Control Engineering Practice 116 (2021) 104916

Contents lists available at ScienceDirect

Control Engineering Practice

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

Fault detection and diagnosis in multivariate systems using multiple

correlation regression
Zhenming Li a , Shiyi Bao b , Xin Peng c , Lijia Luo b ,∗
a College of Chemical Engineering, Zhejiang University of Technology, Hangzhou 310023, China
b
College of Mechanical Engineering, Zhejiang University of Technology, Hangzhou 310023, China
c
Key Laboratory of Smart Manufacturing in Energy Chemical Process, Ministry of Education, East China University of Science and
Technology, Shanghai 200237, China

ARTICLE INFO ABSTRACT

Keywords: Correlations among variables are inherent features of a multivariate system. Making a good use of variable
Multivariate system correlations is important to fault detection and diagnosis (FDD). In this paper, a novel FDD method is proposed
Multiple correlation regression using multiple correlation regression (MCR). MCR builds a regression model using the multiple correlation of
Variable correlation
a variable with other variables. MCR is the best linear regression that produces residuals with the minimal
Fault detection
variance. Based on the MCR residuals of variables, a 𝑇 2 statistic is defined for fault detection. This 𝑇 2 statistic
Fault diagnosis
consists of a set of independent components, with each component corresponding to the MCR residual of a
variable. To reduce the amplification and masking effects caused by fault-free components, three improvements
to the residual-based 𝑇 2 statistic are proposed, called the intersecting 𝑇 2 (IT), weighted 𝑇 2 (WT) and enhanced
𝑇 2 (ET) statistics. In particular, the WT and ET statistics and their control limits vary with samples to adapt
to dynamic changes of the system. An online fault detection strategy is proposed by combining the WT and
ET statistics. The WT statistic is firstly used to detect the occurrence of faults. After faults were detected,
the ET statistic is used to replace the WT statistic in order to enhance the capability of detecting subsequent
faulty samples. A contribution-based fault diagnosis method is also developed. Faulty variables are identified
by comparing variable contributions to the key faulty component of the residual-based 𝑇 2 statistic. The
implementation and advantages of the proposed FDD method are illustrated with a simulation example and
an industrial case study.

1. Introduction well as data storage and analysis techniques. Data-driven FDD iden-
tifies abnormal events mainly by analyzing the operation data of an
The safety, high-quality, high-efficiency are always important issues industrial system, while it requires little engineering knowledge about
of industrial systems. As industrial systems are increasingly sophisti- the system. This makes data-driven FDD methods easy to implement in
cated and operated in more severe conditions, their operational perfor- practical applications, especially for large-scale complex systems. Due
mance is seriously threatened by abnormal events, referred to as faults. to this outstanding advantage, data-driven FDD is attracting more and
Online fault detection and diagnosis (FDD) is an effective way to cope more attention.
with system faults. By detecting the occurrence of faults and diagnosing The key problem of data-driven FDD is to differentiate the fault
the cause of faults in a timely manner, it provides operators a chance to data from the normal data. This problem is often solved by making use
take pre-emptive measures to eliminate faults before they cause serious of data modeling and statistical analysis methods. At first, important
system damage and great economic loss. FDD can be implemented feature information of the data is extracted by data modeling. The
in three different ways of using the mathematical model, engineering extracted feature information is then used to construct statistical in-
knowledge and operation data of an industrial system respectively, dices to measure statistical properties of the data. Finally, fault data are
corresponding to the model-based, knowledge-based and data-driven identified by comparing their statistical indices with the corresponding
FDD methods (Ge, 2017; MacGregor & Cinar, 2012; Park et al., 2020; control limits that are determined from the normal data. To date,
Qin, 2012). Recent years have seen the emergence of data-driven FDD various methods have been used for data modeling in the field of FDD.
methods in a wide range of industries (Ge, 2017; Jiang et al., 2019; Among the most popular there are principal component analysis (PCA)
Yin et al., 2014), thanks to advancements of the sensor technology as (Dong & Qin, 2018), canonical correlation analysis (CCA) (Peng et al.,

∗ Corresponding author.
E-mail address: [email protected] (L. Luo).

https://doi.org/10.1016/j.conengprac.2021.104916
Received 20 January 2021; Received in revised form 5 August 2021; Accepted 5 August 2021
Available online xxxx
0967-0661/© 2021 Elsevier Ltd. All rights reserved.
Z. Li, S. Bao, X. Peng et al. Control Engineering Practice 116 (2021) 104916

