to be constant.
The Simulink model iddemo_cstr implements the CSTR plant as the block CSTR.

open_system('iddemo_cstr');

9-33
9 Online Estimation

The feed temperature input consists of a white noise disturbance on top of a

constant offset. The noise power is 0.0075 . This noise level causes up to 2% deviation
from the desired .

The signal in this example contains a step change from 1.5 [kgmol/m^3] to 2
[kgmol/m^3] at time . In addition to this step change, also contains a
white noise perturbation for t in the [0,200) and [400,600) ranges. The power of this
white noise signal is 0.015. The noise power is adjusted empirically to approximately
give a signal-to-noise ratio of 10. Not having sufficient excitation in the reference signal

9-34
 Online ARMAX Polynomial Model Estimation

in closed-loop identification can lead to not having sufficient information to identify a

unique model. The implementation of is in the iddemo_cstr/CA Reference
block.

Online Estimation for Adaptive Control

It is known from the nonlinear model that the CSTR output is more sensitive to
the control input at higher levels. The Recursive Polynomial Model Estimator
block is used to detect this change in sensitivity. This information is used to adjust the
gains of the PI controller as varies. The aim is to avoid having a a high gain control
loop which may lead to instability.

You estimate a discrete transfer-function from to online with the Recursive

Polynomial Model Estimator block. The adaptive control algorithm uses the DC gain of
this transfer function. The tracking error , is divided by the normalized
DC gain of the estimated transfer function. This normalization is done to have a gain
of 1 on the tracking error at the initial operating point, for which the PI controller is
designed. For instance, the error signal is divided by 2 if the DC gain becomes 2 times its
original value. This corresponds to dividing the PI controller gains by 2. This adpative
controller is implemented in iddemo_cstr/Adaptive PI Controller.

Recursive Polynomial Model Estimator Block Inputs

The 'Recursive Polynomial Model Estimator' block is found under the System
Identification Toolbox/Estimators library in Simulink. You use this block to
estimate linear models with ARMAX structure. ARMAX models have the form:

• The Inputs and Output inport of the recursive polynomial model estimator
block correspond to and respectively. For the CSTR model and are
deviations from the jacket temperature and A concentration trim operating points:
, . It is good to scale and to have a peak
amplitude of 1 to improve the numerical condition of the estimation problem. The
trim operating points, and , are not known exactly before system operation.

9-35
9 Online Estimation

They are estimated and extracted from the measured signals by using a first-
order moving average filter. These preprocessing filters are implemented in the
iddemo_cstr/Preprocess Tj and iddemo_cstr/Preprocess CA blocks.

open_system('iddemo_cstr/Preprocess Tj');

• The optional Enable inport of the Recursive Polynomial Model Estimator block
controls the parameter estimation in the block. Parameter estimation is disabled
when the Enable signal is zero. Parameter estimation is enabled for all other values
of the Enable signal. In this example the estimation is disabled for the time intervals
and . During these intervals the measured input
does not contain sufficient excitation for closed-loop system identification.

Recursive Polynomial Model Estimator Block Setup

Configure the block parameters to estimate a second-order ARMAX model. In the Model
Parameters tab, specify:

• Model Structure: ARMAX. Choose ARMAX since the current and past values of the
disturbances acting on the system, , are expected to impact the CSTR system
output .

• Initial Estimate: None. By default, the software uses a value of 0 for all estimated
parameters.

• Number of parameters in A(q) (na): 2. The nonlinear model has 2 states.

• Number of parameters in B(q) (nb): 2.

9-36
 Online ARMAX Polynomial Model Estimation

• Number of parameters in C(q) (nc): 2. The estimated model corresponds to a

second order model since the maximum of na, nb, and nc are 2.

• Input Delay (nk): 1. Like most physical systems, the CSTR system does not have
direct feedthrough. Also, there are no extra time delays between its I/Os.

• Parameter Covariance Matrix: 1e4. Specify a high covariance value because the
initial guess values are highly uncertain.

• Sample Time: 0.1. The CSTR model is known to have a bandwidth of about 0.25Hz.
chosen such that 1/0.1 is greater than 20 times this bandwidth (5Hz).

9-37
9 Online Estimation

9-38
 Online ARMAX Polynomial Model Estimation

Click Algorithm and Block Options to set the estimation options:

• Estimation Method: Forgetting Factor

• Forgetting Factor: 1-5e-3. Since the estimated parameters are expected to change
with the operating point, set the forgetting factor to a value less than 1. Choose
which corresponds to a memory time constant of
seconds. A 100 second memory time ensures that a significant amount data used for
identification is coming from the 200 second identification period at each operating
point.

• Select the Output estimation error check box. You use this block output to validate
the estimation.

• Select the Add enable port check box. You only want to adapt the estimated
model parameters when extra noise is injected in the reference port. The parameter
estimation n is disabled through this port when the extra noise is no longer injected.

• External reset: None.

9-39
9 Online Estimation

9-40
 Online ARMAX Polynomial Model Estimation

Recursive Polynomial Model Estimator Block Outputs

At every time step, the recursive polynomial model estimator provides an estimate
for , , , and the estimation error . The Error outport of the polynomial
model estimator block contains and is also known as the one-step-ahead prediction
error. The Parameters outport of the block contains the A(q), B(q), and C(q) polynomial
coefficients in a bus signal. Given the chosen polynomial orders ( , , ,
) the Parameters bus elements contain:

The estimated parameters in the A(q), B(q), and C(q) polynomials change during
simulation as follows:

sim('iddemo_cstr');
open_system('iddemo_cstr/ABC');

9-41
9 Online Estimation

The parameter estimates quickly change from their initial values of 0 due to the high
value chosen for the initial parameter covariance matrix. The parameters in the
and polynomials approach their values at rapidly. However, the parameters
in the polynomial show some fluctuations. One reason behind these fluctuations
is that the disturbance to CSTR output is not fully modelled by the ARMAX
structure. The error model is not important for the control problem studied here

since the to relationship is captured by the transfer function . Therefore,

the fluctuation in is not a concern for this identification and control problem.

9-42
 Online ARMAX Polynomial Model Estimation

The parameter estimates are held constant for since the estimator block
was disabled for this interval (0 signal to the Enable inport). The parameter estimation
is enabled at when the CSTR tank starts switching to its new operating point.
The parameters of and converge to their new values by , and then held
constant by setting the Enable port to 0. The convergence of and is slower
at this operating point. This slow convergence is because of the smaller eigenvalues of
the parameter covariance matrix at t=400 compared to the initial 1e4 values set at t=0.
The parameter covariance, which is a measure of confidence in the estimates, is updated
with each time step. The algorithm quickly changed the parameter estimates when the
confidence in estimates were low at t=0. The improved parameter estimates capture
the system behavior better, resulting in smaller one-step-ahead prediction errors and
smaller eigenvalues in the parameter covariance matrix (increased confidence). The
system behavior changes in t=400. However, the block is slower to change the parameter
estimates due to the increased confidence in the estimates. You can use the External
Reset option of the Recursive Polynomial Model Estimator block to provide a new value
for parameter covariance at t=400. To see the value of the parameter covariance, select
the Output parameter covariance matrix check box in the Recursive Polynomial
Model Estimator block.

Validating the Estimated Model

The Error output of the Recursive Polynomial Model Estimator block gives the
one-step-ahead error for the estimated model.

open_system('iddemo_cstr/Error');

9-43
9 Online Estimation

The one-step-ahead error is higher when there are no extra perturbations injected in the
channel for system identification. These higher errors may be caused by the lack
of sufficient information in the input channel that the estimator block relies on.
However, even this higher error is low and bounded when compared to the measured
fluctuations in . This gives confidence in the estimated parameter values.

A more rigorous check of the estimated model is to simulate the estimated model and
compare with the actual model output. The iddemo_cstr/Time-Varying ARMAX block
implements the time-varying ARMAX model estimated by the Online Polynomial Model
Estimator block. The error between the output of the CSTR system and the estimated
time-varying ARMAX model output is:

9-44
 Online ARMAX Polynomial Model Estimation

open_system('iddemo_cstr/Simulation Error');

The simulation error is again bounded and low when compared to the fluctuations in
the . This further provides confidence that the estimated linear models are able to
predict the nonlinear CSTR model behavior.

The identified models can be further analyzed in MATLAB. The model estimates for the
operating points and can be obtained by looking
at the estimated A(q), B(q), and C(q) polynomials at and respectively.
Bode plots of these models are:

Ts = 0.1;
tidx = find(t>=200,1);

9-45
9 Online Estimation

P200 = idpoly(AHat(tidx,:),BHat(tidx,:),CHat(tidx,:),1,1,[],Ts);
tidx = find(t>=600,1);
P600 = idpoly(AHat(tidx,:),BHat(tidx,:),CHat(tidx,:),1,1,[],Ts);
bodemag(P200,'b',P600,'r--',{10^-1,20});
legend('Estimated Model at C_A=1.5 [kgmol/m^3]', ...
'Estimated Model at C_A=2.0 [kgmol/m^3]', ...
'Location', 'Best');

The estimated model has a higher gain at higher concentration levels. This is in
agreement with prior knowledge about the nonlinear CSTR plant. The transfer
function at has a higher gain (double the amplitude) at low
frequencies.

Summary

9-46
 Online ARMAX Polynomial Model Estimation

You estimated two ARMAX models to capture the behavior of the nonlinear CSTR plant
at two operating conditions. The estimation was done during closed-loop operation with
an adaptive controller. You looked at two signals to validate the estimation results: One
step ahead prediction errors and the errors between the CSTR plant output and the
simulation of the estimation model. Both of these errors signals were bounded and small
compared to the CSTR plant output. This provided confidence in the estimated ARMAX
model parameters.
bdclose('iddemo_cstr');

9-47
9 Online Estimation

State Estimation Using Time-Varying Kalman Filter

This example shows how to estimate states of linear systems using time-varying Kalman
filters in Simulink. You use the Kalman Filter block from the System Identification
Toolbox/Estimators library to estimate the position and velocity of a ground vehicle
based on noisy position measurements such as GPS sensor measurements. The plant
model in Kalman filter has time-varying noise characteristics.

Introduction

You want to estimate the position and velocity of a ground vehicle in the north and
east directions. The vehicle can move freely in the two-dimensional space without any
constraints. You design a multi-purpose navigation and tracking system that can be used
for any object and not just a vehicle.

9-48
 State Estimation Using Time-Varying Kalman Filter

and are the vehicle's east and north positions from the origin, is the
vehicle orientation from east and is the steering angle of the vehicle. is the
continuous-time variable.

9-49
9 Online Estimation

The Simulink model consists of two main parts: Vehicle model and the Kalman filter.
These are explained further in the following sections.

open_system('ctrlKalmanNavigationExample');

Vehicle Model

The tracked vehicle is represented with a simple point-mass model:

where the vehicle states are:

the vehicle parameters are:

9-50
 State Estimation Using Time-Varying Kalman Filter

and the control inputs are:

The longitunidal dynamics of the model ignore tire rolling resistance. The lateral
dynamics of the model assume that the desired steering angle can be achieved
instantaneously and ignore the yaw moment of inertia.

The car model is implemented in the ctrlKalmanNavigationExample/Vehicle

Model subsystem. The Simulink model contains two PI controllers for tracking the
desired orientation and speed for the car in the ctrlKalmanNavigationExample/
Speed And Orientation Tracking subsystem. This allows you to specify various
operating conditions for the car test the Kalman filter performance.

Kalman Filter Design

Kalman filter is an algorithm to estimate unknown variables of interest based on a linear

model. This linear model describes the evolution of the estimated variables over time in
response to model initial conditions as well as known and unknown model inputs. In this
example, you estimate the following parameters/variables:

where

9-51
9 Online Estimation

The terms denote velocities and not the derivative operator. is the discrete-time
index. The model used in the Kalman filter is of the form:

where is the state vector, is the measurements, is the process noise, and is the
measurement noise. Kalman filter assumes that and are zero-mean, independent
random variables with known variances , , and .
Here, the A, G, and C matrices are:

where

The third row of A and G model the east velocity as a random walk:
. In reality, position is a continuous-time variable
and is the integral of velocity over time . The first row of the A
and G represent a disrete approximation to this kinematic relationship:
. The second and fourth rows of the A and
G represent the same relationship between the north velocity and position.

The C matrix represents that only position measurements are available. A position
sensor, such as GPS, provides these measurements at the sample rate of 1Hz. The

9-52
 State Estimation Using Time-Varying Kalman Filter

variance of the measurment noise , the R matrix, is specified as . Since R is

specified as a scalar, the Kalman filter block assumes that the matrix R is diagonal, its
diagonals are 50 and is of compatible dimensions with y. If the measurement noise is
Gaussian, R=50 corresponds to 68% of the position measurements being within
or the actual position in the east and north directions. However, this assumption is not
necessary for the Kalman filter.

The elements of capture how much the vehicle velocity can change over one sample
time Ts. The variance of the process noise w, the Q matrix, is chosen to be time-varying.
It captures the intuition that typical values of are smaller when velocity is large.
For instance, going from 0 to 10m/s is easier than going from 10 to 20m/s. Concretely, you
use the estimated north and east velocities and a saturation function to construct Q[n]:

The diagonals of Q model the variance of w inversely proportional to the square of the
estimated velocities. The saturation function prevents Q from becoming too large or
small. The coefficient 250 is obtained from a least squares fit to 0-5, 5-10, 10-15, 15-20,
20-25m/s acceleration time data for a generic vehicle. Note that the diagonal Q implies
a naive approach that assumes that the velocity changes in north and east direction are
uncorrelated.

Kalman Filter Block Inputs and Setup

The 'Kalman Filter' block is in the System Identification Toolbox/Estimators

library in Simulink. It is also in Control System Toolbox library. Configure the block
parameters for discrete-time state estimation. Specify the following Filter Settings
parameters:

• Time domain: Discrete-time. Choose this option to estimate discrete-time states.

• Select the Use current measurement y[n] to improve the xhat[n] check box.
This implements the "current estimator" variant of the discrete-time Kalman filter.

9-53
9 Online Estimation

This option improves the estimation accuracy and is more useful for slow sample
times. However, it increases the computational cost. In addition, this Kalman filter
variant has direct feedthrough, which leads to an algebraic loop if the Kalman filter is
used in a feedback loop that does not contain any delays (the feedback loop itself also
has direct feedthrough). The algebraic loop can further impact the simulation speed.

Click the Options tab to set the block inport and outport options:

• Unselect the Add input port u check box. There are no known inputs in the plant
model.

• Select the Output state estimation error covariance Z check box. The Z matrix
provides information about the filter's confidence in the state estimates.

9-54
State Estimation Using Time-Varying Kalman Filter

9-55
9 Online Estimation

Click Model Parameters to specify the plant model and noise characteristics:

• Model source: Individual A, B, C, D matrices.

• A: A. The A matrix is defined earlier in this example.

• C: C. The C matrix is defined earlier in this example.

• Initial Estimate Source: Dialog

• Initial states x[0]: 0. This represents an initial guess of 0 for the position and
velocity estimates at t=0s.

• State estimation error covariance P[0]: 10. Assume that the error between your
initial guess x[0] and its actual value is a random variable with a standard deviation
.

• Select the Use G and H matrices (defalut G=I and H=0) check box to specify a
non-default G matrix.

• G: G. The G matrix is defined earlier in this example.

• H: 0. The process noise does not impact the measurments y entering the Kalman filter
block.

• Unselect the Time-invariant Q check box. The Q matrix is time-varying and is

supplied through the block inport Q. The block uses a time-varying Kalman filter
due to this setting. You can select this option to use a time-invariant Kalman filter.
Time-invariant Kalman filter performs slightly worse for this problem, but is easier to
design and has a lower computational cost.

• R: R. This is the covariance of the measurement noise . The R matrix is defined

earlier in this example.

• N: 0. Assume that there is no correlation between process and measurement noises.

• Sample time (-1 for inherited): Ts, which is defined earlier in this example.

9-56
State Estimation Using Time-Varying Kalman Filter

9-57
9 Online Estimation

Results

Test the performance of the Kalman filter by simulating a scenario where the vehicle
makes the following maneuvers:

• At t = 0 the vehicle is at , and is stationary.

• Heading east, it accelerates to 25m/s. It decelerates to 5m/s at t=50s.

• At t = 100s, it turns toward north and accelerates to 20m/s.

• At t = 200s, it makes another turn toward west. It accelerates to 25m/s.

• At t = 260s, it decelerates to 15m/s and makes a constant speed 180 degree turn.

