Process Safety and Environmental Protection 135 (2020) 113–125

Contents lists available at ScienceDirect

Process Safety and Environmental Protection

journal homepage: www.elsevier.com/locate/psep

Root causality analysis at early abnormal stage using principal

component analysis and multivariate Granger causality
Hahyung Pyun a , Kyeongsu Kim b , Daegeun Ha c , Chul-Jin Lee d,∗ , Won Bo Lee a,∗∗
a
School of Chemical and Biological Engineering, Seoul National University, Gwanak-ro, Gwanak-gu, Seoul, 08826, Republic of Korea
b
Clean Energy Research Center, Korea Institute of Science and Technology, 5, Hwarang-ro 14-gil, Seongbuk-gu, Seoul, 02792, Republic of Korea
c
Conwell Co., 9-22, Pangyo-ro 255beon-gil, Bundang-gu, Seongnam-si, Gyeonggi-do, 13486, Republic of Korea
d
School of Chemical Engineering and Materials Science, Chung-Ang University, 84 Heukseok-ro, Dongjak-gu, Seoul, 06974, Republic of Korea

a r t i c l e i n f o a b s t r a c t

Article history: As fault detection technologies have been developed, process fault diagnosis at early abnormal stage
Received 21 September 2019 has come to be considered a major problem. In this work, a method to analyze the root cause of faults is
Received in revised form developed to provide proper information at the early abnormal stage. First, principal component analysis
13 November 2019
(PCA) is used for the early detection of the process fault. Then, the contributions, from which the normal
Accepted 12 December 2019
portion is removed, are decomposed by singular value decomposition (SVD) method to select the hierar-
Available online 20 December 2019
chical sensors. Finally, the multivariate Granger causality (MVGC) method is used to construct the sensor
causalities using the hierarchical sensors. The developed methodology is verified using the liquefied nat-
Keywords:
Principal component analysis (PCA)
ural gas fractionation process model, which embeds a sufficient number of highly correlated sensors. The
Multivariate Granger causality (MVGC) results are compared with the conventional principal component analysis method and amplification of
Fault magnitude the residual contribution method to verify the advantages of the proposed method.
Root causality © 2019 Institution of Chemical Engineers. Published by Elsevier B.V. All rights reserved.

1. Introduction process monitoring using PCA, Hotelling’s T 2 and Q-statistic are

used to determine whether a state of a process is a fault or not
The development of physical sensing technology, distributed (Ku et al., 1995). When the fault occurs, the two indicators comple-
control system (DCS), and computing technologies have brought mentarily give information about the fault sample. Because of the
about the development of the plant scale. As a result, an enormous reduction in the size of the data, PCA, using the two indicators, is
amount of data resulting from these huge processes enables the suitable for performing a quick fault detection, visualization, and
detailed analysis of systems to maximize productivity and mini- diagnosis. In the steel industry, the adjusted PCA has been adopted
mize safety issues. In industries, univariate monitoring methods to analyze casting and desulfurization processes for early detec-
are commonly used to analyze plant data. Although they have con- tion and visualization (Dudzic et al., 1999; Dudzic and Miletic,
tributed to the stable operation of plants, this data require analysis 2002). For the detailed analysis of the process state, moving PCA
by skilled operators as the sensors have several limits, such as con- (MPCA) and nonlinear PCA (NLPCA) methods, such as kernel PCA,
trol, warning, and risk limits. For this reason, multivariate statistical have been developed (Chehreghani, 2017; Choi et al., 2005; Geng
process control (MSPC) methods have been developed in the aca- and Zhu, 2005; Kano et al., 2001; Karhunen et al., 1998; Zhu et al.,
demic area and adjusted for use in real industrial plants (Kresta 2012). These advanced methods have strengths in non-Gaussianity
et al., 1991; MacGregor et al., 2007). Principal component analy- of data for early fault detection. However, it also has deficiencies
sis (PCA) is one of the most preferred methods for use in various and breakdowns in its contribution to detecting faults. This can
systems, chemical plants, steel industries, fuel cells, and batch pro- happen when the normal PCA model is constructed with noisy or
cesses (Joe Qin, 2003; Qin, 2012; Santen et al., 2003). This method unimportant data in screening the process state. To solve this prob-
decomposes a large number of sensors into a fewer number of prin- lem, Jeong et al. addressed this problem in the integration of factor
cipal components while maintaining the variation of the data. In analysis and PCA (Jeong et al., 2012). This method sorts out the sen-
sors, which are disturbed in the molten carbon fuel cell, to get an
accurate time for fault detection and diagnostic contribution. For
detailed normal PCA modeling, a multi-mode PCA method is devel-
∗ Corresponding author.
∗∗ Corresponding author. oped, integrating hierarchical clustering with PCA for a global PCA
E-mail addresses: cjlee@cau.ac.kr (C.-J. Lee), wblee@snu.ac.kr (W.B. Lee). combined with individual PCA modes. Jiang et al. proposed the inte-

https://doi.org/10.1016/j.psep.2019.12.010
0957-5820/© 2019 Institution of Chemical Engineers. Published by Elsevier B.V. All rights reserved.
114 H. Pyun et al. / Process Safety and Environmental Protection 135 (2020) 113–125