2020; Song et al., 2020), independent component analysis (ICA) (Yu effects caused by the fault-free components, three improvements to
et al., 2018), and slow feature analysis (SFA) (Shang et al., 2018). the residual-based 𝑇 2 statistic are developed, called the intersecting 𝑇 2
These methods focus on different aspects of data features, such as the (IT), weighted 𝑇 2 (WT) and enhanced 𝑇 2 (ET) statistics, respectively.
variance, correlation, independence and slowly time-varying features. In particular, the WT and ET statistics and their control limits can
They produce a projection matrix to transform data from the original vary with samples to adapt to dynamic changes of the multivariate
variable space to a feature space that is formed by combinations system. An online fault detection strategy is proposed by combining
of original variables. Due to the information redundancy caused by the WT and ET statistics. A contribution-based method is proposed for
the dependence among variables, the most important data features fault diagnosis. For a faulty sample, the component that has the largest
are often contained in a low-dimensional feature subspace. Statistical contribution to the residual-based 𝑇 2 value is considered as a key faulty
indices are then defined in the feature subspace to implement FDD. component, and the variables with larger contributions to this key
Some statistical indices frequently used for FDD include the 𝑇 2 statistic faulty component are identified as faulty variables. The performance
(Mason & Young, 2002), Q statistic (Nomikos & MacGregor, 1995), | S| of the proposed FDD method is illustrated with two case studies.
(generalized variance) statistic (Aparisi et al., 1999), likelihood ratio The rest of this paper is organized as follows. Section 2 presents brief
(Fezai et al., 2020) and Kullback–Leibler divergence (Xie et al., 2015). descriptions of the multiple correlation regression (MCR) and the 𝑇 2
They measure the discrepancy between normal data and fault data statistic. Section 3 introduces the MCR-based FDD method and its com-
in terms of mean, variance, covariance and probability distributions. ponents: fault detection indices, the contribution-based fault diagnosis
Of these, the 𝑇 2 statistic is the most widely used statistical index for method, and an online FDD strategy. Section 4 illustrates advantages
detecting faults that cause deviations in the mean of data (Mason & of the proposed FDD method using two case studies. Conclusions are
Young, 2002). given in Section 5.
An industrial system often contains a large number of variables.
These variables are interrelated due to the mass, energy and informa- 2. Background methods
tion flow within the system. Variable correlations are inherent features
of an industrial system and thus are very useful to FDD. A good mod- 2.1. Multiple correlation regression
eling of variable correlations is therefore a prerequisite of successful
FDD. However, only a good variable correlation model is insufficient Multiple correlation regression (MCR) builds a regression model
to achieve high-performance FDD. It is important to note that fault using the multiple correlation coefficient of variables. For m variables,
scenarios in a multivariate system are complex and changeable. The the multiple correlation coefficient between the kth variable and the
number, distribution and magnitude of faulty variables all vary with other (𝑚 − 1) variables is defined as (Anderson, 2003)
fault scenarios. This requires that statistical indices for FDD can adapt √
√ √ 𝑇 −1
to various fault scenarios. Moreover, it is often the case that faults in 𝜷 𝑇𝑘 𝜮 𝑘 𝜷 𝑘 𝜷 𝑇𝑘 𝜮 𝑘 𝜷 𝑘 √√ 𝝈𝑘 𝜮 𝑘 𝝈𝑘
a multivariate system are caused by the abnormalities of only a few 𝑅𝑘 = √ = = (1)
𝜎𝑘 𝜎𝑘
variables. In this case, a statistical index that contains information of all 𝜎𝑘 𝜷 𝑇𝑘 𝜮 𝑘 𝜷 𝑘
variables may suffer from the amplification and masking effects caused
where 𝜎𝑘 is the variance of the kth variable, 𝜮 𝑘 denotes the covariance
by fault-free variables (Luo, Xu, et al., 2019). Due to the amplification
matrix of the other (𝑚 − 1) variables, 𝝈 𝑘 is a vector that consists of
and masking effects, a statistical index may not exceed the control
covariances between the kth variable and the other (𝑚 − 1) variables,
limit, even though a part of variables is faulty. This reduces the fault { }
and 𝜷 𝑘 = 𝜮 −1 𝝈 𝑘 . Given n samples 𝒙1 , 𝒙2 , … , 𝒙𝑛 of m variables, the
detection capability. Therefore, to get superior FDD performance, a 𝑘
good data model and advanced statistical indices are both needed. 𝜎𝑘 , 𝝈 𝑘 and 𝜮 𝑘 are estimated by
In the past few decades, methods for modeling the multivariate data ∑𝑛 ( )( )𝑇 [ ]
̂ 𝑖=1 𝒙𝑖 − 𝒙 𝒙𝑖 − 𝒙 𝜎̂ 𝑘 𝝈̂ 𝑘
have received huge attention. As mentioned above, the PCA, CCA and 𝜮= = (2)
𝑛−1 𝝈̂ 𝑇 𝜮 ̂
𝑘
ICA have been applied to build data models for FDD. In addition, some 𝑘
nonlinear, dynamic, sparse and probabilistic modeling methods have where 𝜮 ̂ is the estimation of the covariance matrix of m variables, and
been applied to FDD (He & Zhang, 2020; Luo, Bao, & Tong, 2019; ̂ −1 𝝈̂ . A
𝒙 is the mean of n samples. The 𝜷 𝑘 is thus estimated by 𝜷̂ 𝑘 = 𝜮 𝑘 𝑘
Yu & Qin, 2009; Zhang et al., 2020). There also are some methods MCR model is then built for the kth variable
to build multi-block or multi-group models for FDD by dividing a
𝑇
multivariate system into meaningful blocks or groups (Luo et al., 2021, 𝑥̂ 𝑘 = 𝑥𝑘 + 𝜷̂ 𝑘 (𝒙𝑘 − 𝒙𝑘 ) (3)
2020; MacGregor et al., 1994; Wan et al., 2020; Yan et al., 2020).
where 𝑥̂ 𝑘 is the prediction value of the kth variable, 𝒙𝑘 are values of
However, little attention has been paid to the problem of designing
other (𝑚 − 1) variables, 𝑥𝑘 and 𝒙𝑘 are mean values of the kth variable
advanced statistical indices that can adapt to different fault scenarios
and other (𝑚 − 1) variables respectively. The MCR is equivalent to the
as well as reduce the amplification and masking effects. When the data
least square regression, because
model is built by a latent-variable method (e.g., PCA), several methods
𝑇
have been proposed to select appropriate latent variables (e.g., PCs) 𝑥̂ 𝑘 − 𝑥𝑘 = 𝜷̂ 𝑘 (𝒙𝑘 − 𝒙𝑘 )
for the 𝑇 2 statistic to improve fault detection performance (Luo, Bao, & ( −1 )𝑇
⇔ 𝑥̃ 𝑘 = 𝜮 ̂ 𝝈̂ 𝒙̃ 𝑘
Mao, 2019; Valle et al., 1999). Recently, Luo, Xu, et al. (2019) proposed 𝑘 𝑘 (4)
(( )−1 )𝑇
improvements on the 𝑇 2 statistic to gain the adaptive ability and the 𝑇
̃ 𝑿 ̃ ̃ 𝑇 𝒙̃ 𝑘
⇔ 𝑥̃ 𝑘 = 𝑿 𝑘 𝑘
𝑿 𝑘
𝒙̃ 𝑘
ability to reduce the amplification and masking effects. However, the
improvements were made only on the 𝑇 2 statistic itself, and there is where 𝑥̃ 𝑘 and 𝒙̃ 𝑘 denote mean-centered values of m variables, 𝒙̃ 𝑘 is
still a lack of a supporting data modeling method. a vector that consists of n samples of the kth variable, and 𝑿 ̃ is a
𝑘
In this paper, a novel FDD method for multivariate systems is matrix consisting of n samples of other (𝑚 − 1) variables. Note that
proposed by using the multiple correlation regression (MCR). MCR ( )−1
̃ 𝑇𝑿
𝑿 ̃ ̃ 𝑇 𝒙̃ 𝑘 are exactly regression coefficients of the least square
𝑿
builds a regression model for a variable using its multiple correlations 𝑘 𝑘 𝑘
with other variables. Based on the MCR residuals of variables, a 𝑇 2 method. This indicates that the MCR is the best linear regression in the
statistic is defined for fault detection. Each component of this 𝑇 2 sense that the sum of the squares of prediction errors is minimal.
statistic corresponds to the MCR residual of a variable. To minimize the Let 𝑥𝑘 be a measurement of the kth variable. A residual model is
control region of the residual-based 𝑇 2 statistic, MCR models are built built as
for all variables in sequence according to a maximal multiple correla- 𝑟𝑘 = 𝑥𝑘 − 𝑥̂ 𝑘
𝑇 (5)
tion coefficient rule. To further reduce the amplification and masking = 𝑥𝑘 − 𝑥𝑘 − 𝜷̂ 𝑘 (𝒙𝑘 − 𝒙𝑘 )

2
Z. Li, S. Bao, X. Peng et al. Control Engineering Practice 116 (2021) 104916

The variance of 𝑟𝑘 is computed by Second, a contribution-based fault diagnosis method is proposed, which
[ ]2 identifies faulty variables according to variable contributions to fault
1 ∑𝑛 ̂𝑇
𝜎𝑟𝑘 = 𝑛−1 𝑖=1 𝑥𝑘,𝑖 − 𝑥𝑘,𝑖 − 𝜷 𝑘 (𝒙𝑘,𝑖 − 𝒙𝑘,𝑖 ) detection. Third, a strategy is proposed to perform online FDD for a
𝑇 multivariate system.
= 𝜎̂ 𝑘 − 2𝝈̂ 𝑇 𝜷̂ 𝑘 + 𝜷̂ 𝑘 𝜮
̂ 𝜷̂ 𝑘
𝑘 (6)
𝑘
̂ −1 𝝈̂
= 𝜎̂ 𝑘 − 𝝈̂ 𝑇 𝜮
𝑘 𝑘 𝑘 3.1. Fault detection indices
= (1 − 𝑅2𝑘 )𝜎̂ 𝑘
It can be seen from Eq. (6) that the multiple correlation coefficient 𝑅𝑘 Consider a multivariate system of m variables. A MCR model can
represents the fraction of the variance explained by the other (𝑚 − 1) be built for every variable using its multiple correlation with other
variables. The larger 𝑅𝑘 is, the smaller 𝜎𝑟𝑘 is, and thus the higher the (𝑚 − 1) variables. However, this produces m correlated residuals, that
prediction accuracy of the MCR model is. The covariance between the is, m MCR residuals contain redundant information. To eliminate the
residual 𝑟𝑘 and other (𝑚 − 1) variables is computed by redundant information, MCR models are built for m variables in a
( ) [ ] sequential manner. In the sequential modeling procedure, if a MCR
1 ∑𝑛 ̂𝑇 𝑇
𝑐𝑜𝑣 𝑟𝑘 , 𝒙𝑘 = 𝑛−1 𝑖=1 𝑥𝑘,𝑖 − 𝑥𝑘,𝑖 − 𝜷 𝑘 (𝒙𝑘,𝑖 − 𝒙𝑘,𝑖 ) (𝒙𝑘,𝑖 − 𝒙𝑘,𝑖 ) model has been built for a variable, this variable will no longer be
̂ −𝑇 𝜮
= 𝝈̂ 𝑇 − 𝝈̂ 𝑇 𝜮 ̂ (7) used in building MCR models for other variables. Assume the modeling
𝑘 𝑘 𝑘 𝑘
=𝟎 sequence of variables is from 1 to m. The MCR model of the kth variable
is built using variables from k+1 to m. By this way, the residuals of m
This implies that the residual 𝑟𝑘 is uncorrelated with the other (𝑚 − 1) variables are not correlated, because the residual of the kth variable
variables. is uncorrelated with the subsequent (m-k) variables (see Eq.(7)). Let
𝜎𝑟𝑘 denote the variance of the kth residual. The covariance matrix
2.2. The 𝑇 2 statistic of m residuals is a diagonal matrix that consists of variances of m
residuals, denoted as 𝜮 𝑟 = 𝑑𝑖𝑎𝑔(𝜎𝑟1 , 𝜎𝑟2 , … , 𝜎𝑟𝑚 ). Note that the variable
The 𝑇 2 statistic is widely used for monitoring the mean shifts of modeling sequence is important, because it affects the variance of each
a multivariate system. Let 𝜮 and 𝜇 denote the covariance matrix and residual and hence the 𝑇 2 statistic defined by m residuals. According
means of m variables, the 𝑇 2 statistic can be defined as (Mason & to Eq. (11), to minimize the control region of the 𝑇 2 statistic, the
Young, 2002) ∏𝑚
determinant of the matrix 𝜮 𝑟 , i.e., 𝑘=1 𝜎𝑟𝑘 , should be minimized.
When the number of variables is large, it is very time-consuming to
𝑇 2 = (𝒙 − 𝝁)𝑇 𝜮 −1 (𝒙 − 𝝁) (8)
find an optimal sequence of variable modeling that yields the smallest
∏𝑚
where x is a sample of m variables. Note that the 𝑇 2 statistic is 𝑘=1 𝜎𝑟𝑘 . To address this problem, it is suggested to minimize every
the squared Mahalanobis distance, which takes into account variable 𝜎𝑟𝑘 . Assume all variables have been scaled to have variance one. The
correlations measured by the covariance matrix 𝜮. This is advantageous 𝜎𝑟𝑘 of the kth residual is 𝜎𝑟𝑘 = (1 − 𝑅2𝑘 ) (see Eq. (6)), where 𝑅𝑘 is the
to fault detection, because variable correlations are important features multiple correlation coefficient of the kth variable with the subsequent
of a multivariate system. If m variables follow a normal distribution, (m-k) variables. Thus, to minimize 𝜎𝑟𝑘 , the kth MCR model should
the 𝑇 2 statistic follows a 𝜒 2 distribution with m degrees of freedom, be built for the variable that has the largest 𝑅𝑘 . According to this
and its control limit at significance level of 𝛼 is computed by (Mason maximal multiple correlation coefficient rule, it is easy to determine
& Young, 2002) the modeling sequence of variables. This sequence can be denoted as
{ }
1𝑠 , 2𝑠 , … , 𝑚𝑠 , with 𝑅21 ≥ 𝑅22 ≥ ⋯ ≥ 𝑅2𝑚 . Then, a residual-based 𝑇 2
𝑇𝑐2 = 𝜒1−𝛼
2
(𝑚) (9) statistic is defined as
𝑠 𝑠 𝑠