Simulate the Simulink model. Plot the actual, measured and Kalman filter estimates of
vehicle position.

sim('ctrlKalmanNavigationExample');

figure(1);
% Plot results and connect data points with a solid line.
plot(x(:,1),x(:,2),'bx',...
y(:,1),y(:,2),'gd',...
xhat(:,1),xhat(:,2),'ro',...
'LineStyle','-');
title('Position');
xlabel('East [m]');
ylabel('North [m]');
legend('Actual','Measured','Kalman filter estimate','Location','Best');
axis tight;

9-58
 State Estimation Using Time-Varying Kalman Filter

The error between the measured and actual position as well as the error between the
kalman filter estimate and actual position is:

% East position measurement error [m]

n_xe = y(:,1)-x(:,1);
% North position measurement error [m]
n_xn = y(:,2)-x(:,2);
% Kalman filter east position error [m]
e_xe = xhat(:,1)-x(:,1);
% Kalman filter north position error [m]
e_xn = xhat(:,2)-x(:,2);

figure(2);
% East Position Errors

9-59
9 Online Estimation

subplot(2,1,1);
plot(t,n_xe,'g',t,e_xe,'r');
ylabel('Position Error - East [m]');
xlabel('Time [s]');
legend(sprintf('Meas: %.3f',norm(n_xe,1)/numel(n_xe)),sprintf('Kalman f.: %.3f',norm(e_
axis tight;
% North Position Errors
subplot(2,1,2);
plot(t,y(:,2)-x(:,2),'g',t,xhat(:,2)-x(:,2),'r');
ylabel('Position Error - North [m]');
xlabel('Time [s]');
legend(sprintf('Meas: %.3f',norm(n_xn,1)/numel(n_xn)),sprintf('Kalman f: %.3f',norm(e_x
axis tight;

9-60
 State Estimation Using Time-Varying Kalman Filter

The plot legends show the position measurement and estimation error ( and
) normalized by the number of data points. The Kalman filter estimates have
about 25% percent less error than the raw measurements.

The actual velocity in the east direction and its Kalman filter estimate is shown below in
the top plot. The bottom plot shows the estimation error.

e_ve = xhat(:,3)-x(:,3); % [m/s] Kalman filter east velocity error

e_vn = xhat(:,4)-x(:,4); % [m/s] Kalman filter north velocity error
figure(3);
% Velocity in east direction and its estimate
subplot(2,1,1);
plot(t,x(:,3),'b',t,xhat(:,3),'r');
ylabel('Velocity - East [m]');
xlabel('Time [s]');
legend('Actual','Kalman filter','Location','Best');
axis tight;
subplot(2,1,2);
% Estimation error
plot(t,e_ve,'r');
ylabel('Velocity Error - East [m]');
xlabel('Time [s]');
legend(sprintf('Kalman filter: %.3f',norm(e_ve,1)/numel(e_ve)));
axis tight;

9-61
9 Online Estimation

The legend on the error plot shows the east velocity estimation error
normalized by the number of data points.

The Kalman filter velocity estimates track the actual velocity trends correctly. The noise
levels decrease when the vehicle is traveling at high velocities. This is in line with the
design of the Q matrix. The large two spikes are at t=50s and t=200s. These are the times
when the car goes through sudden decelearation and a sharp turn, respectively. The
velocity changes at those instants are much larger than the predictions from the Kalman
filter, which is based on its Q matrix input. After a few time-steps, the filter estimates
catch up with the actual velocity.

Summary

9-62
 State Estimation Using Time-Varying Kalman Filter

You estimated the position and velocity of a vehicle using the Kalman filter block in
Simulink. The process noise dynamics of the model were time-varying. You validated
the filter performance by simulating various vehicle maneuvers and randomly generated
measurement noise. The Kalman filter improved the position measurements and
provided velocity estimates for the vehicle.
bdclose('ctrlKalmanNavigationExample');

9-63
9-64
 10

Model Analysis

• “Validating Models After Estimation” on page 10-2

• “Plotting Models in the App” on page 10-6
• “Simulating and Predicting Model Output” on page 10-8
• “Residual Analysis” on page 10-23
• “Impulse and Step Response Plots” on page 10-31
• “How to Plot Impulse and Step Response Using the App” on page 10-34
• “How to Plot Impulse and Step Response at the Command Line” on page 10-36
• “Frequency Response Plots” on page 10-38
• “How to Plot Bode Plots Using the App” on page 10-42
• “How to Plot Bode and Nyquist Plots at the Command Line” on page 10-44
• “Noise Spectrum Plots” on page 10-46
• “How to Plot the Noise Spectrum Using the App” on page 10-49
• “How to Plot the Noise Spectrum at the Command Line” on page 10-52
• “Pole and Zero Plots” on page 10-54
• “How to Plot Model Poles and Zeros Using the App” on page 10-58
• “How to Plot Poles and Zeros at the Command Line” on page 10-60
• “Analyzing MIMO Models” on page 10-61
• “Customizing Response Plots Using the Response Plots Property Editor” on page
10-67
• “Akaike's Criteria for Model Validation” on page 10-80
• “Computing Model Uncertainty” on page 10-83
• “Troubleshooting Models” on page 10-86
• “Next Steps After Getting an Accurate Model” on page 10-91
10 Model Analysis

Validating Models After Estimation

In this section...
“Ways to Validate Models” on page 10-2
“Data for Model Validation” on page 10-3
“Supported Model Plots” on page 10-3
“Definition of Confidence Interval for Specific Model Plots” on page 10-4

Ways to Validate Models

You can use the following approaches to validate models:

• Comparing simulated or predicted model output to measured output.

See “Simulating and Predicting Model Output” on page 10-8.

To simulate identified models in the Simulink environment, see “Simulating

Identified Model Output in Simulink”.
• Analyzing autocorrelation and cross-correlation of the residuals with input.

See “Residual Analysis” on page 10-23.

• Analyzing model response. For more information, see the following:

• “Impulse and Step Response Plots” on page 10-31

• “Frequency Response Plots” on page 10-38

For information about the response of the noise model, see “Noise Spectrum Plots” on
page 10-46.
• Plotting the poles and zeros of the linear parametric model.

For more information, see “Pole and Zero Plots” on page 10-54.
• Comparing the response of nonparametric models, such as impulse-, step-, and
frequency-response models, to parametric models, such as linear polynomial models,
state-space model, and nonlinear parametric models.

10-2
 Validating Models After Estimation

Note: Do not use this comparison when feedback is present in the system because
feedback makes nonparametric models unreliable. To test if feedback is present in the
system, use the advice command on the data.
• Compare models using Akaike Information Criterion or Akaike Final Prediction
Error.

For more information, see the aic and fpe reference page.
• Plotting linear and nonlinear blocks of Hammerstein-Wiener and nonlinear ARX
models.

Displaying confidence intervals on supported plots helps you assess the uncertainty of
model parameters. For more information, see “Computing Model Uncertainty” on page
10-83.

Data for Model Validation

For plots that compare model response to measured response and perform residual
analysis, you designate two types of data sets: one for estimating the models (estimation
data), and the other for validating the models (validation data). Although you can
designate the same data set to be used for estimating and validating the model, you risk
over-fitting your data. When you validate a model using an independent data set, this
process is called cross-validation.

Note: Validation data should be the same in frequency content as the estimation data. If
you detrended the estimation data, you must remove the same trend from the validation
data. For more information about detrending, see “Handling Offsets and Trends in Data”
on page 2-108.

Supported Model Plots

The following table summarizes the types of supported model plots.

Plot Type Supported Models Learn More

Model Output All linear and nonlinear “Simulating and Predicting
models Model Output” on page
10-8

10-3
10 Model Analysis

Plot Type Supported Models Learn More

Residual Analysis All linear and nonlinear “Residual Analysis” on page
models 10-23
Transient Response • All linear parametric “Impulse and Step Response
models Plots” on page 10-31
• Correlation analysis
(nonparametric) models
• For nonlinear models,
only step response.
Frequency Response All linear models “Frequency Response Plots”
on page 10-38
Noise Spectrum • All linear parametric “Noise Spectrum Plots” on
models page 10-46
• Spectral analysis
(nonparametric) models
Poles and Zeros All linear parametric “Pole and Zero Plots” on page
models 10-54
Nonlinear ARX Nonlinear ARX models only Nonlinear ARX Plots
Hammerstein-Wiener Hammerstein-Wiener Hammerstein-Wiener Plots
models only

Definition of Confidence Interval for Specific Model Plots

You can display the confidence interval on the following plot types:

Plot Type Confidence Interval Corresponds to More Information on Displaying

the Range of ... Confidence Interval
Simulated Output values with a specific Model Output Plots
and probability of being the actual
Predicted output of the system.
Output
Residuals Residual values with a specific Residuals Plots
probability of being statistically
insignificant for the system.

10-4
 Validating Models After Estimation

Plot Type Confidence Interval Corresponds to More Information on Displaying

the Range of ... Confidence Interval
Impulse Response values with a specific Impulse and Step Plots
and Step probability of being the actual
response of the system.
Frequency Response values with a specific Frequency Response Plots
Response probability of being the actual
response of the system.
Noise Power-spectrum values with Noise Spectrum Plots
Spectrum a specific probability of being
the actual noise spectrum of the
system.
Poles and Pole or zero values with a specific Pole-Zero Plots
Zeros probability of being the actual
pole or zero of the system.

10-5
10 Model Analysis

Plotting Models in the App

To create one or more plots of your models, select the corresponding check box in the
Model Views area of the System Identification app. An active model icon has a thick
line in the icon, while an inactive model has a thin line. Only active models appear on the
selected plots.

To include or exclude a model on a plot, click the corresponding icon in the System
Identification app. Clicking the model icon updates any plots that are currently open.

For example, in the following figure, Model output is selected. In this case, the models
n4s4 is not included on the plot because only arx441 is active.

Active model

Inactive model

Plots the model

output of active
models.

Plots Include Only Active Models

To close a plot, clear the corresponding check box in the System Identification app.

10-6
 Plotting Models in the App

Tip To get information about a specific plot, select a help topic from the Help menu in the
plot window.

For general information about working with plots in the System Identification Toolbox
product , see “Working with Plots” on page 14-11.

10-7
10 Model Analysis

Simulating and Predicting Model Output

In this section...
“Why Simulate or Predict Model Output” on page 10-8
“Definition: Simulation and Prediction” on page 10-9
“Simulation and Prediction in the App” on page 10-11
“Simulation and Prediction at the Command Line” on page 10-15
“Compare Simulated Output with Measured Data” on page 10-17
“Simulate Model Output with Noise” on page 10-18
“Simulate a Continuous-Time State-Space Model” on page 10-19
“Predict Using Time-Series Model” on page 10-20

Why Simulate or Predict Model Output

You primarily use a model is to simulate its output, i.e., calculate the output (y(t)) for
given input values. You can also predict model output, i.e., compute a qualified guess of
future output values based on past observations of system’s inputs and outputs. For more
information, see “Definition: Simulation and Prediction” on page 10-9.

You also validate linear parametric models and nonlinear models by checking how well
the simulated or predicted output of the model matches the measured output. You can
use either time or frequency domain data for simulation or prediction. For frequency
domain data, the simulation and prediction results are products of the Fourier transform
of the input and frequency function of the model. For more information, see “Simulation
and Prediction in the App” on page 10-11 and “Simulation and Prediction at the
Command Line” on page 10-15.

Simulation provides a better validation test for the model than prediction. However, how
you validate the model output should match how you plan to use the model. For example,
if you plan to use your model for control design, you can validate the model by predicting
its response over a time horizon that represents the dominating time constants of the
model.

Related Examples
“Compare Simulated Output with Measured Data” on page 10-17

10-8
 Related Examples

“Simulate Model Output with Noise” on page 10-18

“Simulate a Continuous-Time State-Space Model” on page 10-19

“Predict Using Time-Series Model” on page 10-20

Definition: Simulation and Prediction

Simulation means computing the model response using input data and initial conditions.
The time samples of the model response match the time samples of the input data used
for simulation.

For a continuous-time system, simulation means solving a differential equation. For a

discrete-time system, simulation means directly applying the model equations.

For example, consider a dynamic model described by a first-order difference equation

that uses a sampling interval of 1 second:

y(t) + ay(t–1) = bu(t–1),

where y is the output and u is the input. For parameter values a = –0.9 and b = 1.5, the
equation becomes:

y(t) – 0.9y(t–1) = 1.5u(t–1).

Suppose you want to compute the values y(1), y(2), y(3),... for given input values u(0) = 2,
u(1) = 1, u(2) = 4,...Here, y(1) is the value of output at the first sampling instant. Using
initial condition of y(0) = 0, the values of y(t) for times t = 1, 2 and 3 can be computed as:

y(1) = 0.9y(0) + 1.5u(0) = 0.9*0 + 1.5*2 = 3

y(2) = 0.9y(1) + 1.5u(1) = 0.9*3 + 1.5*1 = 4.2

y(3) = 0.9y(2) + 1.5u(2) = 0.9*4.2 + 1.5*4 = 9.78

...

Prediction forecasts the model response k steps ahead into the future using the current
and past values of measured input and output values. k is called the prediction horizon,
and corresponds to predicting output at time kTs, where Ts is the sampling interval.

For example, suppose you use sensors to measure the input signal u(t) and output signal
y(t) of the physical system, described in the previous first-order equation. At the tenth

10-9
10 Model Analysis

sampling instant (t = 10), the output y(10) is 16 mm and the corresponding input u(10)
is 12 N. Now, you want to predict the value of the output at the future time t = 11. Using
the previous equation:

y(11) = 0.9y(10) + 1.5u(10)

Hence, the predicted value of future output y(11) at time t = 10 is:

y(11) = 0.9*16 + 1.5*12 = 32.4

In general, to predict the model response k steps into the future (k≥1) from the current
time t, you should know the inputs up to time t+k and outputs up to time t:

yp(t+k) = f(u(t+k),u(t+k–1),...,u(t),u(t–1),...,u(0)
                y(t),y(t–1),y(t–2),...,y(0))

u(0) and y(0) are the initial states. f() represents the predictor, which is a dynamic
model whose form depends on the model structure. For example, the one-step-ahead
predictor yp of the model y(t) + ay(t–1) = bu(t) is:

yp(t+1) = –ay(t) + bu(t+1)

The difference between prediction and simulation is that in prediction, the past values
of outputs used for calculation are measured values while in simulation the outputs are
themselves a result of calculation using inputs and initial conditions.

The way information in past outputs is used depends on the disturbance model H of
1
the model. For the previous dynamic model, H ( z) = 1 + az-1 . In models of Output-Error
(OE) structure (H(z) = 1), there is no information in past outputs that can be used for
predicting future output values. In this case, predictions and simulations coincide. For
state-space models (idss), output-error structure corresponds to models with K=0. For
polynomial models (idpoly), this corresponds to models with polynomials a=c=d=1.

Note: Prediction with k=∞ means that no previous outputs are used in the computation
and prediction returns the same result as simulation.

Both simulation and prediction require initial conditions, which correspond to the states
of the model at the beginning of the simulation or prediction.

10-10
 Related Examples

Tip If you do not know the initial conditions and have input and output measurements
available, you can estimate the initial condition using this toolbox.

Simulation and Prediction in the App

• “How to Plot Simulated and Predicted Model Output” on page 10-11
• “Interpreting the Model Output Plot” on page 10-11
• “Changing Model Output Plot Settings” on page 10-13
• “Definition: Confidence Interval” on page 10-14

How to Plot Simulated and Predicted Model Output

To create a model output plot for parametric linear and nonlinear models in the System
Identification app, select the Model output check box in the Model Views area. By
default, this operation estimates the initial states from the data and plots the output of
selected models for comparison.

To include or exclude a model on the plot, click the corresponding model icon in the
System Identification app. Active models display a thick line inside the Model Board
icon.

To learn how to interpret the model output plot, see “Interpreting the Model Output Plot”
on page 10-11.

To change plot settings, see “Changing Model Output Plot Settings” on page 10-13.

For general information about creating and working with plots, see “Working with Plots”
on page 14-11.

Interpreting the Model Output Plot

The following figure shows a sample Model Output plot, created in the System
Identification app.

10-11
10 Model Analysis

The model output plot shows different information depending on the domain of the input-
output validation data, as follows:

• For time-domain validation data, the plot shows simulated or predicted model output.
• For frequency-domain data, the plot shows the amplitude of the model response to
the frequency-domain input signal. The model response is equal to the product of the
Fourier transform of the input and the model's frequency function.
• For frequency-response data, the plot shows the amplitude of the model frequency
response.

For linear models, you can estimate a model using time-domain data, and then validate
the model using frequency domain data. For nonlinear models, you can only use time-
domain data for both estimation and validation.

The right side of the plot displays the percentage of the output that the model reproduces
(Best Fit), computed using the following equation:

Ê y - yˆ ˆ
Best Fit = Á 1 - ˜˜ ¥ 100
Á y-y
Ë ¯

10-12
 Related Examples

In this equation, y is the measured output, ŷ is the simulated or predicted model output,
and y is the mean of y. 100% corresponds to a perfect fit, and 0% indicates that the fit is
no better than guessing the output to be a constant ( ŷ = y ).

Because of the definition of Best Fit, it is possible for this value to be negative. A
negative best fit is worse than 0% and can occur for the following reasons:

• The estimation algorithm failed to converge.

• The model was not estimated by minimizing y - yˆ . Best Fit can be negative when
you minimized 1-step-ahead prediction during the estimation, but validate using the
simulated output ŷ .

• The validation data set was not preprocessed in the same way as the estimation data
set.

Changing Model Output Plot Settings

The following table summarizes the Model Output plot settings.

Model Output Plot Settings

Action Command
Display confidence intervals. • To display the dashed lines on either
side of the nominal model curve,
select Options > Show confidence
Note: Confidence intervals are only intervals. Select this option again to
available for simulated model output hide the confidence intervals.
of linear models. Confidence internal
• To change the confidence value, select
are not available for nonlinear ARX and
Options > Set % confidence level,
Hammerstein-Wiener models.
and choose a value from the list.
• To enter your own confidence level,
See “Definition: Confidence Interval” on select Options > Set confidence
page 10-14. level > Other. Enter the value as a
probability (between 0 and 1) or as the
number of standard deviations of a
Gaussian distribution.