gration of Bayesian interference and joint probability with PCA that 2. Preliminaries
adjusts training and identifies the various sub-block normal modes
(Jiang and Yan, 2014). Ha et al. and used the k-nearest neighbor 2.1. Principal component analysis
method to matching the local normal mode and the adjusted PCA
for detection (Ha et al., 2017). PCA is one of the conventional methods for reducing the dimen-
Most relevant studies focus on improving detecting efficiencies; sion of a data set. It is a very powerful method for a large-scale
thus, the PCA model is integrated with other methods to enhance plant because of the presence of an enormous number of inter-
monitoring efficiency. However, root cause analysis is as important correlated variables. A normal data set, in which the process has
as fast monitoring. In general, an analysis of a root cause depends on no abnormal condition, is calculated as a training data set to be
historical data, qualitative knowledge, or an expert system (Dvorak scaled to standardization. The data are composed of n rows ×
and Kuipers, 1991; Gabbar et al., 2014; Nan et al., 2008; Narozny m columns in which the rows are samples (time series) and the
et al., 2001; Qin, 2012; Venkatasubramanian et al., 2003a, 2003b). columns are sensors, X ∈ Rn×m . Then, data set X is decomposed in
Although there is an enormous amount of historical data in the the PCA model, as given in Eq. (1),
data storage system, the storage system mainly comprises normal
operation data or different process conditions compared with the X = TP T + E (1)
present condition. Thus, it is difficult to accurately represent the
where T is the score matrix, P T is the loading vector, and E is the
current process state. Knowledge based qualitative analysis or an
residual matrix. A new sample vector, which means real time data
expert system are very accurate. However, they have a major weak-
for monitoring and diagnosis, can be projected as two parts, a prin-
ness. Abnormal situations are various depending on the condition;
cipal component subspace (PCS) and a residual subspace (RS), as
therefore, there are too many cases to analyze in advance. In addi-
represented in Eqs. (2)–(4),
tion, analysis after a fault can be very time-consuming. To address
these problems, data driven fault detection and root cause diag-
x = x̂ + ∼x (2)
nosis have been developed recently. MacGregor and Kourti, Yue
et al., Qin, and Kitano et al. have proposed a fault reconstruction x̂ = PP T x (3)
scheme that integrates the Hotelling’s T2 and Q-statistic approach  
T
to increase the accuracy of fault contribution (He et al., 2005; Joe ∼x = I − PP x (4)
Qin, 2003; Kitano et al., 2018; Yue and Qin, 2001).
Recently, methods of root-cause diagnosis with this concept where x̂ is the PCS projection, ∼x is the RS projection, and both
have been studied. In terms of root-cause analysis, Granger causal- parameters have their own monitoring variable in each subspace.
ity is one of the most used methodologies. This method is based The Hotelling’s T 2 is the variations in PCS, and squared prediction
on the vector autoregressive model (VAR), which is a linear regress error (SPE) is the variation in RS. The Hotelling’s T 2 refers to how
model.(Barnett and Seth, 2014) Because Granger causality just uses far the observation is from the center of the multivariate normal
the time series data identified fault or abnormal, it is used in distribution. The Q-statistic, which is the SPE, indicates the residual
economics, biology and chemistry (Chen et al., 2018; Ding et al., between the projection of a new sample and PCA modeling. The
2006; Jiang and Yan, 2014; Wen et al., 2013; Yuan and Qin, 2014). limits of the two monitoring variables are defined as Eqs. (5)–(10),
Although this method is simple and powerful, if considering a
large number of variables, the root cause direction can become T 2 = xT PD−1 P T x (5)
twisted or complicated. To solve these problems, Ahmed et al.  
n2 − 1
used singular value decomposition (SVD) to amplify the root cause T2 ≤ F˛ (a, n − a) (6)
variables (Ahmed et al., 2017a, 2017b). Using the residual con- n (n − a)
tribution fault direction, which is modeled by the same exact   2
historical fault data, the contribution data for the propagation path SPE = ∼x2 =  I − PP T x (7)
is enhanced. This methodology is useful for frequently occurring   1⁄h0
faults. However, because it depends on the same exact historical h0 c˛ 22 2 h0 (h0 − 1)
fault data, it is difficult to give root cause information about new SPE ≤ 1 +1+ (8)
1 1
2
faults.
In this study, to solve the problem of PCA and Granger causal-
ity in root cause analysis, a method of selecting fault magnitude 
m
i
i = j (9)
variables is developed. Particularly, the method is focused on the
root-cause analysis of a new fault and the initial fault stage. The PCA j=a+1

method is used to detect the abnormal state using the Hotelling’s T2

21 3
and Q-statistic. When the fault is detected, its contribution data are h0 = 1 − 2
(10)
scaled and analyzed by single value decomposition (SVD), such that 32
the magnitude sensors affected by the fault are identified. These
where a represents the number of selected principal components,
fault magnitude sensors are adjusted to the multivariable Granger
F˛ (a, n − a) is the F-distribution with a and (n−a) degrees of free-
causality (MVGC) method. Through the causality matrix of the mag-
dom, c˛ is the value of the normal distribution, and ˛ is the level of
nitude sensors, the proposed method performs effectively in root
significance (Jackson, 1980; Jackson and Mudholkar, 1979).
cause diagnosis.
This paper is divided into four major sections. The first sec-
tion describes the theory regarding PCA and contribution handling. 2.2. Multivariate Granger causality
In Section 2, the following methodologies are proposed: model-
ing, fault detection, and root cause diagnosis. The next section The Granger causality (GC) is based on linear autoregressive
briefly describes the LNG fractionation process to present a case modeling of stochastic processes (Chen et al., 2018; Ding et al.,
study and fault scenario. Then, the results and discussion of a 2006; Wen et al., 2013). Briefly, if a variable X2 (t) has information
fault scenario are presented. Finally, the last section gives the of the future X1 (t), and there is no information in a group of other
conclusion. variables for predicting the future of X1 (t), then X2 (t) is said to be
 H. Pyun et al. / Process Safety and Environmental Protection 135 (2020) 113–125 115

Fig. 1. Selecting procedure of fault magnitude variables.

a ‘Granger cause’ of X1 (t). This concept is interpreted based on the

vector autoregressive (VAR) approach, as shown in Eqs. (11)–(14),


p

p

X1 (t) = A11 (j) X1 (t − j) + A12 (j) X2 (t − j) + ε1 (t) (11)

j j

var (ε1 (t)) = ˙1 (12)


p

p

X2 (t) = A21 (j) X1 (t − j) + A22 (j) X2 (t − j) + ε2 (t) (13)

j j

var (ε2 (t)) = ˙2 (14)