( ) ( )
The control region of the 𝑇 2 statistic is an ellipsoid in the m-dimensional 𝑇𝑟2 = 𝒓 − 𝒓 𝜮 −1
𝑇
𝒓−𝒓
𝑟
space, given by ∑𝑚𝑠 (𝑟𝑘 −𝑟𝑘 )2 (12)
= 𝑘=1 𝜎
𝑠
(𝒙 − 𝝁)𝑇 𝜮 −1 (𝒙 − 𝝁) ≤ 𝑇𝑐2 𝑟𝑘

⇔ (𝒙 − 𝝁)𝑇 (𝑷 𝜦𝑷 𝑇 )−1 (𝒙 − 𝝁) ≤ 𝑇𝑐2 where r is a vector consisting of MCR residuals of m variables, and 𝒓
⇔ 𝒚 𝑇 𝜦−1 𝒚 ≤ 𝑇𝑐2 (10) and 𝜮 𝑟 are the mean vector and covariance matrix of r. The control
∑ 𝑦2𝑖 limit of 𝑇𝑟2 is computed from the 𝜒 2 (𝑚) distribution.
⇔ 𝑚 𝑖=1 𝜆 ≤ 𝑇𝑐
2
𝑖 Eq. (12) shows that the 𝑇𝑟2 statistic is the summation of m inde-
where 𝜮 = 𝑷 𝜦𝑷 𝑇 is a singular value decomposition, 𝜦 is a diagonal (𝑟 −𝑟 )2
pendent components, where the kth component 𝑘𝜎 𝑘 represents the
matrix consisting of eigenvalues 𝜆1 , . . . , 𝜆𝑚 , and 𝒚 = 𝑷 𝑇 (𝒙 − 𝝁). The 𝑟𝑘
contribution of the kth residual to 𝑇𝑟2 . When only a few components
volume of this ellipsoid is (Luo, Xu, et al., 2019)
are large but many other components are small, the 𝑇𝑟2 statistic may
𝜋2
𝑚
(∏𝑚 ) 1 ( 2) 𝑚 suffer from the amplification and masking effects caused by those small
𝑉 = ( ) 𝜆𝑖 2 𝑇 2
𝑖=1 𝑐
𝛤 𝑚2
+1 components. To reduce the amplification and masking effects, Luo, Xu,
𝑚
√ ( ) (11)
(
𝜋2
) det(𝜮) 𝑇 2 2
𝑚
et al. (2019) have proposed three improvements to the 𝑇 2 statistic: an
= 𝑐
𝛤 𝑚2
+1 adaptive dimension reduction method, the weighted 𝑇 2 (WT) statistic,
and the intersecting 𝑇 2 (IT) statistic. The ideas of intersecting and
where 𝛤 (⋅) denotes the gamma function, and det(𝜮) is the determinant
weighted 𝑇 2 statistics are used in this paper. At first, m components
of the matrix 𝜮. The smaller the det(𝜮) is, the smaller the control region
of 𝑇𝑟2 are divided into two parts
is, and the better the overall fault detection capability of the 𝑇 2 statistic
will be. It has been proved that the det(𝜮) decreases as correlations 2 = ∑𝑙 (𝑟𝑘 −𝑟𝑘 )2
𝑇𝑟,𝑝 𝑘=1𝑠 𝜎𝑟
among m variables increase (Yan et al., 2020). 1
𝑘 (13)
2 ∑𝑚𝑠 (𝑟𝑘 −𝑟𝑘 )2
𝑇𝑟,𝑝 = 𝑘=𝑙+1 𝜎𝑟
2
3. Fault detection and diagnosis using MCR 𝑘

where l is an index after which the 𝑅2𝑘 is small. Clearly, the 𝑇𝑟,𝑝
2 part
1
Based on the MCR, a fault detection and diagnosis (FDD) method for contains most of information on variable correlations, due to the larger
multivariate systems is developed in this section. First, fault detection 𝑅2𝑘 of the first l components. The 𝑇𝑟,𝑝
2 part contains the information
2
indices are defined using MCR residuals of variables, and methods of ‘remaining’ variables. Main correlations of the ‘remaining’ variables
for computing control limits of fault detection indices are presented. 2 part, and therefore
with other variables have been captured by the 𝑇𝑟,𝑝
1

3
Z. Li, S. Bao, X. Peng et al. Control Engineering Practice 116 (2021) 104916

2
correlations among the ‘remaining’ variables are small. The 𝑇𝑟,𝑝 and where 𝑥̃ 𝑘 = 𝑥𝑘 − 𝑥𝑘 , 𝒙̃ 𝑘 = 𝒙𝑘 − 𝒙𝑘 , 𝑥̃ 𝑗,𝑘 and 𝛽̂𝑗,𝑘 are the 𝑗th elements of
1
2
𝑇𝑟,𝑝 parts can be used separately for fault detection, and then their
2
𝒙̃ 𝑘 and 𝜷̂ 𝑘 respectively, and 𝛾𝑗ℎ,𝑘 is the (j, h)th element of the matrix
𝑇
detection results are combined by 𝜷̂ 𝑘 𝜷̂ . The contribution of the kth variable 2 is
𝑣𝑘𝑠 to 𝑇𝑟,𝑘
𝑘
( ) ( )
2 2 2 2 ( )
IT (𝒓is normal) ⇔ 𝑇𝑟,𝑝 ≤ 𝑇 𝑟,𝑝 ∩ 𝑇𝑟,𝑝 ≤ 𝑇 𝑟,𝑝 (14) ∑
𝑚−𝑘
1 1 2 2 1
𝑐𝑘,𝑘 = 𝑥̃ 2𝑘 − 𝑥̃ 𝑘 𝛽̂𝑗,𝑘 𝑥̃ 𝑗,𝑘 (21)
2 2 2 2
𝜎𝑟 𝑘 𝑗=1
where 𝑇 𝑟,𝑝 and 𝑇 𝑟,𝑝 denote control limits of 𝑇𝑟,𝑝 and 𝑇𝑟,𝑝 that are
1 1 1 2
2 2
determined from the distributions 𝜒 (𝑙) and 𝜒 (𝑚 − 𝑙) respectively. The The contribution of} the 𝑗th one of subsequent variables
{
statistic defined in Eq. (14) is named as the intersecting 𝑇 2 (IT) statistic, 2 is
𝑣(𝑘+1)𝑠 , 𝑣(𝑘+2)𝑠 , … , 𝑣𝑚𝑠 to 𝑇𝑟,𝑘
2 and 𝑇 2 are both in control limits; otherwise,
that is, r is normal if 𝑇𝑟,𝑝
1 𝑟,𝑝2 ( )
r is faulty. 1 ∑
𝑚−𝑘