10-13
10 Model Analysis

Action Command
Change between simulated output or • Select Options > Simulated output
predicted output. or Options > k step ahead predicted
output.
• To change the prediction horizon, select
Note: Prediction is only available for time-
Options > Set prediction horizon,
domain validation data.
and select the number of samples.
• To enter your own prediction horizon,
select Options > Set prediction
horizon > Other. Enter the value in
terms of the number of samples.
Display the actual output values (Signal Select Options > Signal plot or Options
plot), or the difference between model > Error plot.
output and measured output (Error plot).
(Time-domain validation data only) Select Options > Customized time
Set the time range for model output and span for fit and enter the minimum and
the time interval for which the Best Fit maximum time values. For example:
value is computed.
[1 20]
(Multiple-output system only) Select the output by name in the Channel
Select a different output. menu.

Definition: Confidence Interval

The confidence interval corresponds to the range of output values with a specific
probability of being the actual output of the system. The toolbox uses the estimated
uncertainty in the model parameters to calculate confidence intervals and assumes the
estimates have a Gaussian distribution.

For example, for a 95% confidence interval, the region around the nominal curve
represents the range of values that have a 95% probability of being the true system
response. You can specify the confidence interval as a probability (between 0 and 1) or as
the number of standard deviations of a Gaussian distribution. For example, a probability
of 0.99 (99%) corresponds to 2.58 standard deviations.

Note: The calculation of the confidence interval assumes that the model sufficiently
describes the system dynamics and the model residuals pass independence tests.

10-14
 Related Examples

In the app, you can display a confidence interval on the plot to gain insight into the
quality of a linear model. To learn how to show or hide confidence interval, see “Changing
Model Output Plot Settings” on page 10-13.

Simulation and Prediction at the Command Line

• “Summary of Simulation and Prediction Commands” on page 10-15
• “Initial States in Simulation and Prediction” on page 10-16

Summary of Simulation and Prediction Commands

Note: If you estimated a linear model from detrended data and want to simulate or
predict the output at the original operation conditions, use the retrend command to the
simulated or predicted output.

Command Description Example

compare Use this command for To plot five-step-ahead predicted
model validation to output of the model mod against
determine how closely the the validation data data, use the
simulated model response following command:
matches the measured
output signal . compare(data,mod,5)

Plots simulated or
predicted output of Note: Omitting the third argument
one or more models on assumes an infinite horizon and
top of the measured results in the comparison of the
output. You should use an simulated response to the input data.
independent validation
data set as input to the
model.
sim Simulate and plot the To simulate the response of the model
model output only. model using input data data, use the
following command:
sim(model,data)

10-15
10 Model Analysis

Command Description Example

predict Predict and plot the model To perform one-step-ahead prediction
output only. of the response for the model model
and input data data, use the
following command:
predict(model,data,1)

Use the following syntax to compute

k-step-ahead prediction of the output
signal using model m:

yhat = predict(m,[y u],k)

Note that predict computes the

prediction results only over the time
range of data. It does not perform
any forecasting of results beyond the
available data range.
forecast Forecast a time series into To forecast the value of a time series
the future. in an arbitrary number of steps
into the future, use the following
command:
forecast(model,past_data,K)

Here, model is a time series model,

past_data is a record of the observed
values of the time series and K is the
forecasting horizon.

Initial States in Simulation and Prediction

The process of computing simulated and predicted responses over a time range starts by
using the initial conditions to compute the first few output values. sim, forecast and
predict commands provide defaults for handling initial conditions.

Simulation: Default initial conditions are zero for polynomial (idpoly), process
(idproc) and transfer-function (idtf) models . For state-space (idss) and grey-
box (idgrey) models, the default initial conditions are the internal model initial

10-16
 Related Examples

states (model property x0). You can specify other initial conditions using the
InitialCondition simulation option (see simOptions).

Use the compare command to validate models by simulation because its algorithm
estimates the initial states of a model to optimize the model fit to a given data set.

If you use sim, the simulated and the measured responses might differ when the initial
conditions of the estimated model and the system that measured the validation data set
differ—especially at the beginning of the response. To minimize this difference, estimate
the initial state values from the data using the findstates command and specify these
initial states as input arguments to the sim command. For example, to compute the
initial states that optimize the fit of the model m to the output data in z:

% Estimate the initial states

X0est = findstates(m,z);
% Simulate the response using estimated initial states
opt = simOptions('InitialCondition',X0est);
sim(m,z.InputData,opt)

See Also: sim (for linear models), sim(idnlarx), sim(idnlgrey), sim(idnlhw)

Prediction: Default initial conditions depend on the type of model. You can specify other
initial conditions using the InitialCondition option (see predictOptions). For
example, to compute the initial states that optimize the 1-step-ahead predicted response
of the model m to the output data z:

opt = predictOptions('InitialCondition','estimate');
[Yp,X0est] = predict(m,z,1,opt);

This command returns the estimated initial states as the output argument X0est. For
information about other ways to specify initials states, see the predict reference page
for the corresponding model type.

See Also: predict

Compare Simulated Output with Measured Data

This example shows how to validate an estimated model by comparing the simulated
model output with measured data.

Create estimation and validation data.

10-17
10 Model Analysis

load iddata1;
ze = z1(1:150);
zv = z1(151:300);

Estimate an ARMAX model.

m = armax(ze,[2 3 1 0]);

Compare simulated model output with measured data.

compare(zv,m);

Simulate Model Output with Noise

10-18
 Related Examples

This example shows how you can create input data and a model, and then use the data
and the model to simulate output data.

In this example, you create the following ARMAX model with Gaussian noise e:

Then, you simulate output data with random binary input u.

Create an ARMAX model.

m_armax = idpoly([1 -1.5 0.7],[0 1 0.5],[1 -1 0.2]);

Create a random binary input.

u = idinput(400,'rbs',[0 0.3]);

Simulate the output data.

opt = simOptions('AddNoise',true);
y = sim(m_armax,u,opt);

The 'AddNoise' option specifies to include in the simulation the Gaussian noise e
present in the model. Set this option to false (default behavior) to simulate the noise-
free response to the input u , which is equivalent to setting e to zero.

Simulate a Continuous-Time State-Space Model

This example shows how to simulate a continuous-time state-space model using a
random binary input u and a sampling interval of 0.1 s.

Consider the following state-space model:

where e is Gaussian white noise with variance 7.

10-19
10 Model Analysis

Create a continuous-time state-space model.

A = [-1 1;-0.5 0]; B = [1; 0.5];

C = [1 0]; D = 0; K = [0.5;0.5];
% Ts = 0 indicates continuous time
model_ss = idss(A,B,C,D,K,'Ts',0,'NoiseVariance',7);

Create a random binary input.

u = idinput(400,'rbs',[0 0.3]);

Create an iddata object with empty output to represent just the input signal.

data = iddata([],u);
data.ts = 0.1;

Simulate the output using the model

opt = simOptions('AddNoise',true);
y=sim(model_ss,data,opt);

Predict Using Time-Series Model

This example shows how to evaluate how well a time-series model predicts the response
for a given prediction horizon.

In this example, y is the original series of monthly sales figures. You use the first half
of the measured data to estimate the time-series model and test the model's ability to
forecast sales six months ahead using the entire data set.

Simulate figures for sales growth every month before settling down.

rng('default');
t = (0:0.2:19)';
yd = 1-exp(-0.2*t) + randn(size(t))/50;
y = iddata(yd,[],0.2);
plot(y);

10-20
 Related Examples

Select the first half as estimation data.

y1 = y(1:48);

Estimate a second-order autoregressive model with noise integration using the data.
m = ar(y1, 2, 'yw', 'IntegrateNoise', true);

Compute 6-step ahead predicted output.

yhat = predict(m,y,6);

Forecast the response 10 steps beyond the available data's time range.
yfuture = forecast(m,y,10);

10-21
10 Model Analysis

Plot the measured, predicted and forecasted outputs.

plot(y,yhat,yfuture);
legend({'measured', 'predicted', 'forecasted'},'location','southeast');

10-22
 Residual Analysis

Residual Analysis
In this section...
“What Is Residual Analysis?” on page 10-23
“Supported Model Types” on page 10-24
“What Residual Plots Show for Different Data Domains” on page 10-24
“Displaying the Confidence Interval” on page 10-25
“How to Plot Residuals Using the App” on page 10-26
“How to Plot Residuals at the Command Line” on page 10-27
“Examine Model Residuals” on page 10-27

What Is Residual Analysis?

Residuals are differences between the one-step-predicted output from the model and the
measured output from the validation data set. Thus, residuals represent the portion of
the validation data not explained by the model.

Residual analysis consists of two tests: the whiteness test and the independence test.

According to the whiteness test criteria, a good model has the residual autocorrelation
function inside the confidence interval of the corresponding estimates, indicating that the
residuals are uncorrelated.

According to the independence test criteria, a good model has residuals uncorrelated
with past inputs. Evidence of correlation indicates that the model does not describe how
part of the output relates to the corresponding input. For example, a peak outside the
confidence interval for lag k means that the output y(t) that originates from the input u(t-
k) is not properly described by the model.

Your model should pass both the whiteness and the independence tests, except in the
following cases:

• For output-error (OE) models and when using instrumental-variable (IV) methods,
make sure that your model shows independence of e and u, and pay less attention to
the results of the whiteness of e.

In this case, the modeling focus is on the dynamics G and not the disturbance
properties H.

10-23
10 Model Analysis

• Correlation between residuals and input for negative lags, is not necessarily an
indication of an inaccurate model.

When current residuals at time t affect future input values, there might be feedback
in your system. In the case of feedback, concentrate on the positive lags in the cross-
correlation plot during model validation.

Supported Model Types

You can validate parametric linear and nonlinear models by checking the behavior of the
model residuals. For a description of residual analysis, see “What Residual Plots Show
for Different Data Domains” on page 10-24.

Note: Residual analysis plots are not available for frequency response (FRD) models.
For time-series models, you can only generate model-output plots for parametric models
using time-domain time-series (no input) measured data.

What Residual Plots Show for Different Data Domains

Residual analysis plots show different information depending on whether you use time-
domain or frequency-domain input-output validation data.

For time-domain validation data, the plot shows the following two axes:

• Autocorrelation function of the residuals for each output

• Cross-correlation between the input and the residuals for each input-output pair

Note: For time-series models, the residual analysis plot does not provide any input-
residual correlation plots.

For frequency-domain validation data, the plot shows the following two axes:

• Estimated power spectrum of the residuals for each output

• Transfer-function amplitude from the input to the residuals for each input-output pair

For linear models, you can estimate a model using time-domain data, and then validate
the model using frequency domain data. For nonlinear models, the System Identification
Toolbox product supports only time-domain data.

10-24
 Residual Analysis

The following figure shows a sample Residual Analysis plot, created in the System
Identification app.

Displaying the Confidence Interval

The confidence interval corresponds to the range of residual values with a specific
probability of being statistically insignificant for the system. The toolbox uses the
estimated uncertainty in the model parameters to calculate confidence intervals and
assumes the estimates have a Gaussian distribution.

For example, for a 95% confidence interval, the region around zero represents the range
of residual values that have a 95% probability of being statistically insignificant. You
can specify the confidence interval as a probability (between 0 and 1) or as the number of
standard deviations of a Gaussian distribution. For example, a probability of 0.99 (99%)
corresponds to 2.58 standard deviations.

You can display a confidence interval on the plot in the app to gain insight into the
quality of the model. To learn how to show or hide confidence interval, see the description
of the plot settings in “How to Plot Residuals Using the App” on page 10-26.

10-25
10 Model Analysis

Note: If you are working in the System Identification app, you can specify a custom
confidence interval. If you are using the resid command, the confidence interval is fixed
at 99%.

How to Plot Residuals Using the App

To create a residual analysis plot for parametric linear and nonlinear models in the
System Identification app, select the Model resids check box in the Model Views area.
For general information about creating and working with plots, see “Working with Plots”
on page 14-11.

To include or exclude a model on the plot, click the corresponding model icon in the
System Identification app. Active models display a thick line inside the Model Board
icon.

The following table summarizes the Residual Analysis plot settings.

Residual Analysis Plot Settings

Action Command
Display confidence intervals • To display the dashed lines on either side of the
around zero. nominal model curve, select Options > Show
confidence intervals. Select this option again to
hide the confidence intervals.
Note: Confidence internal are
• To change the confidence value, select Options >
not available for nonlinear
Set % confidence level and choose a value from
ARX and Hammerstein-Wiener
the list.
models.
• To enter your own confidence level, select Options
> Set confidence level > Other. Enter the
value as a probability (between 0 and 1) or as
the number of standard deviations of a Gaussian
distribution.
Change the number of lags • Select Options > Number of lags and choose the
(data samples) for which to value from the list.
compute autocorellation and • To enter your own lag value, select Options > Set
cross-correlation functions. confidence level > Other. Enter the value as the
number of data samples.

10-26
 Residual Analysis

Action Command
Note: For frequency-domain
validation data, increasing
the number of lags increases
the frequency resolution of
the residual spectrum and the
transfer function.
(Multiple-output system only) Select the input-output by name in the Channel
Select a different input-output menu.
pair.

How to Plot Residuals at the Command Line

The following table summarizes commands that generate residual-analysis plots for
linear and nonlinear models. For detailed information about this command, see the
corresponding reference page.

Note: Apply pe and resid to one model at a time.

Command Description Example

pe Computes and plots model To plot the prediction errors
prediction errors. for the model model using
data data, type the following
command:
pe(model,data)
resid Performs whiteness and To plot residual correlations
independence tests on model for the model model using
residuals, or prediction errors. data data, type the following
Uses validation data input as command:
model input.
resid(model,data)

Examine Model Residuals

This example shows how you can use residual analysis to evaluate model quality.

10-27
10 Model Analysis

Creating Residual Plots

1 To load the sample System Identification app session that contains estimated
models, type the following command in the MATLAB Command Window:

systemIdentification('dryer2_linear_models')
2 To generate a residual analysis plot, select the Model resids check box in the
System Identification app.

This opens an empty plot.

3 In the System Identification app window, click each model icon to display it on the
Residual Analysis plot.

Note: For the nonparametric models, imp and spad, residual analysis plots are not
available.

10-28
 Residual Analysis

Description of the Residual Plot Axes

The top axes show the autocorrelation of residuals for the output (whiteness test). The
horizontal scale is the number of lags, which is the time difference (in samples) between
the signals at which the correlation is estimated. The horizontal dashed lines on the
plot represent the confidence interval of the corresponding estimates. Any fluctuations
within the confidence interval are considered to be insignificant. Four of the models,
arxqs, n4s3, arx223 and amx2222, produce residuals that enter outside the confidence
interval. A good model should have a residual autocorrelation function within the
confidence interval, indicating that the residuals are uncorrelated.

The bottom axes show the cross-correlation of the residuals with the input. A good model
should have residuals uncorrelated with past inputs (independence test). Evidence of
correlation indicates that the model does not describe how the output is formed from the
corresponding input. For example, when there is a peak outside the confidence interval
for lag k, this means that the contribution to the output y(t) that originates from the
input u(t-k) is not properly described by the model. The models arxqs and amx2222
extend beyond the confidence interval and do not perform as well as the other models.

Validating Models Using Analyzing Residuals

To remove models with poor performance from the Residual Analysis plot, click the
model icons arxqs, n4s3, arx223, and amx2222 in the System Identification app.

The Residual Analysis plot now includes only the three models that pass the residual
tests: arx692, n4s6, and amx3322.

10-29
10 Model Analysis

The plots for these models fall within the confidence intervals. Thus, when choosing the
best model among several estimated models, it is reasonable to pick amx3322 because it
is a simpler, low-order model.

10-30
 Impulse and Step Response Plots

Impulse and Step Response Plots

In this section...
“Supported Models” on page 10-31
“How Transient Response Helps to Validate Models” on page 10-31
“What Does a Transient Response Plot Show?” on page 10-32
“Displaying the Confidence Interval” on page 10-33

Supported Models
You can plot the simulated response of a model using impulse and step signals as the
input for all linear parametric models and correlation analysis (nonparametric) models.

You can also create step-response plots for nonlinear models. These step and
impulse response plots, also called transient response plots, provide insight into the
characteristics of model dynamics, including peak response and settling time.

Note: For frequency-response models, impulse- and step-response plots are not available.
For nonlinear models, only step-response plots are available.

Examples
“How to Plot Impulse and Step Response Using the App” on page 10-34

“How to Plot Impulse and Step Response at the Command Line” on page 10-36

How Transient Response Helps to Validate Models

Transient response plots provide insight into the basic dynamic properties of the model,
such as response times, static gains, and delays.

Transient response plots also help you validate how well a linear parametric model,
such as a linear ARX model or a state-space model, captures the dynamics. For example,

10-31
10 Model Analysis

you can estimate an impulse or step response from the data using correlation analysis
(nonparametric model), and then plot the correlation analysis result on top of the
transient responses of the parametric models.

Because nonparametric and parametric models are derived using different algorithms,
agreement between these models increases confidence in the parametric model results.

What Does a Transient Response Plot Show?

Transient response plots show the value of the impulse or step response on the vertical
axis. The horizontal axis is in units of time you specified for the data used to estimate the
model.