Fig. 2. Procedure of normal operation data modeling.
where A(j) is the autoregressive (AR) coefficient, k is the model
order, and ε(t) is the prediction error. These equations provide
This concept can be extended to the multivariate system using the
the definition of the full regression bivariate AR model. If there is
AR coefficients, given as Eq. (20) (Wen et al., 2013).
no dependence between X1 (t) and X2 (t), A12 (j) and A21 (j) are 0.
⎧ ⎫ ⎡ ⎤⎧ ⎫
This concept considers the reduced regression, as expressed in Eqs. ⎪ X1 (t) ⎪ A11 (k) A12 (k) · · · A1m (k)
⎪ X1 (t − k) ⎪
⎪
⎪ ⎪
⎪ ⎪
⎪ ⎪
⎪
(15)–(18). ⎪
⎨ X2 (t) ⎪
⎬ ⎢ ⎥ ⎪ ⎪
∞
⎢ A21 (k) A22 (k) · · · A2m (k) ⎥ ⎨ X2 (t − k) ⎬

p = ⎢ ⎥
⎪ .. ⎪ ⎢ .. . . ⎥⎪. ⎪
X1 (t) =

A 11 (j) X1 (t − j) + ε1(2) (t) (15) ⎪
⎪ ⎪
⎪ ⎣ . . . . ⎦⎪⎪ . ⎪
⎪
⎪
⎩. ⎪
⎭
k=1 .. . . ⎪
⎩. ⎪
⎭
j
Xm (t) Am1 (k) Am2 (k) · · · Amm (k) Xm (t − k)
  ⎧ ⎫
var ε1(2) (t) = ˙1(2) (16) ε1 (t)
⎪
⎪ ⎪
⎪
⎪
⎪ ⎪
⎪

p ⎨ ε2 (t) ⎬

X2 (t) = A 22 (j) X2 (t − j) + ε2(1) (t) (17) + (20)
⎪
⎪ .. ⎪
⎪
j ⎪
⎪ ⎪
⎪
  ⎩. ⎭
εm (t)
var ε2(1) (t) = ˙2(1) (18)

A means reduced AR coefficient, εi(j) refers to the prediction error of GC from Xj (t) to Xi (t) is given in Eq. (21).
the ith variable excluding the jth variable, and ˙i(j) means variance ˙i(j)
ε at the restricted jth variable. GC from X2 (t) to X1 (t) is defined as F Xj →Xi = ln (21)
˙i
a log-likelihood ratio, as given in Eq. (19).
In this multivariate form, ˙i means cov(εi ), which is defined
˙1(2) as the variance from all other m variables, and ˙i(j) represent
F X2 →X1 = ln (19)
˙1 cov(εi(j) ), which is defined for the (m− 1) variables that restrict j.
116 H. Pyun et al. / Process Safety and Environmental Protection 135 (2020) 113–125

In this work, the ‘MVGC tool box’, developed by Barnett et al. in

matlab code, is used for analysis (Barnett and Seth, 2014).

3. Fault detection and root cause analysis

3.1. Selection method of fault magnitude variables

When a fault occurs, a new sample data contains normal and

abnormal portions. These two portions are also reflected in the
principal component subspace (PCS) and residual subspace (RS),
respectively. It is expressed in Eqs. (22)–(24),

x = x∗ + i f (22)

x̂ = x̂∗ + 
ˆ if (23)
∗
∼x = ∼x + ∼i f (24)

where i represents the fault direction,  ˆ i and ∼i are the fault
directions on PCS and RS, respectively, and i refers to the number

of principal components. The strength of the fault is represented
by f , which changes over time. Generally, the portion of con-
tribution about x̂∗ and ∼x∗ is insignificant compared with the fault
strength. Therefore, the contribution of x is about the same as the
contribution of i f . However, it is difficult to ignore that very small
intensity and the initial stage of fault (Ahmed et al., 2017b). There-
fore, the normal variation embedded in the fault data need to be
removed or minimized. For this purpose, the statistics of the nor-
mal contribution data, which is used to train data in PCA modeling,
should be used to scale the fault contribution data. They are given
as Eqs. (25) and (26).

x̂ = x̂ − x̂∗ = 
ˆ if (25)

∼x = ∼x − ∼x∗ = ∼ i f (26)

When the alarm occurred, the fault data set collected form a
fault scenario, which containing k time of samples corresponding
to m sensors, is expressed as Eq. (27).

Xi = [ x1 x2 . . . xk ]T (27)

It can be interpreted as the fault portion from Eqs. (25) and (26),
in which these variables have meaning contributions that are scaled
Fig. 3. Algorithm for monitoring and root cause diagnosis.

Fig. 4. Schematic of LNG fractionation process and fault location.

 H. Pyun et al. / Process Safety and Environmental Protection 135 (2020) 113–125 117

Table 1
Sensor information of LNG fractionation process.