2 and 𝑇 2 can also be combined in the form of the weighted 𝑐𝑗,𝑘 = 𝑥̃ 𝑗,𝑘 𝛾𝑗ℎ,𝑘 𝑥̃ ℎ,𝑘 − 𝑥̃ 𝑘 𝛽̂𝑗,𝑘 𝑥̃ 𝑗,𝑘 (22)
The 𝑇𝑟,𝑝 𝑟,𝑝2 𝜎𝑟 𝑘
1 ℎ=1
𝑇 2 (WT) statistic, given by
For the 𝑇𝑟2 value of a faulty sample, the component with the largest
𝑊 𝑇𝑟2 = 𝜃1 𝑇𝑟,𝑝
2 + 𝜃 𝑇2
2 𝑟,𝑝2 contribution is regarded as the key faulty component. In this key faulty
1
∑ 𝜃 (𝑟 −𝑟 )
2 ∑𝑚𝑠 2
𝜃2 (𝑟𝑘 −𝑟𝑘 ) component, the variable with the largest contribution is identified as
= 𝑙𝑘=1 1 𝜎𝑘 𝑘 + 𝑘=𝑙+1 𝜎𝑟
(15)
𝑠
( )𝑇 𝑟𝑘 −1 ( ) 𝑘 the faulty variable that is responsible for the detected faulty sample.
= 𝒓 − 𝒓 𝜣𝜮 𝑟 𝒓 − 𝒓
where 𝜃1 and 𝜃2 are two weight coefficients, and 𝜣 is a diagonal weight 3.3. Online fault detection and diagnosis strategy
matrix consisting of 𝜃1 and 𝜃2 . Based on results of Box (1954), the 𝑊 𝑇𝑟2
approximately follows a weighted 𝜒 2 distribution For a new sample x, the 𝑊 𝑇𝑟2 value is computed by
2 2
𝑊 𝑇𝑟2 ∼ 𝑔𝜒 2 (ℎ) (16) 2
𝑊 𝑇𝑟,𝑥 = 𝜃1,𝑥 𝑇𝑟,𝑝 + 𝜃2,𝑥 𝑇𝑟,𝑝 (23)
1 ,𝑥 2 ,𝑥

where two parameters g and h are computed by Box (1954) 2

where 𝑇𝑟,𝑝 2
and 𝑇𝑟,𝑝 are computed via Eq. (13) using m MCR
1 ,𝑥 2 ,𝑥

𝑙𝜃12 + (𝑚 − 𝑙)𝜃22 residuals of x, and 𝜃1,𝑥 and 𝜃2,𝑥 are given by

𝑡𝑟(𝜮 𝑟 𝜣𝜮 −1
𝑟 )
2
𝑔= = (17) 2
𝑇𝑟,𝑝
𝑡𝑟(𝜮 𝑟 𝜣𝜮 −1
𝑟 )
𝑙𝜃1 + (𝑚 − 𝑙)𝜃2
𝜃1,𝑥 = 1 ,𝑥
[ ]
−1 2
[ ]2 2
𝑇𝑟,𝑝
1 ,𝑥
2
+𝑇𝑟,𝑝
2 ,𝑥 (24)
𝑡𝑟(𝜮 𝑟 𝜣𝜮 𝑟 ) 𝑙𝜃1 + (𝑚 − 𝑙)𝜃2 2
𝑇𝑟,𝑝 ,𝑥
ℎ= = (18) 𝜃2,𝑥 = 2
2
2
𝑡𝑟(𝜮 𝑟 𝜣𝜮 −1
𝑟 )
2 𝑙𝜃12 + (𝑚 − 𝑙)𝜃22 𝑇𝑟,𝑝
1 ,𝑥
+𝑇𝑟,𝑝
2 ,𝑥

Note that 𝑇𝑟,𝑝 2 2

is important than 𝑇𝑟,𝑝 , because it contains most of
The control limit of 𝑊 𝑇𝑟2 at the significance level 𝛼 is determined 1 ,𝑥 2 ,𝑥
2 (ℎ). Given 𝜃 , 𝜃 and 𝑙 = 𝑚∕2, the value of 𝑊 𝑇 2 may change if information on variable correlations. In most cases, 𝑇𝑟,𝑝 2 is larger
as 𝑔𝜒1−𝛼 1 2 𝑟 1 ,𝑥
2 +𝜃 𝑇 2 , but the control limit of 𝑊 𝑇 2 is 2
than 𝑇𝑟,𝑝 ,𝑥 , and 𝜃1,𝑥 is also larger than 𝜃2,𝑥 . Only when faults have
Eq. (15) is changed to be 𝜃2 𝑇𝑟,𝑝 1 𝑟,𝑝2 𝑟 2
1
2 2
fixed. Therefore, to enhance the fault detection capability of the 𝑊 𝑇𝑟2 larger effects on variables corresponding to 𝑇𝑟,𝑝
2 ,𝑥
, the 𝑇𝑟,𝑝
2 ,𝑥
may be
statistic, the important part should have a larger weight coefficient. 2
larger than 𝑇𝑟,𝑝 ,𝑥 , and hence 𝜃2,𝑥 is larger than 𝜃1,𝑥 . Thus, in the 𝑊 𝑇𝑟,𝑥 2
1
2
Clearly, the 𝑇𝑟,𝑝 part is more important than the 𝑇𝑟,𝑝2 part, because 2
statistic, the contribution of a more important part (either 𝑇𝑟,𝑝 or
1 2 1 ,𝑥
2
most of variable correlations is captured by 𝑇𝑟,𝑝 . A remaining problem 2
𝑇𝑟,𝑝 ) is further magnified by adding a larger weight coefficient. This
1 2 ,𝑥
is how to choose the values of 𝜃1 and 𝜃2 . To solve this problem, a makes the 𝑊 𝑇𝑟,𝑥 2 statistic have better fault detection performance than
method to determine 𝜃1 and 𝜃2 adaptively for every sample will be 2
the 𝑇𝑟 statistic for most of faults. The control limit of 𝑊 𝑇𝑟,𝑥 2 is
present in the online fault detection and diagnosis section. computed from the weighted 𝜒 2 distribution in Eq. (16) by substituting
𝜃1,𝑥 and 𝜃2,𝑥 into Eqs. (17) and (18). It is noteworthy that both the 𝑊 𝑇𝑟,𝑥 2
3.2. Fault diagnosis method value and its control limit vary with the sample x.
Faults are detected if the 𝑊 𝑇𝑟,𝑥 2 values of several successive sam-
After a fault was detected, faulty variables responsible for the fault ples all exceed their control limits. After faults were detected, basic
are identified according to variable contributions to the 𝑇𝑟2 value. information about faults can be obtained by fault diagnosis. Such
Although the 𝑊 𝑇𝑟2 statistic is better than the 𝑇𝑟2 statistic for fault fault information can in turn be used to improve the capability of
detection, it is not used for fault diagnosis. This is because the weight detecting subsequent faulty samples. Based on this idea, an enhanced
coefficients in 𝑊 𝑇𝑟2 change the original variable contributions, which 𝑇 2 (ET) statistic is defined and used to replace the 𝑊 𝑇𝑟,𝑥 2 statistic for

may lead to inaccurate fault diagnosis results. It is important to note monitoring subsequent samples. For a sample x collected during the
2
that the MCR models, corresponding to m components
{ of the 𝑇
}𝑟 statis- faulty operating period, the 𝑇𝑟2 value is computed by Eq. (12). The
tic, are built for variables in the sequence of 𝑣1𝑠 , 𝑣2𝑠 , … , 𝑣𝑚𝑠 , where contribution of the kth component to the 𝑇𝑟2 value of x is
the kth MCR model is built for the kth variable 𝑣𝑘𝑠 using variables ( )2
𝑟𝑘,𝑥 − 𝑟𝑘
from 𝑣(𝑘+1)𝑠 to 𝑣𝑚𝑠 . According to Eq. (12), the contribution of the kth 𝑐𝑘,𝑥 = (25)
component to the 𝑇𝑟2 value is 𝜎𝑟 𝑘
( )2 The component with the largest contribution obviously contains the
𝑟𝑘 − 𝑟𝑘
𝑐𝑘 = (19) key information about the fault, and hence it is a key component that
𝜎𝑟𝑘 deserves more attention. Considering dynamic effects of faults, key
The kth component of the 𝑇𝑟2 statistic is further decomposed as components of l samples collected before x are also taken into account.
The indices of these l +1 components constitute a set 𝛺𝑥 . The ET
2 = (𝑟𝑘 −𝑟𝑘 )2 statistic is then defined as
𝑇𝑟,𝑘 𝜎𝑟
[ 𝑘 𝑇
]2 2 =∑
𝐸𝑇𝑟,𝑥
= 1
𝑥𝑘 − 𝑥𝑘 − 𝜷̂ 𝑘 (𝒙𝑘 − 𝒙𝑘 ) 𝑘∈𝛺𝑥 𝑐𝑘,𝑥 (26)
𝜎𝑟
𝑘 [ 𝑇 𝑇
] (20)
= 1
𝑥̃ 2𝑘 + 𝒙̃ 𝑇 𝜷̂ 𝑘 𝜷̂ 𝑘 𝒙̃ 𝑘 − 2𝑥̃ 𝑘 𝜷̂ 𝑘 𝒙̃ 𝑘 The control limit of the ET statistic is approximately determined from
𝜎𝑟
𝑘 [ ∑
𝑘
∑ ] the distribution 𝜒 2 (𝑙𝑥 ), where 𝑙𝑥 (𝑙𝑥 ≤ 𝑙 + 1) is the number of elements
1
= 𝜎𝑟
𝑥̃ 2𝑘 + 𝑚−𝑘 𝑗,ℎ=1 𝑥 ̃ 𝑗,𝑘 𝛾𝑗ℎ,𝑘 𝑥̃ ℎ,𝑘 − 2𝑥̃ 𝑘 𝑚−𝑘 ̂ ̃
𝑗=1 𝛽𝑗,𝑘 𝑥 𝑗,𝑘 in the set 𝛺𝑥 . Note that both the 𝐸𝑇𝑟,𝑥 2 value and its control limit vary
𝑘