The impulse response of a dynamic model is the output signal that results when the
input is an impulse. That is, u(t) is zero for all values of t except at t=0, where u(0)=1. In
the following difference equation, you can compute the impulse response by setting y(-
T)=y(-2T)=0, u(0)=1, and u(t>0)=0.

y( t) - 1 .5 y(t - T ) + 0.7 y( t - 2 T) =
0 .9u(t) + 0.5u(t - T )

The step response is the output signal that results from a step input, where u(t<0)=0 and
u(t>0)=1.

If your model includes a noise model, you can display the transient response of the noise
model associated with each output channel. For more information about how to display
the transient response of the noise model, see “How to Plot Impulse and Step Response
Using the App” on page 10-34.

The following figure shows a sample Transient Response plot, created in the System
Identification app.

10-32
 Examples

Displaying the Confidence Interval

In addition to the transient-response curve, you can display a confidence interval on the
plot. To learn how to show or hide confidence interval, see the description of the plot
settings in “How to Plot Impulse and Step Response Using the App” on page 10-34.

The confidence interval corresponds to the range of response values with a specific
probability of being the actual response of the system. The toolbox uses the estimated
uncertainty in the model parameters to calculate confidence intervals and assumes the
estimates have a Gaussian distribution.

For example, for a 95% confidence interval, the region around the nominal curve
represents the range where there is a 95% chance that it contains the true system
response. You can specify the confidence interval as a probability (between 0 and 1) or as
the number of standard deviations of a Gaussian distribution. For example, a probability
of 0.99 (99%) corresponds to 2.58 standard deviations.

Note: The calculation of the confidence interval assumes that the model sufficiently
describes the system dynamics and the model residuals pass independence tests.

10-33
10 Model Analysis

How to Plot Impulse and Step Response Using the App

To create a transient analysis plot in the System Identification app, select the Transient
resp check box in the Model Views area. For general information about creating and
working with plots, see “Working with Plots” on page 14-11.

To include or exclude a model on the plot, click the corresponding model icon in the
System Identification app. Active models display a thick line inside the Model Board
icon.

The following table summarizes the Transient Response plot settings.

Transient Response Plot Settings

Action Command
Display step response for linear or Select Options > Step response.
nonlinear model.
Display impulse response for linear Select Options > Impulse response.
model.

Note: Not available for nonlinear models.

Display the confidence interval. • To display the dashed lines on either side
of the nominal model curve, select Options
> Show confidence intervals. Select this
Note: Only available for linear models. option again to hide the confidence intervals.
• To change the confidence value, select
Options > Set % confidence level, and
choose a value from the list.
• To enter your own confidence level, select
Options > Set confidence level > Other.
Enter the value as a probability (between
0 and 1) or as the number of standard
deviations of a Gaussian distribution.
Change time span over which the • Select Options > Time span (time units),
impulse or step response is calculated. and choose a new time span in units of time
For a scalar time span T, the resulting you specified for the model.
response is plotted from -T/4 to T. • To enter your own time span, select Options
> Time span (time units) > Other, and
enter the total response duration.

10-34
 More About

Action Command
Note: To change the time span of • To use the time span based on model
models you estimated using correlation dynamics, type [] or default.
analysis models, select Estimate >
Correlation models and reestimate The default time span is computed based
the model using a new time span. on the model dynamics and might be
different for different models. For nonlinear
models, the default time span is 10.

Toggle between line plot or stem plot. Select Style > Line plot or Style > Stem plot.

Tip Use a stem plot for displaying

impulse response.
(Multiple-output system only) Select the output by name in the Channel
Select an input-output pair to view the menu.
noise spectrum corresponding to those
channels. If the plotted models include a noise model, you
can display the transient response properties
associated with each output channel. The name
of the channel has the format e@OutputName,
where OutputName is the name of the output
channel corresponding to the noise model.
(Step response for nonlinear models Select Options > Step Size, and then chose
only) from two options:
Set level of the input step.
• 0–>1 sets the lower level to 0 and the upper
level to 1.
Note: For multiple-input models, the • Other opens the Step Level dialog box,
input-step level applies only to the where you enter the values for the lower and
input channel you selected to display upper level values.
in the plot.

More About
“Impulse and Step Response Plots” on page 10-31

10-35
10 Model Analysis

How to Plot Impulse and Step Response at the Command Line

You can plot impulse- and step-response plots using the impulseplot and stepplot
commands, respectively. If you want to fetch the response data, use impulse and step
instead.

All plot commands have the same basic syntax, as follows:

• To plot one model, use the syntax command(model).

• To plot several models, use the syntax command(model1,model2,...,modelN).

In this case, command represents any of the plotting commands.

To display confidence intervals for a specified number of standard deviations, use the
following syntax:
h = impulseplot(model);
showConfidence(h,sd);

where h is the plot handle returned by impulseplot. You could also use the plot
handle returned by stepplot. sd is the number of standard deviations of a Gaussian
distribution. For example, a confidence value of 99% for the nominal model curve
corresponds to 2.58 standard deviations.

Alternatively, you can turn on the confidence region view interactively by right-clicking
on the plot and selecting Characteristics > Confidence Region. Use the plot property
editor to specify the number of standard deviations.

The following table summarizes commands that generate impulse- and step-response
plots. For detailed information about each command, see the corresponding reference
page.

Command Description Example

impulse,impulseplotPlot impulse response for To plot the impulse response of
idpoly, idproc, idtf, the model sys, type the following
idss, and idgrey model command:
objects.
impulse(sys)

Note: Does not support

nonlinear models.

10-36
 More About

Command Description Example

step,stepplot Plots the step response of To plot the step response of the
all linear and nonlinear model sys, type the following
models. command:
step(sys)

To specify the step level offset (u0)

and amplitude (A) for a model:

opt = stepDataOptions;
opt.InputOffset = u0;
opt.StepAmplitude = A;

step(sys,opt)

More About
“Impulse and Step Response Plots” on page 10-31

10-37
10 Model Analysis

Frequency Response Plots

In this section...
“What Is Frequency Response?” on page 10-38
“How Frequency Response Helps to Validate Models” on page 10-39
“What Does a Frequency-Response Plot Show?” on page 10-39
“Displaying the Confidence Interval” on page 10-40

What Is Frequency Response?

Frequency response plots show the complex values of a transfer function as a function of
frequency.

In the case of linear dynamic systems, the transfer function G is essentially an operator
that takes the input u of a linear system to the output y:

y = Gu

For a continuous-time system, the transfer function relates the Laplace transforms of the
input U(s) and output Y(s):

Y ( s) = G ( s)U ( s)

In this case, the frequency function G(iw) is the transfer function evaluated on the
imaginary axis s=iw.

For a discrete-time system sampled with a time interval T, the transfer function relates
the Z-transforms of the input U(z) and output Y(z):

Y ( z) = G ( z)U ( z)

In this case, the frequency function G(eiwT) is the transfer function G(z) evaluated on
the unit circle. The argument of the frequency function G(eiwT) is scaled by the sampling
interval T to make the frequency function periodic with the sampling frequency 2 p T .

10-38
 Examples

Examples
“How to Plot Bode Plots Using the App” on page 10-42

“How to Plot Bode and Nyquist Plots at the Command Line” on page 10-44

How Frequency Response Helps to Validate Models

You can plot the frequency response of a model to gain insight into the characteristics
of linear model dynamics, including the frequency of the peak response and stability
margins. Frequency-response plots are available for all linear models.

Note: Frequency-response plots are not available for nonlinear models. In addition,
Nyquist plots do not support time-series models that have no input.

The frequency response of a linear dynamic model describes how the model reacts to
sinusoidal inputs. If the input u(t) is a sinusoid of a certain frequency, then the output
y(t) is also a sinusoid of the same frequency. However, the magnitude of the response is
different from the magnitude of the input signal, and the phase of the response is shifted
relative to the input signal.

Frequency response plots provide insight into linear systems dynamics, such as
frequency-dependent gains, resonances, and phase shifts. Frequency response plots also
contain information about controller requirements and achievable bandwidths. Finally,
frequency response plots can also help you validate how well a linear parametric model,
such as a linear ARX model or a state-space model, captures the dynamics.

One example of how frequency-response plots help validate other models is that you can
estimate a frequency response from the data using spectral analysis (nonparametric
model), and then plot the spectral analysis result on top of the frequency response of
the parametric models. Because nonparametric and parametric models are derived
using different algorithms, agreement between these models increases confidence in the
parametric model results.

What Does a Frequency-Response Plot Show?

System Identification app supports the following types of frequency-response plots for
linear parametric models, linear state-space models, and nonparametric frequency-
response models:

10-39
10 Model Analysis

• Bode plot of the model response. A Bode plot consists of two plots. The top plot shows
the magnitude G by which the transfer function G magnifies the amplitude of the
sinusoidal input. The bottom plot shows the phase j = arg G by which the transfer
function shifts the input. The input to the system is a sinusoid, and the output is also
a sinusoid with the same frequency.
• Plot of the disturbance model, called noise spectrum. This plot is the same as a Bode
plot of the model response, but it shows the output power spectrum of the noise model
instead. For more information, see “Noise Spectrum Plots” on page 10-46.
• (Only in the MATLAB Command Window)
Nyquist plot. Plots the imaginary versus the real part of the transfer function.

The following figure shows a sample Bode plot of the model dynamics, created in the
System Identification app.

Displaying the Confidence Interval

In addition to the frequency-response curve, you can display a confidence interval on
the plot. To learn how to show or hide confidence interval, see the description of the plot
settings in “How to Plot Bode Plots Using the App” on page 10-42

10-40
 Examples

The confidence interval corresponds to the range of response values with a specific
probability of being the actual response of the system. The toolbox uses the estimated
uncertainty in the model parameters to calculate confidence intervals and assumes the
estimates have a Gaussian distribution.

For example, for a 95% confidence interval, the region around the nominal curve
represents the range where there is a 95% chance that it contains the true system
response. You can specify the confidence interval as a probability (between 0 and 1) or as
the number of standard deviations of a Gaussian distribution. For example, a probability
of 0.99 (99%) corresponds to 2.58 standard deviations.

10-41
10 Model Analysis

How to Plot Bode Plots Using the App

To create a frequency-response plot for linear models in the System Identification app,
select the Frequency resp check box in the Model Views area. For general information
about creating and working with plots, see “Working with Plots” on page 14-11.

To include or exclude a model on the plot, click the corresponding model icon in the
System Identification app. Active models display a thick line inside the Model Board
icon.

The following table summarizes the Frequency Function plot settings.

Frequency Function Plot Settings

Action Command
Display the confidence interval. • To display the dashed lines on either side of the
nominal model curve, select Options > Show
confidence intervals. Select this option again to
hide the confidence intervals.
• To change the confidence value, select Options >
Set % confidence level, and choose a value from
the list.
• To enter your own confidence level, select Options
> Set confidence level > Other. Enter the
value as a probability (between 0 and 1) or as
the number of standard deviations of a Gaussian
distribution.
Change the frequency values for Select Options > Frequency range and specify a
computing the noise spectrum. new frequency vector in units of rad/s.

The default frequency vector is Enter the frequency vector using any one of following
128 linearly distributed values, methods:
greater than zero and less
than or equal to the Nyquist • MATLAB expression, such as [1:100]*pi/100 or
frequency. logspace(-3,-1,200). Cannot contain variables
in the MATLAB workspace.
• Row vector of values, such as [1:.1:100]

10-42
 More About

Action Command
Note: To restore the default frequency vector, enter
[].
Change frequency units between Select Style > Frequency (Hz) or Style >
hertz and radians per second. Frequency (rad/s).
Change frequency scale between Select Style > Linear frequency scale or Style >
linear and logarithmic. Log frequency scale.
Change amplitude scale between Select Style > Linear amplitude scale or Style >
linear and logarithmic. Log amplitude scale.
(Multiple-output system only) Select the output by name in the Channel menu.
Select an input-output pair
to view the noise spectrum
corresponding to those channels.

Note: You cannot view cross

spectra between different
outputs.

More About
“Frequency Response Plots” on page 10-38

10-43
10 Model Analysis

How to Plot Bode and Nyquist Plots at the Command Line

You can plot Bode and Nyquist plots for linear models using the bode and nyquist
commands. If you want to customize the appearance of the plot, or turn on the confidence
region programmatically, use bodeplot, and nyquistplot instead.

All plot commands have the same basic syntax, as follows:

• To plot one model, use the syntax command(model).

• To plot several models, use the syntax command(model1,model2,...,modelN).

In this case, command represents any of the plotting commands.

To display confidence intervals for a specified number of standard deviations, use the
following syntax:

h=command(model);
showConfidence(h,sd)

where sd is the number of standard deviations of a Gaussian distribution and command

is bodeplotor nyquistplot. For example, a confidence value of 99% for the nominal
model curve corresponds to 2.58 standard deviations.

The following table summarizes commands that generate Bode and Nyquist plots for
linear models. For detailed information about each command and how to specify the
frequency values for computing the response, see the corresponding reference page.

Command Description Example

bode and bodeplot Plots the magnitude To create the bode plot of the model,
and phase of the sys, use the following command:
frequency response on
a logarithmic frequency bode(sys)
scale.

Note: Does not support

time-series models.
nyquist and Plots the imaginary To plot the frequency response of
nyquistplot versus real part of the the model, sys, use the following
transfer function. command:

10-44
 More About

Command Description Example

Note: Does not support nyquist(sys)
time-series models.

spectrum and Plots the disturbance To plot the output spectrum of

spectrumplot spectra of input-output a time series model, sys, with
models and output 1 standard deviation confidence
spectra of time series region, use the following command:
models.
showConfidence(spectrumplot(sys));

More About
“Frequency Response Plots” on page 10-38

10-45
10 Model Analysis

Noise Spectrum Plots

In this section...
“Supported Models” on page 10-46
“What Does a Noise Spectrum Plot Show?” on page 10-46
“Displaying the Confidence Interval” on page 10-47

Supported Models
When you estimate the noise model of your linear system, you can plot the spectrum of
the estimated noise model. Noise-spectrum plots are available for all linear parametric
models and spectral analysis (nonparametric) models.

Note: For nonlinear models and correlation analysis models, noise-spectrum plots are
not available. For time-series models, you can only generate noise-spectrum plots for
parametric and spectral-analysis models.

Examples
“How to Plot the Noise Spectrum Using the App” on page 10-49

“How to Plot the Noise Spectrum at the Command Line” on page 10-52

What Does a Noise Spectrum Plot Show?

The general equation of a linear dynamic system is given by:

y( t) = G ( z)u(t) + v(t)

In this equation, G is an operator that takes the input to the output and captures the
system dynamics, and v is the additive noise term. The toolbox treats the noise term as
filtered white noise, as follows:

v(t) = H ( z) e( t)

10-46
 Examples

where e(t) is a white-noise source with variance λ.

The toolbox computes both H and l during the estimation of the noise model and stores
these quantities as model properties. The H(z) operator represents the noise model.

Whereas the frequency-response plot shows the response of G, the noise-spectrum plot
shows the frequency-response of the noise model H.

For input-output models, the noise spectrum is given by the following equation:

2
( )
F v (w) = l H eiw

For time-series models (no input), the vertical axis of the noise-spectrum plot is the same
as the dynamic model spectrum. These axes are the same because there is no input for
time series and y = He .

Note: You can avoid estimating the noise model by selecting the Output-Error model
structure or by setting the DisturbanceModel property value to 'None' for a state
space model. If you choose to not estimate a noise model for your system, then H and the
noise spectrum amplitude are equal to 1 at all frequencies.

Displaying the Confidence Interval

In addition to the noise-spectrum curve, you can display a confidence interval on the plot.
To learn how to show or hide confidence interval, see the description of the plot settings
in “How to Plot the Noise Spectrum Using the App” on page 10-49.

The confidence interval corresponds to the range of power-spectrum values with a

specific probability of being the actual noise spectrum of the system. The toolbox uses
the estimated uncertainty in the model parameters to calculate confidence intervals and
assumes the estimates have a Gaussian distribution.

For example, for a 95% confidence interval, the region around the nominal curve
represents the range where there is a 95% chance that the true response belongs.. You
can specify the confidence interval as a probability (between 0 and 1) or as the number of
standard deviations of a Gaussian distribution. For example, a probability of 0.99 (99%)
corresponds to 2.58 standard deviations.

10-47
10 Model Analysis

Note: The calculation of the confidence interval assumes that the model sufficiently
describes the system dynamics and the model residuals pass independence tests.

10-48
 How to Plot the Noise Spectrum Using the App

How to Plot the Noise Spectrum Using the App

To create a noise spectrum plot for parametric linear models in the app, select the Noise
spectrum check box in the Model Views area. For general information about creating
and working with plots, see “Working with Plots” on page 14-11.

To include or exclude a model on the plot, click the corresponding model icon in the
System Identification app. Active models display a thick line inside the Model Board
icon.

The following figure shows a sample Noise Spectrum plot.

The following table summarizes the Noise Spectrum plot settings.

Noise Spectrum Plot Settings

Action Command
Display the confidence • To display the dashed lines on either side of the
interval. nominal model curve, select Options > Show
confidence intervals. Select this option again to hide
the confidence intervals.