Tag No. Tag name Tag description

1 dM-in-T de-Mthanizer inlet stream temperature

2 dM-top-P de-Methanizer top stream pressure
3 dM-top-T de-Methanizer top stream temperature
4 dM-1st-P de-Methanizer stage-1 pressure
5 dM-7st-T de-Methanizer stage-7 temperature
6 dM-13st-P de-Methanizer stage-13 pressure
7 dM-reb-T de-Methanizer reboiler temperature
8 dE-in-F (dM-bot-F) de-Ethanizer inlet stream flow rate (de-Methanizer bottom stream flow rate)
9 dE-in-T (dM-bot-T) de-Ethanizer inlet stream temperature (de-Methanizer bottom stream temperature)
10 dE-top-T de-Ethanizer top stream temperature
11 dE-cond-P de-Ethanizer condenser pressure
12 dE-cond-T de-Ethanizer condenser temperature
13 dE-1st-P de-Ethanizer stage-1 pressure
14 dE-11st-P de-Ethanizer stage-11 pressure
15 dE-11st-T de-Ethanizer stage-11 temperature
16 dE-20st-T de-Ethanizer stage-20 temperature
17 dE-28st-P de-Ethanizer stage-28 pressure
18 dE-reb-T de-Ethanizer reboiler temperature
19 dP-in-F (dE-bot-F) de-Propanizer inlet flow rate (de-Ethanizer bottom stream flow rate)
20 dP-in-T (dE-bot-T) de-Propanizer input temperature (de-Ethanizer bottom stream temperature)
21 dP-cond-P de-Propanizer condenser pressure
22 dP-1st-P de-Propanizer stage-1 pressure
23 dP-19st-T de-Propanizer stage-19 temperature
24 dP-37st-P de-Propanizer stage-37 pressure
25 dP-reb-T de-Propanizer reboiler temperature
26 dB-in-F (dP-bot-F) de-Butanizer inlet flow rate (de-Butanizer bottom stream flow rate)
27 dB-in-T (dP-bot-T) de-Butanizer input temperature (de-Butanizer bottom stream temperature)
28 dB-cond-P de-Butanizer condenser pressure
29 dB-1st-P de-Butanizer stage-1 pressure
30 dB-17st-T de-Butanizer stage-17 temperature
31 dB-34st-P de-Butanizer stage-34 pressure
32 dB-reb-T de-Butanizer reboiler temperature
33 dB-bot-T de-Butanizer bottom stream temperature

Table 2
Product specification of LNG fractionation process.

C1[%] C2[%] C3[%] C4[%] others

deMethanizer top stream 91.60 5.31 2.06 0.80 0.23

deEthanizer top stream 0.00 99.55 0.45
dePropanizer top stream 0.00 0.00 99.61 0.39
deButanizer top stream 0.00 0.00 1.23 98.00 0.77

Table 3
Product specification of LNG fractionation process.

Temperature [◦ ] Pressure [bar] Flow rate[kg/h]

Input −17.8 to −14.8 61.0–63.5 51,000–56,000

by the normal contribution. Scaled contributions are given as Eqs.

(28) and (29),
 T
ˆX i = ˆ i [ f1 f 2 . . . fk ] (28)
 T
∼X i = ∼ i [ f1 f 2 . . . fk ] (29)

where the two equations represent PCS and RS, respectively. In

 T  T
these spaces, ˆX i or ∼X i is converted to the covariance matrix
for analysis of the covariation contributions, as given in Eqs. (30)
and (31).
  T

T
Cov ˆX i =ˆU iˆDiˆV i (30)
  T

Cov ∼X i = ∼U i ∼Di ∼V i T (31)

These two parameters are reconstructed from singular value

decomposition (SVD). This method converts correlated variables
into uncorrelated variables while retaining the singular values. The Fig. 5. Fault detection using (a) Hotelling’s T2 chart and (b) SPE chart in fault scenario
first column of SVD, Ui , represents the fault magnitude of each of 1.
118 H. Pyun et al. / Process Safety and Environmental Protection 135 (2020) 113–125

Fig. 6. T-contribution chart at fault detection time, 240 s, for fault scenario 1.

Fig. 7. RS-amplification chart for fault scenario 1.

the variables, which is the first eigenvalue in diagonal matrix Di . 3.2. Normal operation data modeling
Parameters ˆ i and ∼ i are given as Eqs. (32) and (33).
As general process monitoring steps, the normal operation data
is modeled first, such that the PCA method could construct the pro-
ˆ i = Ûi (:, 1)
 (32) cess limit and information of normal contribution in PCS and RS,
respectively. First, the normal operation data, for which these data
are already known as the normal state, are gathered. These data go
∼ i = ∼U i (:, 1) (33) through preprocessing. Outlier data are eliminated by the 3-sigma
method in statistics, which means that those data outside the 99.7
% normal data range are removed for model accuracy. Then, the
data are scaled by the average and standard deviation. After a sta-
Because it is a new model that reconstructed the SVD method-
ble data set has been prepared, PCA method is used to decompose
ology using the data at the time the abnormality was found, the
them to reduced spaces, PCS and RS. In these two spaces, the limits
first row of Ûi or ∼U i represeants the degree of contribution on
for monitoring and the scale statistics of contribution, respectively,
the current process state. Therefore, sensors with high contribu-
are calculated. Fig. 2 shows the procedure for handling the normal
tion to model construction can be decided as sensors that indicate
data.
abnormalities. These sensors, fault magnitude sensors or hierarchi-
cal sensors, are selected by the procedure shown in Fig. 1. From the
absolute fault magnitude data, 32 % is selected as hierarchical sen- 3.3. Fault detection and root cause analysis
sors for multivariate Granger causality analysis, representing those
that have contributions larger than the sum of the mean and the Fault detection and diagnosis consist of two parts. The first part
1-sigma value. is monitoring the process. New data, real-time data, goes through
 H. Pyun et al. / Process Safety and Environmental Protection 135 (2020) 113–125 119

Fig. 8. Fault magnitude chart for fault scenario 1.

Table 4
Causal matrix using RS-amplification method for fault scenario 1.