4
Z. Li, S. Bao, X. Peng et al. Control Engineering Practice 116 (2021) 104916

Fig. 1. Procedure of fault detection and diagnosis.

with samples. The ET statistic in Eq. (26) is defined using only the Stage I: Offline data analysis
first key component of each sample. In fact, more key components Step I-1: Normalize the training data of variables to have mean zero
of a sample can be used in the ET statistic. The ET statistic is not and variance one.
recommended for monitoring normal samples. This is because the ET Step I-2: Determine the modeling sequence of variables according
statistic overly magnifies the role of key components while totally to the maximal multiple correlation coefficient rule.
ignores other components, which is easier to cause false alarms in nor- Step I-3: Build MCR models for all variables in sequence.
mal operating conditions. Thus, a recommended online fault detection Step I-4: Compute MCR residuals for all of training samples.
scheme is firstly using the 𝑊 𝑇𝑟2 statistic to detect the occurrence of Step I-5: Compute variances of residuals.
faults, and then switching to the ET statistic to improve the capability Stage II: Online fault detection and diagnosis
of detecting faulty samples.
Step II-1: Normalize a new sample 𝒙𝑛𝑒𝑤 using the mean and variance
After faults were detected, another task is to perform online fault di-
of training data.
agnosis using the contribution-based method. For every faulty sample,
Case a: If 𝒙𝑛𝑒𝑤 is collected before faults are detected.
the contribution of each component to the 𝑇𝑟2 value is computed via 2 2
Step II-a2: Compute 𝑇𝑟,𝑝 and 𝑇𝑟,𝑝 via Eq. (13) and coefficients
Eq. (19). The component with the largest contribution is regarded as 1 ,𝑥 2 ,𝑥
𝜃1,𝑥 and 𝜃2,𝑥 via Eq. (24).
the key faulty component. In the key faulty component, the contribu- 2 value via Eq. (23), and determine the
Step II-a3: Compute the 𝑊 𝑇𝑟,𝑥
tion of each variable is computed via Eq. (21) or Eq. (22). The variables
control limit via Eq. (16).
with much larger contributions are identified as faulty variables. Faulty
Step II-a4: Flag 𝒙𝑛𝑒𝑤 as faulty if its 𝑊 𝑇𝑟,𝑥 2 value exceeds the control
variables may vary with samples due to dynamic changes of the system
under fault effects. Therefore, the faulty variables obtained at the time limit. Faults are detected if l successive samples are all faulty.
when faults were just detected are more important for inferring the root Case b: If 𝒙𝑛𝑒𝑤 is collected after faults were detected.
cause of faults. Step II-b2: Determine key components of 𝒙𝑛𝑒𝑤 and the l samples
before 𝒙𝑛𝑒𝑤 , and the indices of l+1 key components constitute a set
3.4. Fault detection and diagnosis procedure 𝛺𝑥 of 𝑙𝑥 elements.
Step II-b3: Compute the 𝐸𝑇𝑟,𝑥 2 value via Eq. (26), and determine the
2
control limit from the 𝜒 (𝑙𝑥 ) distribution.
The complete procedure of fault detection and diagnosis is illus-
Step II-b4: Flag 𝒙𝑛𝑒𝑤 as faulty if its 𝐸𝑇𝑟,𝑥 2 value exceeds the control
trated in Fig. 1. Main steps of this procedure are summarized as
follows. limit.

5
Z. Li, S. Bao, X. Peng et al. Control Engineering Practice 116 (2021) 104916

Table 1
Information of fault data sets.
Data set Mean shifts
𝑿𝑓 ,1 𝛥𝝁𝑖 = [𝛥𝜇1,𝑖 , 𝛥𝜇2,𝑖 , 𝟎6 ]
𝑿𝑓 ,2 𝛥𝝁𝑖 = [𝟎𝟐 , 𝛥𝜇3,𝑖 , 𝛥𝜇6,𝑖 , 𝟎4 ]
𝑿𝑓 ,3 𝛥𝝁𝑖 = [𝟎3 , 𝛥𝜇4,𝑖 , 𝛥𝜇5,𝑖 , 𝟎𝟑 ]
𝑿𝑓 ,4 𝛥𝝁𝑖 = [𝟎𝟔 , 𝛥𝜇7,𝑖 , 𝛥𝜇8,𝑖 ]
𝑿𝑓 ,5 𝛥𝝁𝑖 = [𝟎2 , 𝛥𝜇3,𝑖 , 0, 𝛥𝜇5,𝑖 , 𝟎𝟑 ]
𝑿𝑓 ,6 𝛥𝝁𝑖 = [𝛥𝜇1,𝑖 , 𝟎𝟓 , 𝛥𝜇7,𝑖 , 0]
𝑿𝑓 ,7 𝛥𝝁𝑖 = [𝟎5 , 𝛥𝜇6,𝑗 , 0, 𝛥𝜇8,𝑗 ]

Step II-5: If 𝒙𝑛𝑒𝑤 is faulty, compute component contributions to the

𝑇𝑟2 value of 𝒙𝑛𝑒𝑤 via Eq. (19), flag the component with the largest
contribution as the key faulty component, and then compute variable
contributions to the key faulty component via Eq. (21) or Eq. (22), and
flag variables with much larger contributions as faulty variables.
Step II-6: Return to Step II-1 and monitor the next sample.

4. Case studies

4.1. Simulation examples Fig. 2. Multiple correlation coefficients of variables.

Consider a multivariate system of 8 variables, which follows a

Table 2
normal distribution with the zero mean and the covariance matrix of FSDRs (%) for test data sets.
⎡ 1.0 0.5 0.4 0.2 0.1 0.0 −0.2 0.0 ⎤ Data set MCR PCA CT
⎢ 0.5 1.0 0.6 0.1 0.1 0.1 0.0 0.1 ⎥ 2
𝑇𝑟,𝑝 2
𝑇𝑟,𝑝 IT 𝑊 𝑇𝑟2 𝑇2 SPE 𝑇 2 +SPE
⎢ ⎥ 1 2