10-49
10 Model Analysis

Action Command
• To change the confidence value, select Options > Set
% confidence level, and choose a value from the list.
• To enter your own confidence level, select Options >
Set confidence level > Other. Enter the value as
a probability (between 0 and 1) or as the number of
standard deviations of a Gaussian distribution.
Change the frequency Select Options > Frequency range and specify a new
values for computing the frequency vector in units of radians per second.
noise spectrum.
Enter the frequency vector using any one of following
The default frequency methods:
vector is 128 linearly
distributed values, greater • MATLAB expression, such as [1:100]*pi/100 or
than zero and less than logspace(-3,-1,200). Cannot contain variables in
or equal to the Nyquist the MATLAB workspace.
frequency. • Row vector of values, such as [1:.1:100]

Tip To restore the default frequency vector, enter [].

Change frequency units Select Style > Frequency (Hz) or Style > Frequency
between hertz and radians (rad/s).
per second.
Change frequency scale Select Style > Linear frequency scale or Style > Log
between linear and frequency scale.
logarithmic.
Change amplitude scale Select Style > Linear amplitude scale or Style > Log
between linear and amplitude scale.
logarithmic.
(Multiple-output system Select the output by name in the Channel menu.
only)
Select an input-output pair
to view the noise spectrum
corresponding to those
channels.

10-50
 More About

Action Command
Note: You cannot view cross
spectra between different
outputs.

More About
“Noise Spectrum Plots” on page 10-46

10-51
10 Model Analysis

How to Plot the Noise Spectrum at the Command Line

To plot the disturbance spectrum of an input-output model or the output spectrum of a
time series model, use spectrum. To customize such plots, or to turn on the confidence
region view programmatically for such plots, use spectrumplot instead.

To determine if your estimated noise model is good enough, you can compare the output
spectrum of the estimated noise-model H to the estimated output spectrum of v(t). To
compute v(t), which represents the actual noise term in the system, use the following
commands:
ysimulated = sim(m,data);
v = ymeasured-ysimulated;

ymeasured is data.y. v is the noise term v(t), as described in “What Does a Noise
Spectrum Plot Show?” on page 10-46 and corresponds to the difference between the
simulated response ysimulated and the actual response ymeasured.

To compute the frequency-response model of the actual noise, use spa:

V = spa(v);

The toolbox uses the following equation to compute the noise spectrum of the actual
noise:

•
F v (w) = Â Rv ( t ) e-iwt
t=-•

The covariance function Rv is given in terms of E, which denotes the mathematical

expectation, as follows:

Rv ( t ) = Ev ( t ) v ( t - t )

To compare the parametric noise-model H to the (nonparametric) frequency-response

estimate of the actual noise v(t), use spectrum:
spectrum(V,m)

If the parametric and the nonparametric estimates of the noise spectra are different,
then you might need a higher-order noise model.

10-52
 More About

More About
“Noise Spectrum Plots” on page 10-46

10-53
10 Model Analysis

Pole and Zero Plots

In this section...
“Supported Models” on page 10-54
“What Does a Pole-Zero Plot Show?” on page 10-54
“Reducing Model Order Using Pole-Zero Plots” on page 10-56
“Displaying the Confidence Interval” on page 10-56

Supported Models
You can create pole-zero plots of linear identified models. To study the poles and zeros of
the noise component of an input-output model or a time series model, use noise2meas to
first extract the noise model as an independent input-output model, whose inputs are the
noise channels of the original model.

Examples
“How to Plot Model Poles and Zeros Using the App” on page 10-58

“How to Plot Poles and Zeros at the Command Line” on page 10-60

What Does a Pole-Zero Plot Show?

The following figure shows a sample pole-zero plot of the model with confidence intervals.
x indicate poles and o indicate zeros.

10-54
 Examples

The general equation of a linear dynamic system is given by:

y( t) = G ( z)u(t) + v(t)

In this equation, G is an operator that takes the input to the output and captures the
system dynamics, and v is the additive noise term.

The poles of a linear system are the roots of the denominator of the transfer function
G. The poles have a direct influence on the dynamic properties of the system. The zeros
are the roots of the numerator of G. If you estimated a noise model H in addition to the
dynamic model G, you can also view the poles and zeros of the noise model.

Zeros and the poles are equivalent ways of describing the coefficients of a linear
difference equation, such as the ARX model. Poles are associated with the output side of
the difference equation, and zeros are associated with the input side of the equation. The
number of poles is equal to the number of sampling intervals between the most-delayed
and least-delayed output. The number of zeros) is equal to the number of sampling
intervals between the most-delayed and least-delayed input. For example, there two
poles and one zero in the following ARX model:

10-55
10 Model Analysis

y( t) - 1 .5 y(t - T ) + 0.7 y( t - 2 T) =
0 .9u(t) + 0.5u(t - T )

Reducing Model Order Using Pole-Zero Plots

You can use pole-zero plots to evaluate whether it might be useful to reduce model order.
When confidence intervals for a pole-zero pair overlap, this overlap indicates a possible
pole-zero cancelation.

For example, you can use the following syntax to plot a 1-standard-deviation confidence
interval around model poles and zeros.

showConfidence(iopzplot(model))

If poles and zeros overlap, try estimating a lower order model.

Always validate model output and residuals to see if the quality of the fit changes after
reducing model order. If the plot indicates pole-zero cancellations, but reducing model
order degrades the fit, then the extra poles probably describe noise. In this case, you
can choose a different model structure that decouples system dynamics and noise. For
example, try ARMAX, Output-Error, or Box-Jenkins polynomial model structures with
an A or F polynomial of an order equal to that of the number of uncanceled poles. For
more information about estimating linear polynomial models, see “Identifying Input-
Output Polynomial Models” on page 3-40.

Displaying the Confidence Interval

In addition, you can display a confidence interval for each pole and zero on the plot. To
learn how to show or hide confidence interval, see “How to Plot Model Poles and Zeros
Using the App” on page 10-58.

The confidence interval corresponds to the range of pole or zero values with a specific
probability of being the actual pole or zero of the system. The toolbox uses the estimated
uncertainty in the model parameters to calculate confidence intervals and assumes the
estimates have a Gaussian distribution.

For example, for a 95% confidence interval, the region around the nominal pole or zero
value represents the range of values that have a 95% probability of being the true system
pole or zero value. You can specify the confidence interval as a probability (between 0 and

10-56
 Examples

1) or as the number of standard deviations of a Gaussian distribution. For example, a

probability of 0.99 (99%) corresponds to 2.58 standard deviations.

10-57
10 Model Analysis

How to Plot Model Poles and Zeros Using the App

To create a pole-zero plot for parametric linear models in the System Identification app,
select the Zeros and poles check box in the Model Views area. For general information
about creating and working with plots, see “Working with Plots” on page 14-11.

To include or exclude a model on the plot, click the corresponding model icon in the
System Identification app. Active models display a thick line inside the Model Board
icon.

The following table summarizes the Zeros and Poles plot settings.

Zeros and Poles Plot Settings

Action Command
Display the confidence • To display the dashed lines on either side of the
interval. nominal pole and zero values, select Options > Show
confidence intervals. Select this option again to
hide the confidence intervals.
• To change the confidence value, select Options > Set
% confidence level, and choose a value from the list.
• To enter your own confidence level, select Options >
Set confidence level > Other. Enter the value as
a probability (between 0 and 1) or as the number of
standard deviations of a Gaussian distribution.
Show real and imaginary Select Style > Re/Im-axes. Select this option again to
axes. hide the axes.
Show the unit circle. Select Style > Unit circle. Select this option again to
hide the unit circle. The unit circle is useful as a reference
curve for discrete-time models.
(Multiple-output system Select the output by name in the Channel menu.
only)
Select an input-output pair
to view the poles and zeros
corresponding to those
channels.

10-58
 More About

More About
“Pole and Zero Plots” on page 10-54

10-59
10 Model Analysis

How to Plot Poles and Zeros at the Command Line

You can create a pole-zero plot for linear identified models using the iopzmap and
iopzplot commands.

To display confidence intervals for a specified number of standard deviations, use the
following syntax:
h = iopzplot(model);
showConfidence(h,sd)

where sd is the number of standard deviations of a Gaussian distribution. For example,

a confidence value of 99% for the nominal model curve corresponds to 2.58 standard
deviations.

Command Description Example

iopzmap,iopzplot Plots zeros and poles of To plot the poles and zeros
the model on the S-plane of the model sys, use the
or Z-plane for continuous- following command:
time or discrete-time model,
respectively. iopzmap(sys)

More About
“Pole and Zero Plots” on page 10-54

10-60
 Analyzing MIMO Models

Analyzing MIMO Models

In this section...
“Overview of Analyzing MIMO Models” on page 10-61
“Array Selector” on page 10-62
“I/O Grouping for MIMO Models” on page 10-64
“Selecting I/O Pairs” on page 10-65

Overview of Analyzing MIMO Models

If you import a MIMO system, or an LTI array containing multiple identified linear
models, you can use special features of the right-click menu to group the response plots
by input/output (I/O) pairs, or select individual plots for display. For example, generate a
random 3-input, 3-output MIMO system and then randomly sample it 10 times. Plot the
step response for all the models.

sys_mimo=rsample(idss(rss(3,3,3)),10);
step(sys_mimo);

sys_mimo is an array of ten 3-input, 3-output systems.

A set of 9 plots appears, one from each input to each output, for the ten model samples.

10-61
10 Model Analysis

Array Selector
If you plot an identified linear model array, Array Selector appears as an option in
the right-click menu. Selecting this option opens the Model Selector for LTI Arrays,
shown below.

10-62
 Analyzing MIMO Models

You can use this window to include or exclude models within the LTI array using various
criteria.

Arrays

Select the LTI array for model selection using the Arrays list.

Selection Criteria

There are two selection criteria. The default, Index into Dimensions, allows you to
include or exclude specified indices of the LTI Array. Select systems from the Selection
Criterion Setup and specify whether to show or hide the systems using the pull-down
menu below the Setup lists.

The second criterion is Bound on Characteristics. Selecting this options causes the
Model Selector to reconfigure. The reconfigured window is shown below

10-63
10 Model Analysis

Use this option to select systems for inclusion or exclusion in your response plot based on
their time response characteristics. The panel directly above the buttons describes how
to set the inclusion or exclusion criteria based on which selection criteria you select from
the reconfigured Selection Criteria Setup panel.

I/O Grouping for MIMO Models

You can group the plots by inputs, by outputs, or both by selecting I/O Grouping and
then Inputs, Outputs, or All, respectively, from the right-click menu.

For example, if you select Outputs, the step plot reconfigures into 3 plots, one for each
input.

10-64
 Analyzing MIMO Models

Selecting None returns to the default configuration, where all I/O pairs are displayed
individually.

Selecting I/O Pairs

Another way to organize MIMO system information is to choose I/O Selector from the
right-click menu, which opens the I/O Selector window.

10-65
10 Model Analysis

This window automatically configures to the number of I/O pairs in your MIMO system.
You can select:

• Any individual plot (only one at a time) by clicking on a button

• Any row or column by clicking on Y(*) or U(*)
• All of the plots by clicking [all]

Using these options, you can inspect individual I/O pairs, or look at particular I/O
channels in detail.

10-66
 Customizing Response Plots Using the Response Plots Property Editor

Customizing Response Plots Using the Response Plots Property

Editor

In this section...
“Opening the Property Editor” on page 10-67
“Overview of Response Plots Property Editor” on page 10-68
“Labels Pane” on page 10-70
“Limits Pane” on page 10-70
“Units Pane” on page 10-71
“Style Pane” on page 10-77
“Options Pane” on page 10-77
“Editing Subplots Using the Property Editor” on page 10-78

Opening the Property Editor

After you create a response plot, there are two ways to open the Property Editor:

• Double-click in the plot region

• Select Properties from the right-click menu

Before looking at the Property Editor, open a step response plot using these commands.

sys_dc = idtf([1 -0.8],[1 1 2 1]);

step(sys_dc)

This creates a step plot. Select Properties from the right-click menu. Note that when
you open the Property Editor, a set of black squares appear around the step response,
as this figure shows:

10-67
10 Model Analysis

SISO System Step Response

Overview of Response Plots Property Editor

This figure shows the Property Editor dialog box for a step response.

10-68
 Customizing Response Plots Using the Response Plots Property Editor

The Property Editor for Step Response

In general, you can change the following properties of response plots. Note that only the
Labels and Limits panes are available when using the Property Editor with Simulink
Design Optimization™ software.

• Titles and X- and Y-labels in the Labels pane.

• Numerical ranges of the X and Y axes in the Limits pane.
• Units where applicable (e.g., rad/s to Hertz) in the Units pane.

If you cannot customize units, as is the case with step responses, the Property Editor
will display that no units are available for the selected plot.
• Styles in the Styles pane.

You can show a grid, adjust font properties, such as font size, bold and italics, and
change the axes foreground color
• Change options where applicable in the Options pane.

These include peak response, settling time, phase and gain margins, etc. Plot options
change with each plot response type. The Property Editor displays only the options
that make sense for the selected response plot. For example, phase and gain margins
are not available for step responses.

10-69
10 Model Analysis

As you make changes in the Property Editor, they display immediately in the response
plot. Conversely, if you make changes in a plot using right-click menus, the Property
Editor for that plot automatically updates. The Property Editor and its associated plot
are dynamically linked.

Labels Pane
To specify new text for plot titles and axis labels, type the new string in the field next to
the label you want to change. Note that the label changes immediately as you type, so
you can see how the new text looks as you are typing.

Limits Pane
Default values for the axes limits make sure that the maximum and minimum x and y
values are displayed. If you want to override the default settings, change the values in
the Limits fields. The Auto-Scale box automatically clears if you click a different field.
The new limits appear immediately in the response plot.

To reestablish the default values, select the Auto-Scale box again.

10-70
 Customizing Response Plots Using the Response Plots Property Editor

Units Pane
You can use the Units pane to change units in your response plot. The contents of this
pane depend on the response plot associated with the editor.

The following table lists the options available for the response objects. Use the menus to
toggle between units.

Optional Unit Conversions for Response Plots

10-71
10 Model Analysis

Response Plot Unit Conversions

Bode and • Frequency
Bode Magnitude
By default, shows rad/TimeUnit where TimeUnit
is the system time units specified in the TimeUnit
property of the input system.

Frequency Units Options

• 'Hz'
• 'rad/s'
• 'rpm'
• 'kHz'
• 'MHz'
• 'GHz'
• 'rad/nanosecond'
• 'rad/microsecond'
• 'rad/millisecond'
• 'rad/minute'
• 'rad/hour'
• 'rad/day'
• 'rad/week'
• 'rad/month'
• 'rad/year'
• 'cycles/nanosecond'
• 'cycles/microsecond'
• 'cycles/millisecond'
• 'cycles/hour'
• 'cycles/day'
• 'cycles/week'
• 'cycles/month'

10-72
 Time Units Options

Response Plot Unit Conversions

• 'cycles/year'
• Frequency scale is logarithmic or linear.
• Magnitude in decibels (dB) or the absolute value
• Phase in degrees or radians

Impulse • Time.

By default, shows the system time units specified in

the TimeUnit property of the input system.

Time Units Options

• 'nanoseconds'
• 'microseconds'
• 'milliseconds'
• 'seconds'
• 'minutes'
• 'hours'
• 'days'
• 'weeks'
• 'months'
• 'years'
Nyquist Diagram • Frequency

By default, shows rad/TimeUnit where TimeUnit

is the system time units specified in the TimeUnit
property of the input system.

Frequency Units Options

• 'Hz'
• 'rad/s'

10-73
10 Model Analysis

Response Plot Unit Conversions

• 'rpm'
• 'kHz'
• 'MHz'
• 'GHz'
• 'rad/nanosecond'
• 'rad/microsecond'
• 'rad/millisecond'
• 'rad/minute'
• 'rad/hour'
• 'rad/day'
• 'rad/week'
• 'rad/month'
• 'rad/year'
• 'cycles/nanosecond'
• 'cycles/microsecond'
• 'cycles/millisecond'
• 'cycles/hour'
• 'cycles/day'
• 'cycles/week'
• 'cycles/month'
• 'cycles/year'

Pole/Zero Map • Time.

By default, shows the system time units specified in

the TimeUnit property of the input system.

Time Units Options

• 'nanoseconds'

10-74
 Frequency Units Options

Response Plot Unit Conversions

• 'microseconds'
• 'milliseconds'
• 'seconds'
• 'minutes'
• 'hours'
• 'days'
• 'weeks'
• 'months'
• 'years'
• Frequency

By default, shows rad/TimeUnit where TimeUnit

is the system time units specified in the TimeUnit
property of the input system.

Frequency Units Options

• 'Hz'
• 'rad/s'
• 'rpm'
• 'kHz'
• 'MHz'
• 'GHz'
• 'rad/nanosecond'
• 'rad/microsecond'
• 'rad/millisecond'
• 'rad/minute'
• 'rad/hour'
• 'rad/day'
• 'rad/week'

10-75
10 Model Analysis

Response Plot Unit Conversions

• 'rad/month'
• 'rad/year'
• 'cycles/nanosecond'
• 'cycles/microsecond'
• 'cycles/millisecond'
• 'cycles/hour'
• 'cycles/day'
• 'cycles/week'
• 'cycles/month'
• 'cycles/year'

Step • Time.

By default, shows the system time units specified in

the TimeUnit property of the input system.

Time Units Options

• 'nanoseconds'
• 'microseconds'
• 'milliseconds'
• 'seconds'
• 'minutes'
• 'hours'
• 'days'
• 'weeks'
• 'months'
• 'years'

10-76
 Time Units Options

Style Pane
Use the Style pane to toggle grid visibility and set font preferences and axes foreground
colors for response plots.