Table 5
Causal matrix using fault magnitude method for fault scenario 1.

the scaling process used in normal data statistics. Then, the scaled and then, the covariance of these contribution data is analyzed by
data are projected to the PCA normal model. From the projection, SVD method to select the hierarchical sensors. The fault magnitude
the model is monitored by the Hotelling’s T 2 and SPE. When the pro- method removes the normal portion, Ui (:,1), with the empirical rule
cess alarm occurs, the fault data are collected. Then, the subspace selecting sensors with more than 32 %. Finally, the MVGC method
is determined by which alarm occurred. If the sample data exceed is performed for these sensors, resulting in the construction of a
the Hotelling’s T 2 at t time, the data are gathered from (t − k) time causality matrix. If the alarm occurs from SPE, the process is the
to t time. This data goes into PC-subspace. In this space, the T- same as PCS, except that it is performed in RS. This algorithm is
contributions are calculated and scaled by normal contribution, described in Fig. 3.
120 H. Pyun et al. / Process Safety and Environmental Protection 135 (2020) 113–125

dynamic model is developed using the Aspen hysys® simulator. It

has four main columns; deMethanizer, deEthanizer, dePropanizer,
and deButanizer, separating methane, ethane, propane, and
butane, respectively. They are monitored by a total of 33 sensors,
which monitor flow rate [m3 /h], pressure [bar], and temperature
[◦ C]. Because the process involves small carbons, the operating
conditions of the column are low temperature and high pressure.
Therefore, the correlation between the sensors is very sensitive.
A schematic of the LNG fractionation process is shown in Fig. 4.
Sensor information is shown in Table 1, and product specifications
are shown in Table 2. The LNG dynamic model simulation referred
to Kim’s thesis.(Kim, 2011) For the PCA normal modeling, normal
operation data are collected for which the condition of the input
Fig. 9. Causality flow using (a) RS amplification and (b) fault magnitude for fault
scenario 1.
stream is shown in Table 3. Temperature, pressure, and flow rate
are given as sinusoidal variations within the parameters of the
process specifications.
4. Case studies

4.1. Liquefied natural gas (LNG) plant fractionation process

4.2. Development of fault scenarios

The processing in the LNG plant involves four major tasks: gas
In this study, two cases are developed to verify the algorithm.
treatment, liquefaction, fractionation, and storage. In this work, to
The first case involves reboiler overheating case. Overheating is
generate data for verifying the algorithm, the fractionation process

Fig. 10. Visualized causality in process diagram for fault scenario 1.

Table 6
Causal matrix using SPE contribution for fault scenario 2.
 H. Pyun et al. / Process Safety and Environmental Protection 135 (2020) 113–125 121

Table 7
Causal matrix using RS-amplification method for fault scenario 2.

Table 8
Causal matrix using fault magnitude method for fault scenario 2.

Fig. 11. Fault detection using (a) Hotelling’s T2 chart and (b) SPE chart in fault scenario 2.

one of the most common faults in column operation. This fault is increase of 3 % in the heat duty is very small compared with the
caused by several reasons, such as improper change in the control normal value, but if it lasts, flooding can occur in the column. The
limit, accurate temperature not detected, malfunction in the con- second case is the condenser overcooling in the deEthanizer col-
trol logic, steam valve opening, and human error (Bahadori, 2014; umn. It is also a common fault in column operation (Bahadori, 2014;
Stichlmair and Fair, 1998). To simulate this fault, the heat duty of Stichlmair and Fair, 1998). This fault can be compared with the first
deEthanizer reboiler is increased to 3 % over the normal state. An case in that propagation path is expected to be in the opposite direc-
122 H. Pyun et al. / Process Safety and Environmental Protection 135 (2020) 113–125

Fig. 12. T-contribution chart at fault detection time, 165 s, for fault scenario 2.

Fig. 13. RS-amplification chart for fault scenario 2.

tion. To simulate this fault, the heat duty of deEthanizer condenser perature sensors around the deEthanizer reboiler, such as dE-reb-T,
is increased to 3 % above the normal state. Fig. 4 is described the dP-in-T (dE-bot-T), or dE-20st-T should provide larger values than
location of fault 1 and 2. those from the other sensors. However, Fig. 6 shows the misinter-
pretation about the fault represented in the T-contribution chart.
The chart points to irrelevant variables, such as dM-in-T, dM-top-
4.3. Fault scenario 1 T, and dP-reb-T, which show higher contributions than the entire
group of deEthanizer sensors. This contribution trend is due to the
In the first scenario, reboiler overheating, an alarm is generated intensity of fault that represents a small-scale fault or initial stage.
by the Hotelling’s T2 at 240 s after the fault occurs. In RS, the SPE Therefore, a similar or greater portion is used to represent normal
undergoes an alarm at 289 s. Because PCS detects the alarm ear- information than fault information. Because there are no major sen-
lier than RS does, the analysis should occur through the PCS. Fig. 5 sors associated with the fault, conventional PCA is not performed
shows the Hotelling’s T2 chart and SPE chart. in the MVGC analysis.
To demonstrate the excellence of the algorithm, conventional The RS amplification method uses the fault direction, which
PCA contribution and RS amplification are compared with the is constructed from the historical SPE contribution of same fault,
developed algorithm. Generally, a contribution chart is used at to enhance affected sensors. Therefore, as shown in Fig. 7, this
alarm time for finding sensors affected by fault. If the contribution method causes dE-reb-T and dP-in-T(dE-bot-T) to be sufficiently
chart provide proper information about reboiler overheating, tem- higher than the other sensors. However dM-in-T, dM-top-T, and
 H. Pyun et al. / Process Safety and Environmental Protection 135 (2020) 113–125 123

Fig. 14. Fault magnitude chart for fault scenario 2.