⎢ 0.4 0.6 1.0 0.2 −0.1 0.0 0.2 0.0 ⎥ 𝑿𝑛,0 1.1 0.9 2.0 2.1 1.0 1.1 2.1 1.1
⎢ 0.2 0.1 0.2 1.0 0.7 0.4 0.2 0.1 ⎥ 𝑿𝑓 ,1 27.6 1.0 28.3 29.7 8.2 18.2 24.8 20.8
𝜮=⎢ ⎥ (27)
⎢ 0.1 0.1 −0.1 0.7 1.0 0.6 0.2 0.3 ⎥ 𝑿𝑓 ,2 9.9 19.9 27.6 31.5 11.4 18.4 27.7 25.7
⎢ 0.0 0.1 0.0 0.4 0.6 1.0 −0.1 0.2 ⎥ 𝑿𝑓 ,3 20.2 0.9 20.9 22.3 8.2 11.4 18.6 16.2
⎢ −0.2 0.0 0.2 0.2 0.2 −0.1 1.0 0.6 ⎥ 𝑿𝑓 ,4 4.6 9.0 13.3 14.9 8.9 3.9 12.5 9.5
⎢ ⎥ 𝑿𝑓 ,5 9.9 20.3 28.0 32.1 11.3 19.7 28.5 26.5
⎣ 0.0 0.1 0.0 0.1 0.3 0.2 0.6 1.0 ⎦ 𝑿𝑓 ,6 66.8 1.0 67.1 68.2 6.6 66.3 68.3 59.4
𝑿𝑓 ,7 8.1 16.6 23.3 26.9 11.0 13.4 22.8 20.8
A total of 2 × 104 samples are randomly generated from the multivari-
ate normal distribution 𝑁 (𝟎, 𝜮). These samples constitute a training
data set 𝑿0 and a normal test data set 𝑿𝑛,0 , with 1 × 104 samples
in each data set. Fault test data sets are also generated by adding The detection performance is evaluated by the faulty sample detection
random mean shifts to some variables, that is, faulty samples follow the rate (FSDR), i.e., the percentage of samples that are detected as faulty.
distribution 𝑁 (𝛥𝝁, 𝜮) with 𝛥𝝁 ≠ 𝟎. Table 1 shows the information of The FSDRs for test data sets are shown in Table 2. The column
fault data sets. In the data set 𝑿𝑓 ,1 , samples are separately generated 𝑇 2 +SPE shows joint detection results of the PCA-based 𝑇 2 and SPE
( )
from the normal distributions 𝑁 𝛥𝝁𝑖 , 𝜮 (𝑖 = 1, . . . , 1 × 104 ) with statistics, similar to the IT statistic. For fault data sets 𝑿𝑓 ,1 , 𝑿𝑓 ,3 and
𝛥𝝁𝑖 = [𝛥𝜇1,𝑖 , 𝛥𝜇2,𝑖 , 𝟎6 ]. Values of 𝛥𝜇1,𝑖 and 𝛥𝜇2,𝑖 are randomly drawn 𝑿𝑓 ,6 , with faults caused by mean shifts of variables corresponding
from the uniform distribution 𝑈 (1, 2). Similar to 𝑿𝑓 ,1 , the other six fault 2 , only the 𝑇 2
to 𝑇𝑟,𝑝 2
( ) 1 𝑟,𝑝1 detects faulty samples, while the 𝑇𝑟,𝑝2 almost
test data sets are generated from 𝑁 𝛥𝝁𝑖 , 𝜮 with mean shifts added to cannot detect faulty samples because variables used in 𝑇𝑟,𝑝 contain 2
2
different variables. no fault information. Moreover, for 𝑿𝑓 ,1 , 𝑿𝑓 ,3 and 𝑿𝑓 ,6 , the 𝑇𝑟,𝑝 2 has
1
The MCR models are built for 8 variables in sequence. The modeling
{ } larger FSDRs that the CT statistic. This is because the 𝑇𝑟,𝑝 2 reduces
sequence of variables is 𝑣5 , 𝑣7 , 𝑣2 , 𝑣4 , 𝑣1 , 𝑣6 , 𝑣3 , 𝑣8 . The MCR model of 1
the amplification and masking effects 20 by excluding the redundant
each variable is built using all the subsequent variables. Fig. 2 shows 2
information contained in 𝑇𝑟,𝑝 . Faults in 𝑿𝑓 ,2 , 𝑿𝑓 ,4 , 𝑿𝑓 ,5 , and 𝑿𝑓 ,7
the multiple correlation coefficient of each variable with subsequent 2
2
are caused by mean shifts of variables corresponding to only 𝑇𝑟,𝑝 or
variables. Clearly, the multiple correlation coefficient sequence is in 2 2 2
2
2
the descending order. The first MCR model is built for the variable both 𝑇𝑟,𝑝 and 𝑇𝑟,𝑝 . For these fault data sets, both the 𝑇𝑟,𝑝 and 𝑇𝑟,𝑝
1 2 1 2
detect faulty samples, while the 𝑇𝑟,𝑝 2 has higher FSDRs. This is because
5, because it has the largest multiple correlation coefficient. Most of 2
2
the 𝑇𝑟,𝑝 contains only 3 variables, and hence it suffers less from the
variable correlations are captured by the first 5 MCR models. Multiple 2
amplification and masking effects than the 𝑇𝑟,𝑝 2 . However, a part of
correlation coefficients among the last three variables are very low. 1
2
variable information is lost in either one of 𝑇𝑟,𝑝 and 𝑇𝑟,𝑝 2 . This is the
Especially, the correlation between variables 3 and 8 is zero. Therefore, 1 2
2
the 𝑇𝑟,𝑝 in Eq. (23) is defined using the residuals of first five variables 2
reason why the IT and 𝑊 𝑇𝑟 statistics always perform better than 𝑇𝑟,𝑝 2
1 ,𝑥 1
2
𝑣5 , 𝑣7 , 𝑣2 , 𝑣4 and 𝑣1 , and the 𝑇𝑟,𝑝 is defined using the residuals of the and 𝑇𝑟,𝑝2 . For all the seven fault data sets, both the IT and 𝑊 𝑇 2 statistics
2 ,𝑥 2 𝑟
last three variables 𝑣6 , 𝑣3 and 𝑣8 . Based on 𝑇𝑟,𝑝 2 2
and 𝑇𝑟,𝑝 , both the IT 2
have larger FSDRs than the CT statistic, especially the 𝑊 𝑇𝑟 statistic.
1 ,𝑥 2 ,𝑥
2
and 𝑊 𝑇𝑟 statistics are defined to detect the number of faulty samples In general, the FSDR of the 𝑊 𝑇𝑟2 statistic increases by more than 5%
in every test data set. For each sample, two weight coefficients of the as compared to the CT statistic. However, this improvement in fault
𝑊 𝑇𝑟2 statistic are computed via Eq. (24). Performance of the IT and detection performance is at the price of a slight increase (about 1%)
𝑊 𝑇𝑟2 statistics are compared with the classical 𝑇 2 (CT) statistic (see in the FSDR for the normal test data 𝑿𝑛,0 , which is also known as the
Eq. (8)) and the PCA-based method (Nomikos & MacGregor, 1995). In false alarm rate. The performance of the PCA-based 𝑇 2 +SPE statistics
the PCA-based method, a 𝑇 2 statistic is defined using the first 5 PCs is comparable to the IT statistic but is not as good as the 𝑊 𝑇𝑟2 statistic.
corresponding to the largest eigenvalues, and a SPE statistic is defined These results indicate that the 𝑊 𝑇𝑟2 statistic has the best fault detection
using the last 3 PCs. The 99% control limit is used for all of statistics. capability.

6
Z. Li, S. Bao, X. Peng et al. Control Engineering Practice 116 (2021) 104916

Table 3
Monitored variables in the TE process.
Label Variable Label Variable
𝑣1 A feed rate (stream 1) 𝑣18 Stripper temperature
𝑣2 D feed rate (stream 2) 𝑣19 Stripper steam flow
𝑣3 E feed rate (stream 3) 𝑣20 Compress work
𝑣4 A and C feed rate (stream 4) 𝑣21 Reactor cooling water outlet temperature
𝑣5 Recycle flow (stream 8) 𝑣22 Condenser cooling water outlet temperature
𝑣6 Reactor feed rate (stream 6) 𝑣23 D feed flow valve (stream 2)
𝑣7 Reactor pressure 𝑣24 E feed flow valve (stream 3)
𝑣8 Reactor level 𝑣25 A feed flow valve (stream 1)
𝑣9 Reactor temperature 𝑣26 A and C feed flow valve (stream 4)
𝑣10 Purge rate (stream 9) 𝑣27 Compressor recycle value
𝑣11 Product separator temperature 𝑣28 Purge valve (stream 9)
𝑣12 Product separator level 𝑣29 Separator pot liquid flow valve (stream 10)
𝑣13 Product separator pressure 𝑣30 Stripper liquid product flow valve (stream 11)
𝑣14 Product separator underflow 𝑣31 Stripper steam valve
𝑣15 Stripper level 𝑣32 Reactor cooling water flow valve
𝑣16 Stripper pressure 𝑣33 Condenser cooling water flow valve
𝑣17 Stripper underflow (stream 11)