You have the following choices:

• Grid — Activate grids by default in new plots.

• Fonts — Set the font size, weight (bold), and angle (italic) for fonts used in response
plot titles, X/Y-labels, tick labels, and I/O-names.
• Colors — Specify the color vector to use for the axes foreground, which includes the
X-Y axes, grid lines, and tick labels. Use a three-element vector to represent red,
green, and blue (RGB) values. Vector element values can range from 0 to 1.

If you do not want to specify RGB values numerically, click the Select button to open
the Select Color dialog box.

Options Pane
The Options pane allows you to customize response characteristics for plots. Each
response plot has its own set of characteristics and optional settings; the table below lists
them. Use the check boxes to activate the feature and the fields to specify rise or settling
time percentages.

10-77
10 Model Analysis

Response Characteristic Options for Response Plots

Plot Customizable Feature

Bode Diagram and Bode Select lower magnitude limit
Magnitude Adjust phase offsets to keep phase close to a particular
value, within a range of ±180º, at a given frequency.
Unwrap phase (default is unwrapped)
Impulse Show settling time within xx% (specify the percentage)
Nyquist Diagram None
Pole/Zero Map None
Step Show settling time within xx% (specify the percentage)
Show rise time from xx to yy% (specify the percentages)

Editing Subplots Using the Property Editor

If you create more than one plot in a single figure window, you can edit each plot
individually. For example, the following code creates a figure with two plots, a step and
an impulse response with two randomly selected systems:

subplot(2,1,1)
step(rss(2,1))
subplot(2,1,2)
impulse(rss(1,1))

10-78
 Time Units Options

After the figure window appears, double-click in the upper (step response) plot to activate
the Property Editor. You will see a set of small black squares appear around the
step response, indicating that it is the active plot for the editor. To switch to the lower
(impulse response) plot, just click once in the impulse response plot region. The set of
black squares switches to the impulse response, and the Property Editor updates as
well.

10-79
10 Model Analysis

Akaike's Criteria for Model Validation

In this section...
“Definition of FPE” on page 10-80
“Computing FPE” on page 10-81
“Definition of AIC” on page 10-81
“Computing AIC” on page 10-82

Definition of FPE
Akaike's Final Prediction Error (FPE) criterion provides a measure of model quality
by simulating the situation where the model is tested on a different data set. After
computing several different models, you can compare them using this criterion.
According to Akaike's theory, the most accurate model has the smallest FPE.

Note: If you use the same data set for both model estimation and validation, the fit
always improves as you increase the model order and, therefore, the flexibility of the
model structure.

Akaike's Final Prediction Error (FPE) is defined by the following equation:

Ê1 + d N ˆ
FPE = V Á
Á 1 - d ˜˜
Ë N¯

where V is the loss function, d is the number of estimated parameters, and N is the
number of values in the estimation data set.

The toolbox assumes that the final prediction error is asymptotic for d<<N and uses the
following approximation to compute FPE:

FPE = V (1 + 2 d N )

The loss function V is defined by the following equation:

10-80
 Akaike's Criteria for Model Validation

Ê N ˆ
 e ( t, q N ) (e ( t, q N ))
T
1
V = det Á N ˜
Á ˜
Ë 1 ¯

where q N represents the estimated parameters.

Computing FPE
You can compute Akaike's Final Prediction Error (FPE) criterion for linear and nonlinear
models.

Note: FPE for nonlinear ARX models that include a tree partition nonlinearity is not
supported.

To compute FPE, use the fpe command, as follows:

FPE = fpe(m1,m2,m3,...,mN)

According to Akaike's theory, the most accurate model has the smallest FPE.

You can also access the FPE value of an estimated model , m, as follows:

• If m is a linear model, type m.Report.Fit.FPE

• If m is a nonlinear model, type m.EstimationInfo.FPE

Definition of AIC
Akaike's Information Criterion (AIC) provides a measure of model quality by simulating
the situation where the model is tested on a different data set. After computing several
different models, you can compare them using this criterion. According to Akaike's
theory, the most accurate model has the smallest AIC.

Note: If you use the same data set for both model estimation and validation, the fit
always improves as you increase the model order and, therefore, the flexibility of the
model structure.

Akaike's Information Criterion (AIC) is defined by the following equation:

10-81
10 Model Analysis

2d
AIC = log V +
N

where V is the loss function, d is the number of estimated parameters, and N is the
number of values in the estimation data set.

The loss function V is defined by the following equation:

Ê N ˆ
 e ( t, q N ) (e ( t, q N ))
T
1
V = det Á N ˜
Á ˜
Ë 1 ¯

where q N represents the estimated parameters.

For d<<N:

Ê Ê 2d ˆ ˆ
AIC = log Á V Á 1 + ˜
Ë Ë N ˜¯ ¯

Note: AIC is approximately equal to log(FPE).

Computing AIC
Use the aic command to compute Akaike's Information Criterion (AIC) for one or more
linear or nonlinear models, as follows:
AIC = aic(m1,m2,m3,...,mN)

According to Akaike's theory, the most accurate model has the smallest AIC.

10-82
 Computing Model Uncertainty

Computing Model Uncertainty

In this section...
“Why Analyze Model Uncertainty?” on page 10-83
“What Is Model Covariance?” on page 10-83
“Types of Model Uncertainty Information” on page 10-84

Why Analyze Model Uncertainty?

In addition to estimating model parameters, the toolbox algorithms also estimate
variability of the model parameters that result from random disturbances in the output.

Understanding model variability helps you to understand how different your model
parameters would be if you repeated the estimation using a different data set (with the
same input sequence as the original data set) and the same model structure.

When validating your parametric models, check the uncertainty values. Large
uncertainties in the parameters might be caused by high model orders, inadequate
excitation, and poor signal-to-noise ratio in the data.

Note: You can get model uncertainty data for linear parametric black-box models, and
both linear and nonlinear grey-box models. Supported model objects include idproc,
idpoly, idss, idtf, idgrey, idfrd, and idnlgrey.

What Is Model Covariance?

Uncertainty in the model is called model covariance.

When you estimate a model, the covariance matrix of the estimated parameters is stored
with the model. Use getcov to fetch the covariance matrix. Use getpvec to fetch the
list of parameters and their individual uncertainties that have been computed using the
covariance matrix. The covariance matrix is used to compute all uncertainties in model
output, Bode plots, residual plots, and pole-zero plots.

Computing the covariance matrix is based on the assumption that the model structure
gives the correct description of the system dynamics. For models that include a
disturbance model H, a correct uncertainty estimate assumes that the model produces

10-83
10 Model Analysis

white residuals. To determine whether you can trust the estimated model uncertainty
values, perform residual analysis tests on your model, as described in “Residual
Analysis” on page 10-23. If your model passes residual analysis tests, there is a good
chance that the true system lies within the confidence interval and any parameter
uncertainties results from random disturbances in the output.

For output-error models, such as transfer function models, state-space with K=0
and polynomial models of output-error form, with the noise model H fixed to 1, the
covariance matrix computation does not assume white residuals. Instead, the covariance
is estimated based on the estimated color of the residual correlations. This estimation of
the noise color is also performed for state-space models with K=0, which is equivalent to
an output-error model.

Types of Model Uncertainty Information

You can view the following uncertainty information from linear and nonlinear grey-box
models:

• Uncertainties of estimated parameters.

Type present(model) at the prompt, where model represents the name of a linear
or nonlinear model.
• Confidence intervals on the linear model plots, including step-response, impulse-
response, Bode, Nyquist, noise spectrum and pole-zero plots.

Confidence intervals are computed based on the variability in the model parameters.
For information about displaying confidence intervals, see the corresponding plot
section.
• Covariance matrix of the estimated parameters in linear models and nonlinear grey-
box models.

Use getcov.
• Estimated standard deviations of polynomial coefficients, poles/zeros, or state-space
matrices using idssdata, tfdata, zpkdata, and polydata.
• Simulated output values for linear models with standard deviations using the sim
command.

Call the sim command with output arguments, where the second output argument
is the estimated standard deviation of each output value. For example, type

10-84
 Computing Model Uncertainty

[ysim,ysimsd]=sim(model,data), where ysim is the simulated output, ysimsd

contains the standard deviations on the simulated output, and data is the simulation
data.
• To perform Monte-Carlo analysis, use rsample to generate a random sampling of
an identified model in a given confidence region. An array of identified systems of
the same structure as the input system is returned. The parameters of the returned
models are perturbed about their nominal values in a way that is consistent with the
parameter covariance.
• To simulate the effect of parameter uncertainties on a model's response, use simsd.

10-85
10 Model Analysis

Troubleshooting Models
In this section...
“About Troubleshooting Models” on page 10-86
“Model Order Is Too High or Too Low” on page 10-86
“Nonlinearity Estimator Produces a Poor Fit” on page 10-87
“Substantial Noise in the System” on page 10-87
“Unstable Models” on page 10-88
“Missing Input Variables” on page 10-89
“Complicated Nonlinearities” on page 10-89

About Troubleshooting Models

During validation, you might find that your model output fits the validation data poorly.
You might also find some unexpected or undesirable model characteristics.

If the tips suggested in these sections do not help improve your models, then a good
model might not be possible for this data. For example, your data might have poor signal-
to-noise ratio, large and nonstationary disturbances, or varying system properties.

Model Order Is Too High or Too Low

When the Model Output plot does not show a good fit, there is a good chance that you
need to try a different model order. System identification is largely a trial-and-error
process when selecting model structure and model order. Ideally, you want the lowest-
order model that adequately captures the system dynamics.

You can estimate the model order as described in “Preliminary Step – Estimating Model
Orders and Input Delays” on page 3-46. Typically, you use the suggested order as a
starting point to estimate the lowest possible order with different model structures. After
each estimation, you monitor the Model Output and the Residual Analysis plots, and
then adjust your settings for the next estimation.

When a low-order model fits the validation data poorly, try estimating a higher-order
model to see if the fit improves. For example, if a Model Output plot shows that a fourth-
order model gives poor results, try estimating an eighth-order model. When a higher-

10-86
 Troubleshooting Models

order model improves the fit, you can conclude that higher-order models might be
required and linear models might be sufficient.

You should use an independent data set to validate your models. If you use the same
data set to both estimate and validate a model, the fit always improves as you increase
model order, and you risk overfitting. However, if you use an independent data set to
validate your models, the fit eventually deteriorates if your model orders are too high.

High-order models are more expensive to compute and result in greater parameter
uncertainty.

Nonlinearity Estimator Produces a Poor Fit

In the case of nonlinear ARX and Hammerstein-Wiener models, the Model Output plot
does not show a good fit when the nonlinearity estimator has incorrect complexity.

You specify the complexity of piece-wise-linear, wavelet, sigmoid, and custom networks
using the number of units (NumberOfUnits nonlinear estimator property). A high
number of units indicates a complex nonlinearity estimator. In the case of neural
networks, you specify the complexity using the parameters of the network object. For
more information, see the Neural Network Toolbox documentation.

To select the appropriate complexity of the nonlinearity estimator, start with a low
complexity and validate the model output. Next, increate the complexity and validate the
model output again. The model fit degrades when the nonlinearity estimator becomes too
complex.

Note: To see the model fit degrade when the nonlinearity estimator becomes too complex,
you must use an independent data set to validate the data that is different from the
estimation data set.

Substantial Noise in the System

There are a couple of indications that you might have substantial noise in your system
and might need to use linear model structures that are better equipped to model noise.

One indication of noise is when a state-space model is better than an ARX model at
reproducing the measured output; whereas the state-space structure has sufficient
flexibility to model noise, the ARX model structure is less able to model noise because the

10-87
10 Model Analysis

A polynomial must account for both the system dynamics and the noise. The following
equation represents the ARX model and shows that A couples the dynamics and the noise
by appearing in the denominator of both the dynamics term and the noise terms:

B 1
y= u+ e
A A

Another indication that a noise model is needed appears in residual analysis plots when
you see significant autocorrelation of residuals at nonzero lags. For more information
about residual analysis, see “Residual Analysis” on page 10-23.

To model noise more carefully, use the ARMAX or the Box-Jenkins model structure,
where the dynamics term and the noise term are modeled by different polynomials.

Unstable Models
Unstable Linear Model

You can test whether a linear model is unstable is by examining the pole-zero plot of the
model, which is described in “Pole and Zero Plots” on page 10-54. The stability threshold
for pole values differs for discrete-time and continuous-time models, as follows:

• For stable continuous-time models, the real part of the pole is less than 0.
• For stable discrete-time models, the magnitude of the pole is less than 1.

Note: Linear trends might cause linear models to be unstable. However, detrending the
model does not guarantee stability.

When an unstable model is OK: In some cases, an unstable model is still a useful
model. For example, your system might be unstable without a controller, and you plan to
use your model for control design. In this case, you can import your unstable model into
Simulink or Control System Toolbox products.

Forcing stability during estimation: If you believe that your system is stable, but
your model is unstable, then you can estimate the model with the Focus estimation
option set to Stability. This setting might result in a reduced model quality. For more
information about Focus, see the various estimation option configuration commands
such as tfestOptions, ssestOptions, procestOptions etc..

10-88
 Troubleshooting Models

Allowing for some instability: A more advanced approach to achieving a stable model
is by setting the stability threshold estimation option to allow a margin of error. The
threshold estimation options are advanced properties of an estimation option set.

• For continuous-time models, set the value of

opt.Advanced.StabilityThreshold.s. The model is considered stable if the pole
on the far right is to the left of s stability threshold.
• For discrete-time models, set the value of opt.Advanced.StabilityThreshold.z.
The model is considered stable if all poles inside the circle centered at the origin and
with a radius of equal to the z stability threshold.

Unstable Nonlinear Models

To test if a nonlinear model is unstable is to plot the simulated model output on top of
the validation data. If the simulated output diverges from measured output, the model
is unstable. However, agreement between model output and measured output does not
guarantee stability.

Missing Input Variables

If the Model Output plot and Residual Analysis plot shows a poor fit and you have
already tried different structures and orders and modeled noise, it might be that there
are one or more missing inputs that have a significant effect on the output.

Try including other measured signals in your input data, and then estimating the models
again.

Inputs need not be control signals. Any measurable signal can be considered an input,
including measurable disturbances.

Complicated Nonlinearities
If the Model Output plot and Residual Analysis plot shows a poor fit, consider if
nonlinear effects are present in the system.

You can model the nonlinearities by performing a simple transformation on the signals
to make the problem linear in the new variables. For example, if electrical power is the
driving stimulus in a heating process and temperature is the output, you can form a
simple product of voltage and current measurements.

10-89
10 Model Analysis

If your problem is sufficiently complex and you do not have physical insight into the
problem, you might try fitting nonlinear black-box models.

10-90
 Next Steps After Getting an Accurate Model

Next Steps After Getting an Accurate Model

For linear parametric models, you can perform the following operations:

• Transform between continuous-time and discrete-time representation.

See “Transforming Between Discrete-Time and Continuous-Time Representations”.

• Transform between linear model representations, such as between polynomial, state-
space, and zero-pole representations.

See “Transforming Between Linear Model Representations”.

• Extract numerical data from transfer functions, pole-zero models, and state-space
matrices.

See “Extracting Numerical Model Data”.

For nonlinear black-box models (idnlarx and idnlhw objects), you can compute a linear
approximation of the nonlinear model. See “Linear Approximation of Nonlinear Black-
Box Models”.

System Identification Toolbox models in the MATLAB workspace are immediately

available to other MathWorks® products. However, if you used the System Identification
app to estimate models, you must first export the models to the MATLAB workspace.

Tip To export a model from the app, drag the model icon to the To Workspace rectangle.
Alternatively, right-click the model to open the Data/model Info dialog box. Click
Export.

If you have the Control System Toolbox software installed, you can import your linear
plant model for control-system design. For more information, see “Using Identified
Models for Control Design Applications” on page 12-2.

Finally, if you have Simulink software installed, you can exchange data between the
System Identification Toolbox software and the Simulink environment. For more
information, see “Simulating Identified Model Output in Simulink”.

10-91
10-92
 11

Setting Toolbox Preferences

• “Toolbox Preferences Editor” on page 11-2

• “Units Pane” on page 11-4
• “Style Pane” on page 11-7
• “Options Pane” on page 11-8
• “SISO Tool Pane” on page 11-9
11 Setting Toolbox Preferences

Toolbox Preferences Editor

In this section...
“Overview of the Toolbox Preferences Editor” on page 11-2
“Opening the Toolbox Preferences Editor” on page 11-2

Overview of the Toolbox Preferences Editor

The Toolbox Preferences editor allows you to set plot preferences that will persist from
session to session.

Opening the Toolbox Preferences Editor

To open the Toolbox Preferences editor, select Toolbox Preferences from the File
menu of the LTI Viewer or the SISO Design Tool. Alternatively, you can type
identpref

at the MATLAB prompt.

Control System Toolbox Preferences Editor

11-2
 Toolbox Preferences Editor

Note: The SISO Design Tool requires the Control System Toolbox software.

• “Units Pane” on page 11-4

• “Style Pane” on page 11-7
• “Options Pane” on page 11-8
• “SISO Tool Pane” on page 11-9

11-3
11 Setting Toolbox Preferences

Units Pane

Use the Units pane to set preferences for the following:

• Frequency

The default auto option uses rad/TimeUnit as the frequency units relative to
the system time units, where TimeUnit is the system time units specified in the
TimeUnit property of the system on frequency-domain plots. For multiple systems
with different time units, the units of the first system is used.