Fig. 15. Causality flow using (a) SPE contribution (b) RS amplification and (c) fault magnitude for fault scenario 2.

dB-in-F(dP-bot-F), which have the next largest values, screen the reb-T. They are reasonable processes from the interpretation point
relevant sensors that are affected by the fault. Because the alarm of view because the initial root causes are located near the reboiler.
occurs in PCS, RS cannot contain the overall fault information. Thus, On the other hand, RS amplification shows the main causality is
it is clear that an insufficient amount of information would be pro- from dP-in-T(dE-bot-T) and dM-in-T to dM-top-T.
vided. These screening sensors act as misleading elements in the Intuitively, these causalities are not reasonable because the
MVGC analysis. physical distance between the sensors is too far or these are not sen-
Compared with conventional PCA contribution and RS amplifi- sor relationships that can affect or be affected. These compared root
cation, the developed algorithm shows a much more reasonable causalities are described in Fig. 9. For a visual comparison, the fault
result, as indicated in Fig. 8. The sensors of deEthanizer mainly magnitude method and RS amplification method are expressed in
affected by fault are increased, such as dE-20st-T, dE-reb-T, and the process diagram, as shown in Fig. 10. Green arrows indicate
dP-in-T(dE-bot-T). In addition, portions of unrelated fault sensors the RS amplification causality, and red arrows represent the fault
are decreased, such as dM-in-T, dM-top-T, and dP-cond-P. This is magnitude causality. The developed method should be recognized
because a new model was constructed from the data at the time more clearly in the process diagram.
of detecting the fault. Therefore, ignoring a portion of the normal
contribution enhances the related sensors and diminishes the nor-
4.4. Fault scenario 2
mal behavior of the unaffected sensors, such that the hierarchical
sensors can stand out.
In the second scenario, condenser overcooling, the SPE alarm
Using the method mentioned in Fig. 1, the magnitude sensors
occurs earlier than the Hotelling’s T 2 limit. The detection time
are selected for analysis in the MVGC method. From these sen-
is 165 s in RS, and the alarm does not occur until approximately
sors, MVGC analysis constructs the causality matrix. This matrix
2000 s in PCS. Fig. 11 shows the Hotelling’s T2 chart and SPE chart.
shows that proposed method can provide the root causality more
Therefore, the root causality should be analyzed in RS. To prove the
clearly than the RS amplification method can. The causality matrix
excellence of the algorithm, just as in fault scenario 1, the conven-
is described in Tables 4 and 5 for the RS amplification method and
tional PCA contribution and RS amplification are compared with
fault magnitude method, respectively. Row variables means causal
the developed algorithm.
sensors and column variables are the affected sensors. A significant
Using the SPE contribution at 165 s, Fig. 12 shows the SPE
amount of data, through a comparison of internal data in its own
contribution chart at the alarm time. Sensors dE-cond-P, dE-
table, is shaded thickly, and slightly larger data shaded thinly. The
1st-P, dE-11st-P, dE-28st-P, and dP-in-F(dE-bot-F) are the most
proposed method shows the greatest strength of cause and effect
affected. Using RS amplification makes this result clearer. Fig. 13
from dE-reb-T to dE-20st-T. Next is from dP-in-T(dE-bot-T) to dE-
124 H. Pyun et al. / Process Safety and Environmental Protection 135 (2020) 113–125

Fig. 16. Visualized causality in process diagram for fault scenario 2.

shows that dE-cond-P, dE-1st-P, dE-11st-P, dE-28st-P, and dP- difference of just two sensors, compared with the SPE contribu-
19st-T are enhanced, and the other sensors are weaken by the tion and RS amplification. It can be interpreted that causality must
RS amplification method. This is a suitable result because these include all the major variables about the fault. If one key variable
sensors are related to the condenser. The proposed method, is missing, the result can be misleading, as results show. There-
fault magnitude, shows that dE-top-T, dE-cond-P, dE-1st-P, dE- fore, the fault magnitude algorithm can select hierarchical sensors
11st-P, and dE-28st-P are the major fault sensors, as shown in properly and find the accurate root cause at the initial stage of
Fig. 14. These results appear to be similar except for one or two fault.
sensors; dP-in-F(dE-bot-F) in SPE-contribution, dP-19st-T in RS-
amplification, and dE-top-T and dE-1st-P in the fault magnitude
method. 5. Conclusion
A small difference in the result of the three methodologies, pro-
duces a completely different result in the MVGC analysis. Tables 6–8 Because of multivariate statistical analysis, process fault detec-
shows the causality from the three methods, SPE-contribution, RS- tion technology has been developed rapidly. On the other hand,
amplification, and fault magnitude, respectively. fault diagnosis systems depend on specific historical fault data or
First, the SPE contribution shows that dE-cond-P affects dP- long analysis time. To solve this problem, in this study, a method
in-F(dE-bot-F), dE-28st-P affects dE-cond-P and dE-11st-P, and of selecting fault magnitude variables while considering the fault
dP-in-F(dE-bot-F) affects dE-cond-P and dE-28st-P. These causal- detection time is proposed for root-cause analysis. In particular, the
ities can be interpreted that the relationship between causality method focuses on providing an accurate root causality, even when
sensors is a bit far, broad, and crossed, such that it is ambigu- the fault has a small intensity. Using calculated variables from PCA,
ous to define the root cause. RS amplification provides the main a fault can be divided into a normal and abnormal portion. In PCS
causalities as dE-11st-P to dE-28st-P and dE-cond-P. The weak or RS, these portions can be separated by subspace to minimize the
causalities are from dE-28st-P to dE-11st-P, from dE-cond-P to effect of the normal portion on the fault information. The covari-
dE-11st-P, and from dP-19st-T to dE-28st-P. These results can be ance of the scaled fault contribution decomposed by SVD analysis
interpreted as the fault starts from the column internal pressure and hierarchical variables can be selected by a fault magnitude
problem. Finally, the fault magnitude methodology shows that dE- algorithm. MVGC calculated these major variables to generate a
top-T affects dE-1st-P primarily, and weak causalities are given matrix of causality. To verify the performance, an LNG fractionation
by dE-cond-P to dE-1st-P and dE-top-T. This result indicates that dynamic model is developed. From the normal and fault operation
the root cause starts from the condenser area. These results are data, the fault magnitude methodology can better determine the
described as a flow diagram in Fig. 15 and can be more clearly visu- proper root cause than conventional contribution analysis or RS
alized in the process diagram in Fig. 16. This result comes from the amplification can. In terms of conventional PCA, because the 33 sen-
 H. Pyun et al. / Process Safety and Environmental Protection 135 (2020) 113–125 125