Table 4 correlations are captured by the first 20 MCR models, especially the
Faults in the TE process. first 10 models. For the remaining 13 variables, the largest multiple
No. Faulty variable Fault type correlation coefficient is only about 0.22 (the variable 𝑣3 ). Therefore,
1 A/C feed ratio, B composition constant (stream 4) Step 2
the 𝑇𝑟,𝑝 in Eq. (23) is defined using residuals of the first 20 MCR
1 ,𝑥
2 B composition, A/C feed ratio constant (stream 4) Step 2
models, and the 𝑇𝑟,𝑝 is defined using residuals of the last 13 MCR
3 D feed temperature (stream 2) Step 2 ,𝑥
4 Reactor cooling water inlet temperature Step models. The 𝑊 𝑇𝑟2 statistic (see Eq. (23)) is used to detect faults. Faults
5 Condenser cooling water inlet temperature Step are detected if the 𝑊 𝑇𝑟2 values of 3 consecutive samples all exceed their
6 A feed loss (stream 1) Step control limits. After faults were detected, the 𝑊 𝑇𝑟2 statistic is replaced
7 C header pressure loss-reduced availability (stream 4) Step
by the ET statistic (see Eq. (26)) to monitor subsequent samples. The
8 A, B, C feed composition (stream 4) Random
9 D feed temperature (stream 2) Random
index set 𝛺𝑥 in the ET statistic consists of the indices of key components
10 C feed temperature (stream 4) Random of 4 consecutive samples. This online fault detection scheme is denoted
11 Reactor cooling water inlet temperature Random as 𝑊 𝑇𝑟2 +ET. This scheme is compared with the case of using only
12 Condenser cooling water inlet temperature Random the 𝑊 𝑇𝑟2 statistic and two other methods using the classical 𝑇 2 (CT)
13 Reaction kinetics Slow drift
statistic (see Eq. (8)) and the PCA-based 𝑇 2 and SPE statistics. In the
14 Reactor cooling water valve Sticking
15 Condenser cooling water valve Sticking PCA-based method, the 𝑇 2 statistic is defined using the first 16 PCs
16 Unknown Unknown that correspond to the largest eigenvalues, and the remaining 17 PCs
17 Unknown Unknown are used to define the SPE statistic. The 99% control limit is used
18 Unknown Unknown for all of statistics. Fault detection performance is evaluated by the
19 Unknown Unknown
false alarm rate (FAR), faulty sample detection rate (FSDR), and fault
20 Unknown Unknown
21 Valve for stream 4 fixed at steady state position Constant detection delay (FDD). The FAR is the percentage of samples that are
falsely detected as faulty in the normal operating period. The FSDR
is the percentage of samples that are detected as faulty in the faulty
operating period. The FDD is the number of samples collected in the
4.2. An industrial case study time period after a fault occurred (at the 161st sample) but before the
fault is detected.
4.2.1. Process description Table 5 shows detection results of different methods for 21 fault
The proposed FDD method is tested by the Tennessee Eastman (TE) data sets. The 𝑊 𝑇𝑟2 , 𝑊 𝑇𝑟2 +ET and PCA-based 𝑇 2 +SPE statistics have
chemical process simulation (Down & Vogel, 1993). A flowchart of slightly higher FARs than the CT statistic. For most of faults, the 𝑊 𝑇𝑟2 ,
the TE process is shown in Fig. 3. The TE process consists of five 𝑊 𝑇𝑟2 +ET, CT and PCA-based 𝑇 2 +SPE statistics have comparable FDDs.
major unit operations: a reactor, a product condenser, a vapor–liquid For fault data sets 3, 9 and 15, the FDSRs of 𝑊 𝑇𝑟2 +ET exceed 18%,
separator, a recycle compressor and a product stripper. In this process, but the CT and PCA-based 𝑇 2 +SPE statistics have very low FSDRs. It is
four reactants (A, C, D and E) are converted into three products (F, noteworthy that faulty samples in data set 3, 9 and 15 are difficult to be
G and H). Down and Vogel (1993) have built a model for the TE detected, because these three faults have little effects on the TE process.
process and developed the simulation code. The benchmark simulation For fault data sets 10, 11, 16, 19 and 21, the 𝑊 𝑇𝑟2 +ET statistics also
data of the TE process are available at the website http://web.mit. have larger FSDRs than other statistics. Especially for fault data sets 11
edu/braatzgroup/links.html. These data contain a normal data set and and 21, the FSDRs of 𝑊 𝑇𝑟2 +ET increase by about 10% as compared
21 fault data sets. Each data set consists of 960 measurements of to the CT statistic and the PCA-based 𝑇 2 +SPE statistics. For these five
52 process variables. The data of 33 variables are used in this case fault data sets, the 𝑊 𝑇𝑟2 statistic also has slightly higher FSDRs than the
study. The selected 33 variables are shown in Table 3. There are 21 CT statistic and the PCA-based 𝑇 2 +SPE statistics. This indicates that the
disturbances that cause faults, as shown in Table 4. In each fault data 𝑊 𝑇𝑟2 statistic is better than the CT statistic and the PCA-based method.
set, a disturbance was introduced into the TE process after the 160th In addition, after faults were detected, the use of the ET statistic can
measurement. further improve the capability of detecting faulty samples. For fault
data sets 1, 2, 4–8, 12–14, 17, 18 and 20, the 𝑊 𝑇𝑟2 , 𝑊 𝑇𝑟2 +ET, CT and
4.2.2. Fault detection results PCA-based 𝑇 2 +SPE statistics have comparable performance and their
The MCR models are built for 33 variables in sequence. Fig. 4 shows FSDRs are all very high. The reason is that fault magnitudes in these
the modeling sequence of variables and the multiple correlation coeffi- data sets are very large such that faulty samples are easy to be detected.
cient of each variable with the subsequent variables. Clearly, multiple For comparison purposes, Table 5 also shows the results of three
correlation coefficients are in the descending order. Most of variable recently proposed fault detection methods, including the sparsity and

7
Z. Li, S. Bao, X. Peng et al. Control Engineering Practice 116 (2021) 104916

Fig. 3. Flowchart of the Tennessee Eastman process.

reasons. First, the MCR is better than other three methods in modeling
the multiple correlation among variables. Second, unlike the traditional
𝑇 2 and SPE statistics used in other three methods, the 𝑊 𝑇𝑟2 and ET
statistics can effectively reduce the amplification and masking effects
caused by fault-free variables.
Fig. 5 and Fig. 6 show monitoring charts for fault data sets 11
and 21 respectively. The below plots of Fig. 5a and Fig. 6a show the
key component of every sample. Fig. 5 shows that all of statistics
detect the fault occurrence in the data set 11 at the 166th sample.
During the faulty operating period from the 166th sample to the 960th
sample, about 20% of faulty samples are not detected by the WT, CT
or PCA-based 𝑇 2 and SPE statistics, as shown in Fig. 5b–d and Table 5.
However, the ET statistic only misses 10% faulty samples (see Fig. 5a).
The key components of faulty samples mainly alternate between two
components 14 and 17, corresponding to variables 𝑣9 and 𝑣32 . For
the fault data set 21, Fig. 6 b–d show that a large part of samples in
the period between the 420th sample and the 560th sample are not
detected by the WT, CT or PCA-based 𝑇 2 and SPE statistics, while only
12 faulty samples are missed by the ET statistic. As shown in Fig. 6a, the
component 12 is the key component for most of faulty samples. These
Fig. 4. Multiple correlation coefficients of variables in the TE process.
two examples illustrate that, during the faulty operating period, the ET
statistic is much better than other statistics in detecting faulty samples.
The reason is that the ET statistic focuses only on key components that
manifold regularized convolutional auto-encoders (SMRCAE) (Zhang are most important to detect a faulty sample, unlike other statistics that
et al., 2021), just-in-time feature analysis (JITFA) (Qin et al., 2021) consider all of components simultaneously. This significantly reduces
and hierarchical hybrid distributed PCA (HDPCA) (Cao et al., 2021). the interference from fault-free components, and hence enables the ET
As shown in Table 5, the MCR-based 𝑊 𝑇𝑟2 +ET statistics have better statistic to have a better capability of detecting faulty samples.
fault detection performance than other three methods, by virtue of its
higher FSDR, smaller FDD and lower FAR. For fault data sets 10, 11, 4.2.3. Fault diagnosis results
16, 19 and 21, the 𝑊 𝑇𝑟2 +ET statistics have much higher FSDRs than The contribution-based fault diagnosis method is used to identify
other three methods. The FDDs of 𝑊 𝑇𝑟2 +ET are generally smaller than faulty variables in each fault data set. The performance of this method
those of JITFA. For all of 21 fault data sets, the 𝑊 𝑇𝑟2 +ET statistics is illustrated by fault data sets 6 and 11. The occurrence of faults in the
have much smaller FARs than the HDPCA method. It is evident by data set 6 was detected by the WT statistic at the 161st sample. Fig. 7a
Table 5 that the MCR-based 𝑊 𝑇𝑟2 +ET statistics perform better than the shows contributions of 33 components to the 𝑇𝑟2 value of the 161st
SMRCAE, JITFA and HDPCA methods. This may be attributed to two sample. Clearly, the 5th component has a much larger contribution

