FDA Batch - Jan 2005
Uploaded by
rishita26FDA Batch - Jan 2005
Uploaded by
rishita26Analysis, Monitoring
and Control of Batch
Processes:
A perspective from historical
use in the chemical industry
John MacGregor
McMaster Advanced Control Consortium (MACC)
Chemical Engineering Dept.
McMaster University
Hamilton, ON, Canada
Problems to be solved for Batch
Processes
Analysis of historical problems with the batch
operations:
Have records on history of past batches
Which batches were different and why?
Diagnose and fix the problems
On-line monitoring of new batches
Collect data on new batches with the improved process
Build PCA/PLS model for good batches (common cause
varaition only)
Set-up on-line monitoring based on this model
Problems to be solved for Batch
Processes
Control of final product quality for each batch
Process behavior can show that if the batch is to
continue as is, then product quality will not be good.
Make small mid-course adjustment to bring quality back
into control
Use LV models and control in LV space
Product Design &Transfer between Plants
How to operate a batch to achieve a specified product
quality
Data to build a LV model – Invert model to get conditions
Typical Industrial Batch
Reactor
Production of an 2.
agricultural chemical
Differential Pressure
PIC
3. Dryer Pressure
1. Level in Collector
Collector Tank
4. Power
5. Torque Resistance
6. Speed
Heating Medium
Manually
Dryer Tank Determined
10. Dryer Temperature
8. Jacket Temperature TIC TIC 9. Dryer Temperature SP
7. Jacket Temperature SP
A Batch Run
Collector Tank Level
Dryer Temp
Agitator Speed
A Batch Run
Collector Tank Level
Dryer Temp
Agitator Speed
A Batch Run
Collector Tank Level
Dryer Temp
Agitator Speed
A Batch Run
Collector Tank Level
Dryer Temp
Agitator Speed
A Batch Run
Collector Tank Level
Dryer Temp
Agitator Speed
Available Data
Initial Conditions (Z)
Initial Chemical Analysis of wet cake
Weight of Wet Cake
Process (X)
10 Trajectories
Batch Duration
Final (Y)
Chemical analysis of product (~10 properties)
Residual solvent concentration
On-Spec/Off-Spec/Hi-solvent classification
Nature of Industrial Batch
Data
End Properties
e
it m variables
batches
Z X Y
Initial Conditions Variable Trajectories
Usually 200,000 – 500,000 observations in data sets
Process Variable Trajectories:
X Data
Multi-way PCA for Batch
Trajectory Data
Unfold the 3-way array of data so that each batch
corresponds to one row
Multi-way PLS for Batch Data
LV
scores
batche
Z X Y T
s
VT WT QT
Loadings
Models the time varying covariance structure
among all the process variables over the entire
time history of the batch
Every batch summarized by a few LV scores (t1,
t2, t3)
Relates the IC’s and time varying trajectory
information to the final product quality
A. Trouble-shooting a Batch
Agricultural Chemical Process
Problem
Molecule was great for purpose intended
But, could not produce it consistently in industrial batch
process
Almost 50% of batches were bad
Thought the problem related to chemistry –initial charges and
reaction variations
Data available Z, X, Y on many past batches
PLS analysis showed problems were simple
operational problems
Easily fixed
PLS model for Y vs. Z (IC data)
Score plot of Z space for
historical batches Variable importance plot for
model
See that chemistry variables aren’t the problem – operating
variables are!
PLS model for operating data (X) vs. Y
Each batch summarized by 2 LV scores (t1,
t2)
t2
t1
Good separation of batches. Good batches have high t 1=Xw1
Interpretation using PLS loading plot for
w1
Each variable has 350 loading weights associated with the 350 time intervals of the batch
• Good batches have: -low CTTANKL over entire batch, high x1,Jtempsp and Dtemp
during early batch stage, and low batch times. Consistently doing this – good product!
B. Process Monitoring
Once have fixed the process then want to maintain
gains and detect and remove other problems
Build a LV model from historical data with only
acceptable operation
Any deviation from this model will reveal
unacceptable behavior
Statistics to plot:
a
T t / s
Joint t1-t2 plots or 2 2 2
Hotelling’s T2: l l
l 1
k
Residual SPE:
SPEi ( xij xij ) 2
j 1
B. Process Monitoring – Ag
Chemical
Plots: Hotelling’s T2 and SPE
Monitoring of new batch #73
T2 plot SPE plot
Contribution plots to diagnose
the problem
Problem: Variable x6 diverged above its nominal trajectory at time
277
Industrial Usage
Analysis and monitoring of batch process by
multivariate PCA/PLS methods is becoming well
established
Chemicals
Steel
Microelectronics
Forest products
Product release is often based on these charts
rather than QC lab analyses
C: Process Control
If can successfully monitor batch processes using
initial condition and trajectory data, then why not
take active control if the product is predicted to
be poor?
Mid-course correction strategies:
Adjust few discrete variables at some point (e.g.
initiator, batch time)
Adjust MV trajectories for remainder of batch (eg. Temp
profile)
Only do so for those few batches where MSPC indicates
final product will be unacceptable
Concept: Mid-Course Corrective
Control
Final PSD
Prediction
On-line
Measurements (Tr)
Trajectories
o
t2
*
t1
Reduce Space
uc0 uc1 time
Off-line
measurements
Mid-Course Correction Control Of
PSD
Adjustments: Emulsifier shot and batch
time adjustment
Control results in LV space PSD results for batch #2
-8
x 10 Control of normalized PSD
9
2
8
Targe After
t control
Particle number distribution (mol/l-nm)
7
3
No Control
2
0
0 10 20 30 40 50 60 70
radius (nm)
Good Industrial Results (Mitsubishi
Chemicals): Mid-Course Control of Dp
Before and After Implemention
Year # of Std. # Off
Batches Deviation Spec.
1 75 .030 9
2 68 .027 6
3 76 .026 4
4 72 .030 5
current 44 .015 0
D. PRODUCT DESIGN
PROBLEM:
Given new product specifications ydes find recipe and
process conditions xdes which will produce this product.
Possible Approaches:
Theoretical models and optimization
Response surface methods – DOE
Statistical methods based on historical data
Operating Conditions for New
Product – Product Design
X Temperatures
Pressures
Concentrations
Recipes
Flows
Trajectories
Y Density
Tensile Strength
Mw, Mn
Transparency
Biological activity
Toxicity
Hydrophobicity
Design of batch emulsion
polymerization process for new
product
13 variables in ydes
y1 to y6 and y8 are to be kept in normal range
y7 = y7des and y9 = y9des
y10< y10const, y11< y11const, y12 < y12const and y13 < y13const
With the minimal possible batch time
Batch emulsion polymerization
data (pilot plant)
Solutions using different weights on
time
All solutions yield same ydes
Case 1 to 5 Increase weight on time-usage
E. Product transfer between
plants
Source Site Target Site
Process Conditions
Process Conditions
Chemical or Physical Properties Product 4
Product 1 For each product Chemical or Physical Properties
For each product
Product 5
Product 2
Product 3
Summary – Batch Processes
Latent Variable methods are now becoming well
established in industry for analysis, monitoring
& control
Useful for product design and transfer
Latent variable models easy to build
Incorporate information on all measured process
variables
Key is to work in the reduced latent variable space
Latent variable scores summarize all the important
information
Process variable trajectories and final quality can be
predicted from these