sors have been reduced to two variables, the PCA model has a large Geng, Z., Zhu, Q., 2005. Multiscale Nonlinear Principal Component Analysis (NLPCA)
portion in the normal state. Therefore, as shown in results for sce- and its application for chemical process monitoring. Ind. Eng. Chem. Res. 44,
3585–3593, http://dx.doi.org/10.1021/ie0493107.
nario 1 and 2, contribution charts contain an insufficient number of Ha, D., Ahmed, U., Pyun, H., Lee, C.-J., Baek, K.H., Han, C., 2017. Multi-mode operation
sensors affected by fault, such that they cannot determine the root of principal component analysis with k-nearest neighbor algorithm to moni-
cause properly. In contrast, RS amplification exhibits good perfor- tor compressors for liquefied natural gas mixed refrigerant processes. Comput.
Chem. Eng. 106, 96–105, http://dx.doi.org/10.1016/J.COMPCHEMENG.2017.05.
mance in enhancing the contribution-affected variables. However, 029.
this method uses only RS and contributions that reflect normal He, Q.P., Qin, S.J., Wang, J., 2005. A new fault diagnosis method using fault directions
data. Therefore, it cannot perform well in the MVGC analysis for in Fisher discriminant analysis. AIChE J. 51, 555–571, http://dx.doi.org/10.1002/
aic.10325.
finding the root cause. This method can provide a misleading result
Jackson, J.E., 1980. Principal components and factor analysis: part I—principal com-
because of a failure to select the key variables of the fault. Unlike the ponents. Int. J. Qual. Assur. Eng. Technol. Educ. 12, 201–213, http://dx.doi.org/
two methodologies, the fault magnitude method shows proper root 10.1080/00224065.1980.11980967.
Jackson, J.E., Mudholkar, G.S., 1979. Control procedures for residuals associated with
cause at the initial fault stage, even with small intensity. Because
principal component analysis. Technometrics 21, 341–349, http://dx.doi.org/10.
the normal portion is removed from the fault of small intensity, 1080/00401706.1979.10489779.
affected sensors can be selected and analyzed properly in the MVGC Jeong, H., Cho, S., Kim, D., Pyun, H., Ha, D., Han, C., Kang, M., Jeong, M., Lee, S.,
method. The fault magnitude algorithm is based on just normal 2012. A heuristic method of variable selection based on principal component
analysis and factor analysis for monitoring in a 300 kW MCFC power plant.
operation data, therefore, it can be expanded for the overall process Int. J. Hydrogen Energy 37, 11394–11400, http://dx.doi.org/10.1016/J.IJHYDENE.
system. 2012.04.135.
Jiang, Q., Yan, X., 2014. Monitoring multi-mode plant-wide processes by using
mutual information-based multi-block PCA, joint probability, and Bayesian
Declaration of Competing Interest inference. Chemometr. Intell. Lab. Syst. 136, 121–137, http://dx.doi.org/10.1016/
j.chemolab.2014.05.012.
None. Joe Qin, S., 2003. Statistical process monitoring: basics and beyond. J. Chemom. 17,
480–502, http://dx.doi.org/10.1002/cem.800.
Kano, M., Hasebe, S., Hashimoto, I., Ohno, H., 2001. A new multivariate statistical
Acknowledgements process monitoring method using principal component analysis. Comput. Chem.
Eng. 25, 1103–1113, http://dx.doi.org/10.1016/S0098-1354(01)00683-4.
Karhunen, J., Pajunen, P., Oja, E., 1998. The nonlinear PCA criterion in blind source
This work was supported by the Human Resources Development separation: relations with other approaches. Neurocomputing 22, 5–20, http://
(No. 20184030202070) of the Korea Institute of Energy Technology dx.doi.org/10.1016/S0925-2312(98)00046-0.
Evaluation and Planning (KETEP) grant funded by the Korea govern- Kim, K., 2011. Modeling and Simulation of Fractionation Process in LNG Plant. Model.
Simul. fractionation Process LNG plant / Kyeongjin Kim.
ment Ministry of Trade, Industry and Energy and the Korea Agency Kitano, K., Kano, M., Gopaluni, B., 2018. Fault identification with modified
for Infrastructure Technology Advancement grant funded by the reconstruction-based contribution based on kernel principal component analy-
Ministry of Land, Infrastructure and Transport (17IFIP-B133622- sis. In: 2017 Asian Control Conference, ASCC 2017. IEEE, pp. 1493–1498, http://
dx.doi.org/10.1109/ASCC.2017.8287394.
01).
Kresta, J.V., Macgregor, J.F., Marlin, T.E., 1991. Multivariate statistical monitoring of
process operating performance. Can. J. Chem. Eng. 69, 35–47, http://dx.doi.org/
References 10.1002/cjce.5450690105.
Ku, W., Storer, R.H., Georgakis, C., 1995. Disturbance detection and isolation
by dynamic principal component analysis. Chemometr. Intell. Lab. Syst. 30,
Ahmed, U., Ha, D., An, J., Zahid, U., Han, C., 2017a. Fault propagation path estimation
179–196, http://dx.doi.org/10.1016/0169-7439(95)00076-3.
in NGL fractionation process using principal component analysis. Chemometr.