For the frequency axis, you can select logarithmic or linear scales.

Other Frequency Units Options

• 'Hz'
• 'rad/s'
• 'rpm'
• 'kHz'
• 'MHz'
• 'GHz'
• 'rad/nanosecond'
• 'rad/microsecond'

11-4
 Other Time Units Options

• 'rad/millisecond'
• 'rad/minute'
• 'rad/hour'
• 'rad/day'
• 'rad/week'
• 'rad/month'
• 'rad/year'
• 'cycles/nanosecond'
• 'cycles/microsecond'
• 'cycles/millisecond'
• 'cycles/hour'
• 'cycles/day'
• 'cycles/week'
• 'cycles/month'
• 'cycles/year'
• Magnitude — Decibels (dB) or absolute value (abs)
• Phase — Degrees or radians
• Time

The default auto option uses the time units specified in the TimeUnit property of the
system on the time- and frequency-domain plots. For multiple systems with different
time units, the units of the first system is used.

Other Time Units Options

• 'nanoseconds'
• 'microseconds'
• 'milliseconds'
• 'seconds'
• 'minutes'
• 'hours'
11-5
11 Setting Toolbox Preferences

• 'days'
• 'weeks'
• 'months'
• 'years'

11-6
 Style Pane

Style Pane
Use the Style pane to toggle grid visibility and set font preferences and axes foreground
colors for all plots you create. This figure shows the Style pane.

You have the following choices:

• Grid — Activate grids by default in new plots.

• Fonts — Set the font size, weight (bold), and angle (italic).
• Colors — Specify the color vector to use for the axes foreground, which includes the
X-Y axes, grid lines, and tick labels. Use a three-element vector to represent red,
green, and blue (RGB) values. Vector element values can range from 0 to 1.

If you do not want to specify RGB values numerically, click the Select button to open
the Select Colors dialog box.

11-7
11 Setting Toolbox Preferences

Options Pane
The Options pane has selections for time responses and frequency responses. This figure
shows the Options pane with default settings.

The following are the available options for the Options pane:

• Time Response:

• Show settling time within xx%— You can set the threshold of the settling time
calculation to any percentage from 0 to 100%. The default is 2%.
• Specify rise time from xx% to yy%— The standard definition of rise time is the
time it takes the signal to go from 10% to 90% of the final value. You can choose
any percentages you like (from 0% to 100%), provided that the first value is
smaller than the second.
• Frequency Response:

• Only show magnitude above xx—Specify a lower limit for magnitude values in
response plots so that you can focus on a region of interest.
• Unwrap phase—By default, the phase is unwrapped. Wrap the phrase by clearing
this box. If the phase is wrapped, all phase values are shifted such that their
equivalent value displays in the range [-180°, 180°).

11-8
 SISO Tool Pane

SISO Tool Pane

The SISO Tool pane has settings for the SISO Design Tool. This figure shows the SISO
Tool pane with default settings.

You can make the following selections:

• Compensator Format — Select the time constant, natural frequency, or zero/pole/

gain format. The time constant format is a factorization of the compensator transfer
function of the form

(1 + Tz1 s)
DC ¥ L
(1 + Tp1 s )

where DC is compensator DC gain, Tz1, Tz2, ..., are the zero time constants, and Tp1,
Tp2, ..., are the pole time constants.

The natural frequency format is

DC ¥
(1 + s wz ) L
1

(1 + s w p )
1

11-9
11 Setting Toolbox Preferences

where DC is compensator DC gain, ωz1, and ωz2, ... and ωp1, ωp2, ..., are the natural
frequencies of the zeros and poles, respectively.

The zero/pole/gain format is

( s + z1 )
K¥
( s + p1 )

where K is the overall compensator gain, and z1, z2, ... and p1, p2, ..., are the zero and
pole locations, respectively.
• Bode Options — By default, the SISO Design Tool shows the plant and sensor poles
and zeros as blue x's and o's, respectively. Clear this box to eliminate the plant's poles
and zeros from the Bode plot. Note that the compensator poles and zeros (in red) will
still appear.

11-10
 12

Control Design Applications

• “Using Identified Models for Control Design Applications” on page 12-2

• “Create and Plot Identified Models Using Control System Toolbox Software” on page
12-6
12 Control Design Applications

Using Identified Models for Control Design Applications

In this section...
“How Control System Toolbox Software Works with Identified Models” on page 12-2
“Using balred to Reduce Model Order” on page 12-2
“Compensator Design Using Control System Toolbox Software” on page 12-3
“Converting Models to LTI Objects” on page 12-3
“Viewing Model Response Using the LTI Viewer” on page 12-4
“Combining Model Objects” on page 12-5

How Control System Toolbox Software Works with Identified Models

System Identification Toolbox software integrates with Control System Toolbox software
by providing a plant for control design.

Control System Toolbox software also provides the LTI Viewer to extend System
Identification Toolbox functionality for linear model analysis.

Control System Toolbox software supports only linear models. If you identified a
nonlinear plant model using System Identification Toolbox software, you must linearize
it before you can work with this model in the Control System Toolbox software. For more
information, see the linapp, idnlarx/linearize, or idnlhw/linearize reference
page.

Note: You can only use the System Identification Toolbox software to linearize nonlinear
ARX (idnlarx) and Hammerstein-Wiener (idnlhw) models. Linearization of nonlinear
grey-box (idnlgrey) models is not supported.

Using balred to Reduce Model Order

In some cases, the order of your identified model might be higher than necessary to
capture the dynamics. If you have the Control System Toolbox software, you can use
balred to compute a state-spate model approximation with a reduced model order.

To learn how you can reduce model order using pole-zero plots, see “Reducing Model
Order Using Pole-Zero Plots” on page 10-56.

12-2
 Using Identified Models for Control Design Applications

Compensator Design Using Control System Toolbox Software

After you estimate a plant model using System Identification Toolbox software, you can
use Control System Toolbox software to design a controller for this plant.

System Identification Toolbox models in the MATLAB workspace are immediately

available to Control System Toolbox commands. However, if you used the System
Identification app to estimate models, you must first export the models to the MATLAB
workspace. To export a model from the app, drag the model icon to the To Workspace
rectangle. Alternatively, right-click the icon to open the Data/model Info dialog box. Click
Export to export the model.

Control System Toolbox software provides both the SISO Design Tool and commands for
working at the command line. You can import linear models directly into SISO Design
Tool using the following command:
sisotool(model)

You can also identify a linear model from measured SISO data and tune a PID controller
for the resulting model in the PID Tuner. You can interactively adjust the identified
parameters to obtain an LTI model whose response fits your response data. The PID
Tuner automatically tunes a PID controller for the identified model. You can then
interactively adjust the performance of the tuned control system, and save the identified
plant and tuned controller. To access the PID Tuner, enter pidTuner at the MATLAB
command line. For more information, see “PID Controller Tuning”.

Converting Models to LTI Objects

You can convert linear identified models into numeric LTI models (ss, tf, zpk) of
Control System Toolbox software.

The following table summarizes the commands for transforming linear state-space and
polynomial models to an LTI object.

Commands for Converting Models to LTI Objects

Command Description Example
frd Convert to frequency-response ss_sys = frd(model)
representation.
ss Convert to state-space ss_sys = ss(model)
representation.

12-3
12 Control Design Applications

Command Description Example

tf Convert to transfer-function tf_sys = tf(model)
form.
zpk Convert to zero-pole form. zpk_sys = zpk(model)

The following code converts the noise component of a linear identified model, sys, to a
numeric state-space model:
noise_model_ss = idss(sys,'noise');

To convert both the measured and noise components of a linear identified model, sys, to
a numeric state-space model:
model_ss = idss(sys,'augmented');

For more information about subreferencing the dynamic or the noise model, see
“Separation of Measured and Noise Components of Models” on page 3-117.

Viewing Model Response Using the LTI Viewer

• “What Is the LTI Viewer?” on page 12-4
• “Displaying Identified Models in the LTI Viewer” on page 12-4

What Is the LTI Viewer?

If you have the Control System Toolbox software, you can plot models in the LTI Viewer
from either the System Identification app or the MATLAB Command Window.

The LTI Viewer is a graphical user interface for viewing and manipulating the response
plots of linear models.

Note: The LTI Viewer does not display model uncertainty.

For more information about working with plots in the LTI Viewer, see the “Control
System Toolbox ” documentation.

Displaying Identified Models in the LTI Viewer

When the MATLAB software is installed, the System Identification app contains the To
LTI Viewer rectangle. To plot models in the LTI Viewer, do one of the following:

12-4
 Using Identified Models for Control Design Applications

• Drag and drop the corresponding icon to the To LTI Viewer rectangle in the System
Identification app.
• Right-click the icon to open the Data/model Info dialog box. Click Show in LTI
Viewer to plot the model in the LTI Viewer.

Alternatively, use the following syntax when working at the command line to view a
model in the LTI Viewer:
ltiview(model)

Combining Model Objects

If you have the Control System Toolbox software, you can combine linear model objects,
such as idtf, idgrey, idpoly, idproc, and idss model objects, similar to the way you
combine LTI objects. The result of these operations is a numeric LTI model that belongs
to the Control System Toolbox software. The only exceptions are the model stacking and
model concatenation operations, which deliver results as identified models.

For example, you can perform the following operations on identified models:

• G1+G2
• G1*G2
• append(G1,G2)
• feedback(G1,G2)

12-5
12 Control Design Applications

Create and Plot Identified Models Using Control System Toolbox

Software
This example shows how to create and plot models using the System Identification
Toolbox software and Control System Toolbox software. The example requires a Control
System Toolbox license.

Construct a random numeric model using the Control System Toolbox software.

rng('default');
sys0 = drss(3,3,2);

rng('default') specifies the setting of the random number generator as its default
setting.

sys0 is a third-order numeric state-space model with three outputs and two inputs.

Convert sys0 to an identified state-space model and set its output noise variance.

sys = idss(sys0);
sys.NoiseVariance = 0.1*eye(3);

Generate input data for simulating the output.

u = iddata([],idinput([800 2],'rbs'));

Simulate the model output with added noise.

opt = simOptions('AddNoise',true);
y = sim(sys,u,opt);

opt is an option set specifying simulation options. y is the simulated output for sys0.

Create an input-output ( iddata ) object.

data = [y u];

Estimate the state-space model from the generated data using ssest .

estimated_ss = ssest(data(1:400));

estimated_ss is an identified state-space model.

12-6
 Create and Plot Identified Models Using Control System Toolbox Software

Convert the identified state-space model to a numeric transfer function.

sys_tf = tf(estimated_ss);

Plot the model output for identified state-space model.

compare(data(401:800),estimated_ss)

Plot the response of identified model using the LTI Viewer.

ltiview(estimated_ss);

12-7
12 Control Design Applications

12-8
 13

System Identification Toolbox Blocks

• “Using System Identification Toolbox Blocks in Simulink Models” on page 13-2

• “Preparing Data” on page 13-3
• “Identifying Linear Models” on page 13-4
• “Simulating Identified Model Output in Simulink” on page 13-5
• “Simulate Identified Model Using Simulink Software” on page 13-8
13 System Identification Toolbox Blocks

Using System Identification Toolbox Blocks in Simulink Models

System Identification Toolbox software provides blocks for sharing information between
the MATLAB and Simulink environments.

You can use the System Identification Toolbox block library to perform the following
tasks:

• Stream time-domain data source (iddata object) into a Simulink model.

• Export data from a simulation in Simulink software as a System Identification
Toolbox data object (iddata object).
• Import estimated models into a Simulink model, and simulate the models with or
without noise.

The model you import might be a component of a larger system modeled in Simulink.
For example, if you identified a plant model using the System Identification Toolbox
software, you can import this plant into a Simulink model for control design.
• Estimate parameters of linear polynomial models during simulation from single-
output data.

To open the System Identification Toolbox block library, on the Home tab, in the
Simulink section, click Simulink Library. In the Library Browser, select System
Identification Toolbox.

You can also open the System Identification Toolbox block library directly by typing the
following command at the MATLAB prompt:
slident

To get help on a block, right-click the block in the Library Browser, and select Help.

13-2
 Preparing Data

Preparing Data
The following table summarizes the blocks you use to transfer data between the
MATLAB and Simulink environments.

After you add a block to the Simulink model, double-click the block to specify block
parameters. For an example of bringing data into a Simulink model, see the tutorial on
estimating process models in the “System Identification Toolbox Getting Started Guide”.

Block Description
Iddata Sink Export input and output signals to the MATLAB
workspace as an iddata object.
Iddata Source Import iddata object from the MATLAB workspace.

Input and output ports of the block correspond to input

and output signals of the data. These inputs and outputs
provide signals to blocks that are connected to this data
block.

For information about configuring each block, see the corresponding reference pages.

13-3
13 System Identification Toolbox Blocks

Identifying Linear Models

The following table summarizes the blocks you use to recursively estimate model
parameters in a Simulink model during simulation and export the results to the
MATLAB environment.

After you add a block to the model, double-click the block to specify block parameters.

Block Description
Recursive Least Squares Estimate model coefficients using recursive least squares
Estimator (RLS) algorithm
Recursive Polynomial Model Estimate input-output and time-series model coefficients
Estimator
Kalman Filter Estimate states of discrete-time or continuous-time linear
system
Model Type Converter Convert polynomial model coefficients to state-space
model matrices

For information about configuring each block, see the corresponding reference pages.

13-4
 Simulating Identified Model Output in Simulink

Simulating Identified Model Output in Simulink

In this section...
“When to Use Simulation Blocks” on page 13-5
“Summary of Simulation Blocks” on page 13-5
“Specifying Initial Conditions for Simulation” on page 13-6

When to Use Simulation Blocks

Add model simulation blocks to your Simulink model from the System Identification
Toolbox block library when you want to:

• Represent the dynamics of a physical component in a Simulink model using a data-

based nonlinear model.
• Replace a complex Simulink subsystem with a simpler data-based nonlinear model.

You use the model simulation blocks to import the models you identified using System
Identification Toolbox software from the MATLAB workspace into the Simulink
environment. For a list of System Identification Toolbox simulation blocks, see
“Summary of Simulation Blocks” on page 13-5.

Summary of Simulation Blocks

The following table summarizes the blocks you use to import models from the MATLAB
environment into a Simulink model for simulation. Importing a model corresponds to
entering the model variable name in the block parameter dialog box.

Block Description
“Idmodel” Simulate a linear identified model in Simulink software.
The model can be a process (idproc), linear polynomial
(idpoly), state-space (idss), grey-box (idgrey) and
transfer-function (idtf) model.
“Nonlinear ARX Model” Simulate idnlarx model in Simulink.
“Hammerstein-Wiener Simulate idnlhw model in Simulink.
Model”

13-5
13 System Identification Toolbox Blocks

Block Description
“Nonlinear Grey-Box Model” Simulate nonlinear ODE (idnlgrey model object) in
Simulink.

After you import the model into Simulink software, use the block parameter dialog box to
specify the initial conditions for simulating that block. (See “Specifying Initial Conditions
for Simulation” on page 13-6.) For information about configuring each block, see the
corresponding reference pages.

Specifying Initial Conditions for Simulation

For accurate simulation of a linear or a nonlinear model, you can use default initial
conditions or specify the initial conditions for simulation using the block parameters
dialog box.

• “Specifying Initial States of Linear Models” on page 13-6

• “Specifying Initial States of Nonlinear ARX Models” on page 13-7
• “Specifying Initial States of Hammerstein-Wiener Models” on page 13-7

Specifying Initial States of Linear Models

Specify the initial states for simulation in the Initial states (state space only: idss,
idgrey) field of the Function Block Parameters: Idmodel dialog box:

• For idss and idgrey models, initial states must be a vector of length equal to the
order of the model.
• For models other than idss and idgrey, initial conditions are zero.
• In some situations, you may want to match the simulated response of the model to a
certain input/output data set:

1 Convert the identified model into state-space form (idss model), and use the
state-space model in the block.
2 Compute the initial state values that produce the best fit between the model
output and the measured output signal using findstates(idParametric).
3 Specify the same input signal for simulation that you used as the validation data
in the app or in the compare plot.

For example:

13-6
 Simulating Identified Model Output in Simulink

% Convert to state-space model

mss = idss(m);
% Estimate initial states from data
X0 = findstates(mss,z);

z is the data set you used for validating the model m. Use the model mss and
initial states X0 in the Idmodel block to perform the simulation.

Specifying Initial States of Nonlinear ARX Models

The states of a nonlinear ARX model correspond to the dynamic elements of the
nonlinear ARX model structure, which are the model regressors. Regressors can be the
delayed input/output variables (standard regressors) or user-defined transformations
of delayed input/output variables (custom regressors). For more information about the
states of a nonlinear ARX model, see the idnlarx reference page.

For simulating nonlinear ARX models, you can specify the initial conditions as input/
output values, or as a vector. For more information about specifying initial conditions for
simulation, see the IDNLARX Model reference page.

Specifying Initial States of Hammerstein-Wiener Models

The states of a Hammerstein-Wiener model correspond to the states of the embedded

linear (idpoly or idss) model. For more information about the states of a Hammerstein-
Wiener model, see the idnlhw reference page.