MacGregor, J.F., Kourti, T., Liu, J., Bradley, J., Dunn, K., Yu, H., 2007. Multivariate
Intell. Lab. Syst. 162, 73–82, http://dx.doi.org/10.1016/j.chemolab.2017.01.006.
methods for the analysis of data-bases, process monitoring, and control in the
Ahmed, U., Ha, D., Shin, S., Shaukat, N., Zahid, U., Han, C., 2017b. Estimation of
material processing industries. IFAC Proc. Vol. 40, 193–198, http://dx.doi.org/
disturbance propagation path using Principal Component Analysis (PCA) and
10.3182/20070821-3-CA-2919.00028.
Multivariate Granger Causality (MVGC) techniques. Ind. Eng. Chem. Res. 56,
Nan, C., Khan, F., Iqbal, M.T., 2008. Real-time fault diagnosis using knowledge-based
7260–7272, http://dx.doi.org/10.1021/acs.iecr.6b02763.
expert system. Process Saf. Environ. Prot. 86, 55–71, http://dx.doi.org/10.1016/
Bahadori, A., 2014. Natural Gas Processing: Technology and Engineering Design,
J.PSEP.2007.10.014.
Natural Gas Processing: Technology and Engineering Design. Elsevier, http://dx.
Narozny, J.A., Phohole, A.D.L., Elnir, M., 2001. An expert system for machine
doi.org/10.1016/C2013-0-13070-X.
fault diagnosis. IFAC Proc. Vol. 34, 95–99, http://dx.doi.org/10.1016/S1474-
Barnett, L., Seth, A.K., 2014. The MVGC multivariate Granger causality toolbox: a new
6670(17)33188-9.
approach to Granger-causal inference. J. Neurosci. Methods 223, 50–68, http://
Qin, S.J., 2012. Survey on data-driven industrial process monitoring and diagnosis.
dx.doi.org/10.1016/J.JNEUMETH.2013.10.018.
Annu. Rev. Control 36, 220–234, http://dx.doi.org/10.1016/j.arcontrol.2012.09.
Chehreghani, M.H., 2017. Classification with Minimax Distance Measures. Aaai., pp.
004.
1784–1790.
Santen, A., Koot, G.L.M., Zullo, L.C., 2003. Statistical data analysis of a chemical
Chen, H.S., Yan, Z., Zhang, X., Liu, Y., Yao, Y., 2018. Root cause diagnosis of process
plant. Comput. Chem. Eng. 21, S1123–S1129, http://dx.doi.org/10.1016/s0098-
faults using conditional granger causality analysis and maximum spanning tree.
1354(97)87653-3.
IFAC-PapersOnLine 51, 381–386, http://dx.doi.org/10.1016/j.ifacol.2018.09.330.
Stichlmair, J., Fair, J.R., 1998. Distillation: Principles and Practices. Wiley.
Choi, S.W., Lee, C., Lee, J.-M., Park, J.H., Lee, I.-B., 2005. Fault detection and identifi-
Venkatasubramanian, V., Rengaswamy, R., Kavuri, S.N., 2003a. A review of pro-
cation of nonlinear processes based on kernel PCA. Chemometr. Intell. Lab. Syst.
cess fault detection and diagnosis part II: qualitative models and search
75, 55–67, http://dx.doi.org/10.1016/J.CHEMOLAB.2004.05.001.
strategies. Comput. Chem. Eng. 27, 313–326, http://dx.doi.org/10.1016/S0098-
Ding, M., Chen, Y., Bressler, S.L., 2006. Granger causality: basic theory and application
1354(02)00161-8.
to neuroscience. In: Handbook of Time Series Analysis: Recent Theoreti-
Venkatasubramanian, V., Rengaswamy, R., Kavuri, S.N., Yin, K., 2003b. A review of
cal Developments and Applications., pp. 437–460, http://dx.doi.org/10.1002/
process fault detection and diagnosis. Comput. Chem. Eng. 27, 327–346, http://
9783527609970.ch17.
dx.doi.org/10.1016/s0098-1354(02)00162-x.
Dudzic, M., Miletic, I., 2002. Multivariate statistical monitoring of a continuous
Wen, X., Rangarajan, G., Ding, M., 2013. Multivariate Granger causality: an estima-
steel slab caster. In: Proceedings of the American Control Conference, American
tion framework based on factorization of the spectral density matrix. Philos.
Automatic Control Council, pp. 600–601, http://dx.doi.org/10.1109/ACC.2002.
Trans. R. Soc. A Math. Phys. Eng. Sci. 371, http://dx.doi.org/10.1098/rsta.2011.
1024872.
0610.
Dudzic, M., Vaculik, V., Miletic, I., 1999. Applications of multivariate statistics at
Yuan, T., Qin, S.J., 2014. Root cause diagnosis of plant-wide oscillations using Granger
Dofasco. In: Record of Workshop Papers - IEEE Industry Applications Society
causality. J. Process Control 24, 450–459, http://dx.doi.org/10.1016/j.jprocont.
(IAS): 1999 Advanced Process Control Applications for Industry Workshop, IEEE,
2013.11.009.
pp. 27–29, http://dx.doi.org/10.1109/APCA.1999.805022.
Yue, H.H., Qin, S.J., 2001. Reconstruction-based fault identification using a com-
Dvorak, D., Kuipers, B., 1991. Process monitoring and diagnosis: a model-based
bined index. Ind. Eng. Chem. Res. 40, 4403–4414, http://dx.doi.org/10.1021/
approach. IEEE Expert. Syst. their Appl. 6, 67–74, http://dx.doi.org/10.1109/64.
ie000141+.
87688.
Zhu, Z., Song, Z., Palazoglu, A., 2012. Process pattern construction and multi-mode
Gabbar, H.A., Hussain, S., Hosseini, A.H., 2014. Simulation-based fault propagation
monitoring. J. Process Control 22, 247–262, http://dx.doi.org/10.1016/j.jprocont.
analysis—application on hydrogen production plant. Process Saf. Environ. Prot.
2011.08.002.
92, 723–731, http://dx.doi.org/10.1016/J.PSEP.2013.12.006.