The default initial state for simulating a Hammerstein-Wiener model is 0. For more
information about specifying initial conditions for simulation, see the IDNLHW Model
reference page.

13-7
13 System Identification Toolbox Blocks

Simulate Identified Model Using Simulink Software

This example shows how to set the initial states for simulating a model such that the
simulation provides a best fit to measured input-output data.

Prerequisites
Estimate a model, M, using a multiple-experiment data set, Z, which contains data from
three experiments — z1, z2, and z3:
% Load multi-experiment data.
load(fullfile(matlabroot, 'toolbox', 'ident', 'iddemos',...
'data', 'twobodiesdata'));

% Create an iddata object to store the multi-experiment data.

z1=iddata(y1, u1, 0.005,'Tstart',0);
z2=iddata(y2, u2, 0.005,'Tstart',0);
z3=iddata(y3, u3, 0.005,'Tstart',0);
Z = merge(z1,z2,z3);

% Estimate a 5th order state-space model.

opt = n4sidOptions('Focus','simulation');
[M,x0] = n4sid(Z,5,opt);

To simulate the model using input u2, use x0(:,2) as the initial states. x0(:,2) is
computed to maximize the fit between the measured output, y2, and the response of M.

To compute initial states that maximizes the fit to the corresponding output y2, and
simulate the model using the second experiment:

1 Extract the initial states that correspond to the second experiment for simulation:
X0est = x0(:,2);
2 Open the System Identification Toolbox library by typing the following command at
the MATLAB prompt:
slident
3 Open a new Simulink model window. Then, drag and drop an Idmodel block from
the library into the model window.
4 Open the Function Block Parameters dialog box by double-clicking the Idmodel
block. Specify the following block parameters:

13-8
 Prerequisites

a In the Model variable field, type M to specify the estimated model.

b In the Initial state field, type X0est to specify the estimated initial states.
Click OK.
5 Drag and drop an Iddata Source block into the model window. Then, configure the
model, as shown in the following figure.

6 Simulate the model for 2 seconds, and compare the simulated output ysim with the
measured output ymeasured using the Scope block.

13-9
13-10
 14

System Identification App

• “Steps for Using the System Identification App” on page 14-2

• “Working with System Identification App” on page 14-3
14 System Identification App

Steps for Using the System Identification App

A typical workflow in the System Identification app includes the following steps:

1 Import your data into the MATLAB workspace, as described in “Representing Data
in MATLAB Workspace” on page 2-9.
2 Start a new session in the System Identification app, or open a saved session. For
more information, see “Starting a New Session in the App” on page 14-3.
3 Import data into the app from the MATLAB workspace. For more information, see
“Represent Data”.
4 Plot and preprocess data to prepare it for system identification. For example, you can
remove constant offsets or linear trends (for linear models only), filter data, or select
data regions of interest. For more information, see “Preprocess Data”.
5 Specify the data for estimation and validation. For more information, see “Specify
Estimation and Validation Data in the App” on page 2-30.
6 Select the model type to estimate using the Estimate menu.
7 Validate models. For more information, see “Model Validation”.
8 Export models to the MATLAB workspace for further analysis. For more
information, see “Exporting Models from the App to the MATLAB Workspace” on
page 14-10.

14-2
 Working with System Identification App

Working with System Identification App

In this section...
“Starting and Managing Sessions” on page 14-3
“Managing Models” on page 14-7
“Working with Plots” on page 14-11
“Customizing the System Identification App” on page 14-15

Starting and Managing Sessions

• “What Is a System Identification Session?” on page 14-3
• “Starting a New Session in the App” on page 14-3
• “Description of the System Identification App Window” on page 14-4
• “Opening a Saved Session” on page 14-6
• “Saving, Merging, and Closing Sessions” on page 14-6
• “Deleting a Session” on page 14-7

What Is a System Identification Session?

A session represents the total progress of your identification process, including any data
sets and models in the System Identification app.

You can save a session to a file with a .sid extension. For example, you can save
different stages of your progress as different sessions so that you can revert to any stage
by simply opening the corresponding session.

To start a new session, see “Starting a New Session in the App” on page 14-3.

For more information about the steps for using the System Identification app, see “Steps
for Using the System Identification App” on page 14-2.

Starting a New Session in the App

To start a new session in the System Identification app, type systemIdentification

in the MATLAB Command Window:

systemIdentification

14-3
14 System Identification App

Alternatively, you can start a new session by selecting the Apps tab of MATLAB
desktop. In the Apps section, click System Identification. This action opens the
System Identification app.

Note: Only one session can be open at a time.

You can also start a new session by closing the current session using File > Close
session. This toolbox prompts you to save your current session if it is not already saved.

Description of the System Identification App Window

The following figure describes the different areas in the System Identification app.

14-4
 Working with System Identification App

Data Board Model Board

Select check
boxes to display Select check
data plots. boxes to display
model plots.

The layout of the window organizes tasks and information from left to right. This
organization follows a typical workflow, where you start in the top-left corner by
importing data into the System Identification app using the Import data menu and
end in the bottom-right corner by plotting the characteristics of your estimated model on
model plots. For more information about using the System Identification app, see “Steps
for Using the System Identification App” on page 14-2.

The Data Board area, located below the Import data menu in the System
Identification app, contains rectangular icons that represent the data you imported into
the app.

The Model Board, located to the right of the <--Preprocess menu in the System
Identification app, contains rectangular icons that represent the models you estimated or
imported into the app. You can drag and drop model icons in the Model Board into open
dialog boxes.

14-5
14 System Identification App

Opening a Saved Session

You can open a previously saved session using the following syntax:
systemIdentification(session,path)

session is the file name of the session you want to open and path is the location of
the session file. Session files have the extension .sid. When the session file in on the
matlabpath, you can omit the path argument.

If the System Identification app is already open, you can open a session by selecting File
> Open session.

Note: If there is data in the System Identification app, you must close the current session
before you can open a new session by selecting File > Close session.

Saving, Merging, and Closing Sessions

The following table summarizes the menu commands for saving, merging, and closing
sessions in the System Identification app.

Task Command Comment

Close the current File > Close session You are prompted to save the current
session and start session before closing it.
a new session.
Merge the File > Merge session You must start a new session and
current session import data or models before you can
with a previously select to merge it with a previously
saved session. saved session. You are prompted to
select the session file to merge with
the current. This operation combines
the data and the models of both
sessions in the current session.
Save the current File > Save Useful for saving the session
session. repeatedly after you have already
saved the session once.
Save the current File > Save As Useful when you want to save your
session under a work incrementally. This command
new name.

14-6
 Working with System Identification App

Task Command Comment

lets you revert to a previous stage, if
necessary.

Deleting a Session

To delete a saved session, you must delete the corresponding session file.

Managing Models
• “Importing Models into the App” on page 14-7
• “Viewing Model Properties” on page 14-8
• “Renaming Models and Changing Display Color” on page 14-8
• “Organizing Model Icons” on page 14-9
• “Deleting Models in the App” on page 14-10
• “Exporting Models from the App to the MATLAB Workspace” on page 14-10

Importing Models into the App

You can import System Identification Toolbox models from the MATLAB workspace into
the System Identification app. If you have Control System Toolbox software, you can also
import any models (LTI objects) you created using this toolbox.

The following procedure assumes that you begin with the System Identification app
already open. If this window is not open, type the following command at the prompt:

systemIdentification

To import models into the System Identification app:

1 Select Import from the Import models list to open the Import Model Object dialog
box.
2 In the Enter the name field, type the name of a model object. Press Enter.
3 (Optional) In the Notes field, type any notes you want to store with this model.
4 Click Import.
5 Click Close to close the Import Model Object dialog box.

14-7
14 System Identification App

Viewing Model Properties

You can get information about each model in the System Identification app by right-
clicking the corresponding model icon.

The Data/model Info dialog box opens. This dialog box describes the contents and the
properties of the corresponding model. It also displays any associated notes and the
command-line equivalent of the operations you used to create this model.

Tip To view or modify properties for several models, keep this window open and right-
click each model in the System Identification app. The Data/model Info dialog box
updates when you select each model.

Renaming Models and Changing Display Color

You can rename a model and change its display color by double-clicking the model icon in
the System Identification app.

The Data/model Info dialog box opens. This dialog box describes both the contents
and the properties of the model. The object description area displays the syntax of the
operations you used to create the model in the app.

To rename the model, enter a new name in the Model name field.

You can also specify a new display color using three RGB values in the Color field.
Each value is between 0 to 1 and indicates the relative presence of red, green, and blue,
respectively. For more information about specifying default data color, see “Customizing
the System Identification App” on page 14-15.

Tip As an alternative to using three RGB values, you can enter any one of the following
letters in single quotes:
'y' 'r' 'b' 'c' 'g' 'm' 'k'

These strings represent yellow, red, blue, cyan, green, magenta, and black, respectively.

Finally, you can enter comments about the origin and state of the model in the Diary
And Notes area.

14-8
 Working with System Identification App

To view model properties in the MATLAB Command Window, click Present.

Organizing Model Icons

You can rearrange model icons in the System Identification app by dragging and
dropping the icons to empty Model Board rectangles.

Note: You cannot drag and drop a model icon into the data area on the left.

When you need additional space for organizing model icons, select Options > Extra
model/data board in the System Identification app. This action opens an extra session
window with blank rectangles. The new window is an extension of the current session
and does not represent a new session.

Tip When you import or estimate models and there is insufficient space for the icons, an
additional session window opens automatically.

You can drag and drop model icons between the main System Identification app and any
extra session windows.

Type comments in the Notes field to describe the models. When you save a session, as
described in “Saving, Merging, and Closing Sessions” on page 14-6, all additional
windows and notes are also saved.

14-9
14 System Identification App

Deleting Models in the App

To delete models in the System Identification app, drag and drop the corresponding icon
into Trash. You can also use the Delete key on your keyboard to move items to the
Trash. Moving items to Trash does not permanently delete these items.

To restore a model from Trash, drag its icon from Trash to the Model Board in the
System Identification app. You can view the Trash contents by double-clicking the
Trash icon.

Note: You must restore a model to the Model Board; you cannot drag model icons to the
Data Board.

To permanently delete all items in Trash, select Options > Empty trash.

Exiting a session empties Trash automatically.

Exporting Models from the App to the MATLAB Workspace

The models you create in the System Identification app are not available in the MATLAB
workspace until you export them. Exporting is necessary when you need to perform an
operation on the model that is only available at the command line. Exporting models to
the MATLAB workspace also makes them available to the Simulink software or another
toolbox, such as the Control System Toolbox product.

To export a model to the MATLAB workspace, do one of the following:

14-10
 Working with System Identification App

• Drag and drop the corresponding icon to the To Workspace rectangle.

• Right-click the icon to open the Data/model Info dialog box. Click Export to export
the model.

When you export models to the MATLAB workspace, the resulting variables have the
same name as in the System Identification app.

Working with Plots

• “Identifying Data Sets and Models on Plots” on page 14-11
• “Changing and Restoring Default Axis Limits” on page 14-12
• “Selecting Measured and Noise Channels in Plots” on page 14-14
• “Grid and Line Styles in Plots” on page 14-14
• “Opening a Plot in a MATLAB Figure Window” on page 14-15
• “Printing Plots” on page 14-15

Identifying Data Sets and Models on Plots

You can identify data sets and models on a plot by color: the color of the line in the data
or model icon in the System Identification app matches the line color on the plots.

You can also display data tips for each line on the plot by clicking a plot curve and
holding down the mouse button.

Note: You must disable zoom by selecting Style > Zoom before you can display data tips.
For more information about enabling zoom, see “Magnifying Plots” on page 14-12.

The following figure shows an example of a data tip, which contains the name of the data
set and the coordinates of the data point.

14-11
14 System Identification App

Data Tip on a Plot

Changing and Restoring Default Axis Limits

There are two ways to change which portion of the plot is currently in view:

• “Magnifying Plots” on page 14-12

• “Setting Axis Limits” on page 14-13

Magnifying Plots

Enable zoom by selecting Style > Zoom in the plot window. To disable zoom, select Style
> Zoom again.

Tip To verify that zoom is active, click the Style menu. A check mark should appear next
to Zoom.

You can adjust magnification in the following ways:

14-12
 Working with System Identification App

• To zoom in default increments, left-click the portion of the plot you want to center in
the plot window.
• To zoom in on a specific region, click and drag a rectangle that identifies the region for
magnification. When you release the mouse button, the selected region is displayed.
• To zoom out, right-click on the plot.

Note: To restore the full range of the data in view, select Options > Autorange in the
plot window.

Setting Axis Limits

You can change axis limits for the vertical and the horizontal axes of the input and
output channels that are currently displayed on the plot.

1 Select Options > Set axes limits to open the Limits dialog box.
2 Specify a new range for each axis by editing its lower and upper limits. The limits
must be entered using the format [LowerLimit UpperLimit]. Click Apply. For
example:

[0.1 100]

Note: To restore full axis limits, select the Auto check box to the right of the axis
name, and click Apply.
3 To plot data on a linear scale, clear the Log check box to the right of the axis name,
and click Apply.

Note: To revert to base-10 logarithmic scale, select the Log check box to the right of
the axis name, and click Apply.
4 Click Close.

Note: To view the entire data range, select Options > Autorange in the plot window.

14-13
14 System Identification App

Selecting Measured and Noise Channels in Plots

Model inputs and outputs are called channels. When you create a plot of a multivariable
input-output data set or model, the plot only shows one input-output channel pair at a
time. The selected channel names are displayed in the title bar of the plot window.

Note: When you select to plot multiple data sets, and each data set contains several
input and output channels, the Channel menu lists channel pairs from all data sets.

You can select a different input-output channel pair from the Channel menu in any
System Identification Toolbox plot window.

The Channel menu uses the following notation for channels: u1->y2 means that the
plot displays a transfer function from input channel u1 to output channel y2. System
Identification Toolbox estimates as many noise sources as there are output channels. In
general, e@ynam indicates that the noise source corresponds to the output with name
ynam.

For example, e@y3->y1 means that the transfer function from the noise channel
(associated with y3) to output channel y2 is displayed. For more information about noise
channels, see “Separation of Measured and Noise Components of Models” on page 3-117.

Tip When you import data into the System Identification app, it is helpful to assign
meaningful channel names in the Import Data dialog box. For more information about
importing data, see “Represent Data”.

Grid and Line Styles in Plots

There are several Style options that are common to all plot types. These include the
following:

• “Grid Lines” on page 14-14

• “Solid or Dashed Lines” on page 14-15

Grid Lines

To toggle showing or hiding grid lines, select Style > Grid.

14-14
 Working with System Identification App

Solid or Dashed Lines

To display currently visible lines as a combination of solid, dashed, dotted, and dash-
dotted line style, select Style > Separate linestyles.

To display all solid lines, select Style > All solid lines. This choice is the default.

All line styles match the color of the corresponding data or model icon in the System
Identification app.

Opening a Plot in a MATLAB Figure Window

The MATLAB Figure window provides editing and printing commands for plots that are
not available in the System Identification Toolbox plot window. To take advantage of this
functionality, you can first create a plot in the System Identification app, and then open
it in a MATLAB Figure window to fine-tune the display.

After you create the plot, as described in “Plotting Models in the App” on page 10-6,
select File > Copy figure in the plot window. This command opens the plot in a
MATLAB Figure window.

Printing Plots

To print a System Identification Toolbox plot, select File > Print in the plot window. In
the Print dialog box, select the printing options and click OK.

Customizing the System Identification App

• “Types of App Customization” on page 14-15
• “Saving Session Preferences” on page 14-16
• “Modifying idlayout.m” on page 14-16

Types of App Customization

The System Identification app lets you customize the window behavior and appearance.
For example, you can set the size and position of specific dialog boxes and modify the
appearance of plots.

You can save the session to save the customized app state.

You might choose to edit the file that controls default settings, as described in “Modifying
idlayout.m” on page 14-16 (advanced usage).

14-15
14 System Identification App

Saving Session Preferences

Use Options > Save preferences to save the current state of the System Identification
app. This command saves the following settings to a preferences file, idprefs.mat:

• Size and position of the System Identification app

• Sizes and positions of dialog boxes
• Four recently used sessions
• Plot options, such as line styles, zoom, grid, and whether the input is plotted using
zero-order hold or first-order hold between samples

You can only edit idprefs.mat by changing preferences in the app.

The idprefs.mat file is located in the same folder as startup.m, by default. To change
the location where your preferences are saved, use the midprefs command with the new
path as the argument. For example:
midprefs('c:\matlab\toolbox\local\')

You can also type midprefs and browse to the desired folder.

To restore the default preferences, select Options > Default preferences.

Modifying idlayout.m

You might want to customize the default plot options by editing idlayout.m (advanced
usage).

To customize idlayout.m defaults, save a copy of idlayout.m to a folder in your

matlabpath just above the ident folder level.

Caution Do not edit the original file to avoid overwriting the idlayout.m defaults
shipped with the product.

You can customize the following plot options in idlayout.m:

• Order in which colors are assigned to data and model icons

• Line colors on plots
• Axis limits and tick marks

14-16
 Working with System Identification App

• Plot options, set in the plot menus

• Font size

Note: When you save preferences using Options > Save preferences to idprefs.mat,
these preferences override the defaults in idlayout.m. To give idlayout.m precedence
every time you start a new session, select Options > Default preferences.

14-17
14-18