8
Z. Li, S. Bao, X. Peng et al. Control Engineering Practice 116 (2021) 104916

Table 5
Detection results of different methods for 21 fault data sets.
Fault MCR PCA CT SMRCAE JITFA HDPCA
(Zhang et al., 2021) (Qin et al., 2021) (Cao et al., 2021)
𝑊 𝑇𝑟2 𝑊 𝑇𝑟2 +ET 𝑇 2 +SPE 𝑇2 SPE
FSDR FAR FDD FSDR FAR FDD FSDR FAR FDD FSDR FAR FDD FSDR FAR FSDR FAR FSDR FDD FSDR FAR
1 100.0 1.3 0 100.0 1.3 0 99.9 2.5 0 100.0 0.6 0 99.8 1.3 99.8 1.9 99.9 2 99.9 8.1
2 98.5 0.6 12 98.5 0.6 12 98.4 0.6 2 98.4 0.0 0 98.6 1.9 98.3 1.9 98.4 13 99.5 13.1
3 6.0 8.8 26 24.6 8.8 26 5.5 6.3 14 4.4 6.3 13 17.4 3.1 16.4 3.8 / / 27.3 21.3
4 100.0 2.5 0 100.0 2.5 0 100.0 0.6 55 100.0 0.6 57 94.3 0.6 100.0 0.0 100.0 0 100.0 12.5
5 100.0 2.5 0 100.0 2.5 0 100.0 0.6 0 100.0 0.6 0 96.4 0.6 100.0 0.0 100.0 0 44.3 15.6
6 100.0 0.0 0 100.0 0.0 0 100.0 0.0 0 100.0 0.0 0 100.0 0.6 100.0 0.6 100.0 0 100.0 5.0
7 100.0 0.0 0 100.0 0.0 0 100.0 0.0 0 100.0 0.0 0 100.0 2.5 100.0 2.5 100.0 0 100.0 5.6
8 98.3 1.3 15 98.3 1.3 15 98.3 0.6 0 98.0 0.0 0 99.1 24.4 99.1 16.9 97.9 19 98.6 13.8
9 4.8 7.5 0 22.8 7.5 0 4.4 5.0 17 3.6 5.0 19 12.4 18.1 10.9 21.3 / / 20.9 35.0
10 91.4 0.0 21 94.5 0.0 21 90.3 0.6 6 89.6 0.0 4 81.4 3.8 90.5 3.1 88.4 21 78.4 11.3
11 84.0 1.3 5 91.9 1.3 5 81.4 1.9 27 81.1 0.6 27 77.3 2.5 80.6 2.5 78.7 5 85.5 12.5
12 99.9 0.6 1 99.9 0.6 1 99.9 0.6 5 99.9 0.6 5 99.6 2.5 99.5 0.6 99.9 1 99.1 12.5
13 95.4 0.0 37 95.4 0.0 37 95.4 0.6 1 95.3 0.0 1 94.8 1.3 94.8 0.6 95.3 40 95.9 5.6
14 100.0 0.6 0 100.0 0.6 0 100.0 0.0 37 100.0 0.6 37 100.0 1.9 100.0 0.6 100.0 0 100.0 15.0
15 10.6 5.0 561 18.8 5.0 561 7.8 5.0 0 7.3 3.4 0 12.8 2.5 15.1 1.9 / / 32.9 8.8
16 93.5 5.6 6 96.0 5.6 6 92.9 5.6 564 92.9 3.8 563 84.1 23.8 93.3 27.5 91.8 6 70.6 36.3
17 97.6 5.6 19 98.0 5.6 19 97.0 5.6 6 97.3 1.9 6 95.5 3.1 97.0 1.9 97.1 19 96.5 21.3
18 90.8 1.3 77 90.8 1.3 77 90.5 0.6 19 90.4 1.3 19 90.9 3.8 90.1 0.6 90.4 78 93.5 10.6
19 94.6 1.3 1 97.1 1.3 1 93.1 1.3 77 94.4 0.6 78 44.4 1.3 83.0 0.6 92.5 9 82.6 9.4
20 91.4 1.3 62 92.3 1.3 62 91.4 1.9 1 91.0 0.6 1 66.3 0.6 81.9 0.6 90.9 66 78.0 11.3
21 59.0 5.6 256 66.9 5.6 256 56.8 4.4 62 56.8 3.1 62 52.8 6.3 48.9 6.9 55.1 261 55.6 16.3

Fig. 5. Monitoring charts for the fault data set 11. (a) WT+ET, (b) WT, (c) CT, and (d) PCA-based 𝑇 2 and SPE.

than other components, and hence it is regarded as the key faulty The 5th component is the key faulty component for samples in the
component. Contributions of variables to this component are shown period from 161 to 340 (see Fig. 8a), and the faulty variable is either
in Fig. 7b. Variables 1 and 25 are identified as faulty, because their 𝑣1 or 𝑣25 (see Fig. 8b). Because variables 𝑣1 (A feed rate) and 𝑣25 (A
contributions are much larger than other variables. Fig. 8 shows the feed flow valve) correspond to the control loop of the A feed flow (see
identified faulty components and faulty variables for all faulty samples. Fig. 3), it is a natural inference that the fault in the data set 6 was
In Fig. 8, contributions are normalized by dividing the largest value. caused by the disturbance in the A feed flow. This diagnosis result

9
Z. Li, S. Bao, X. Peng et al. Control Engineering Practice 116 (2021) 104916

Fig. 6. Monitoring charts for the fault data set 21. (a) WT+ET, (b) WT, (c) CT, and (d) PCA-based 𝑇 2 and SPE.

Fig. 7. Contribution plots at the 161st sample in the fault data set 6. (a) component contribution, (b) variable contribution.

agrees well with the real cause of the fault data set 6 (see Table 4). is 𝑣32 (see Fig. 9b). Fig. 10 shows the identified faulty components
Fig. 8a also shows that the 3rd component becomes the key faulty and faulty variables for all faulty samples in the data set 11. In the
component for samples in the period from 340 to 960, and the variable faulty operating period from the 166th sample to the 960th sample,
𝑣33 is the faulty variable. This is because of the fault propagation caused the key faulty component mainly alternates between the 14th and
by the mass, energy and information flow in the TE process (Luo et al., 17th components, and the faulty variable is either 𝑣9 or 𝑣32 . Note that
2017). variables 𝑣9 (Reactor temperature) and 𝑣32 (Reactor cooling water flow
The fault in the data set 11 was caused by random changes of the valve) correspond to the control loop of the reactor temperature (see
reactor cooling water inlet temperature. Because this faulty variable Fig. 3). It is reasonable to infer that this fault was most likely caused by
was not monitored, it cannot be identified directly. This increases the disturbance in the reactor cooling system. This diagnosis result is
the difficulty of fault diagnosis. The fault was detected by the WT very close to the real cause of the fault. These two examples illustrate
statistic at the 166th sample. At this sample, the 14th component that the contribution-based method can produce reliable fault diagnosis
is the key faulty component (see Fig. 9a), and the faulty variable results.

10
Z. Li, S. Bao, X. Peng et al. Control Engineering Practice 116 (2021) 104916

Fig. 8. Contribution plots for all faulty samples in the fault data set 6. (a) component contribution, (b) variable contribution.

Fig. 9. Contribution plots at the 166th sample in the fault data set 11. (a) Component contribution, (b) variable contribution.

Fig. 10. Contribution plots for all faulty samples in the fault data set 11. (a) Component contribution, (b) variable contribution.

5. Conclusion minimal variance. Based on MCR residuals of variables, a 𝑇 2 statistic

is defined for fault detection. To minimize the control region of this 𝑇 2
Variable correlations are inherent features of a multivariate system statistic, MCR models are built for variables in sequence by following
and hence are very useful to fault detection and diagnosis (FDD). In this a maximal multiple correlation coefficient rule. Three improvements
paper, multiple correlation regression (MCR) is applied for modeling to the residual-based 𝑇 2 statistic are also developed, called the inter-
correlations among variables. The MCR is the best linear regression in secting 𝑇 2 (IT), weighted 𝑇 2 (WT) and enhanced 𝑇 2 (ET) statistics
the least square sense, which produces the regression residuals with the respectively. These statistics have better fault detection performance,

11
Z. Li, S. Bao, X. Peng et al. Control Engineering Practice 116 (2021) 104916

by virtue of the ability to reduce the amplification and masking effects Luo, L., Bao, S., & Tong, C. (2019). Sparse robust principal component analysis with
caused by fault-free components. In particular, the WT and ET statistics applications to fault detection and diagnosis. Industrial and Engineering Chemistry
Research, 58(3), 1300–1309.
and their control limits vary with samples adaptively. This enables the
Luo, L., Lovelett, R. J., & Ogunnaike, B. A. (2017). Hierarchical monitoring of industrial
WT and ET statistics to adapt to dynamic changes of the multivariate processes for fault detection, fault grade evaluation, and fault diagnosis. AIChE
system. An online fault detection strategy is proposed by combining Journal, 63(7), 2781–2795.
the WT and ET statistics. The WT statistic is firstly used to detect Luo, L., Peng, X., & Tong, C. (2021). A multigroup framework for fault detection and
the fault occurrence, and then the WT statistic is replaced by the ET diagnosis in large-scale multivariate systems. Journal of Process Control, 100, 65–79.
Luo, L., Wang, J., Tong, C., & Zhu, J. (2020). Multivariate fault detection and diagnosis
statistic to further enhance the capability of detecting subsequent faulty based on variable grouping. Industrial and Engineering Chemistry Research, 59(16),
samples. A contribution-based method is proposed for fault diagnosis. 7693–7705.
Faulty variables are identified as the variables with larger contributions Luo, L., Xu, M., Bao, S., Mao, J., & Tong, C. (2019). Improvements to the 𝑇 2 statistic for
to the key faulty component of the residual-based 𝑇 2 statistic. It is multivariate fault detection. Industrial and Engineering Chemistry Research, 58(45),
20692–20709.
illustrated by two case studies that the WT statistic has better fault
MacGregor, J. F., & Cinar, A. (2012). Monitoring, fault diagnosis, fault-tolerant control
detection performance than the classical 𝑇 2 statistic and the PCA-based and optimization: Data driven methods. Computers & Chemical Engineering, 47,
method. The online fault detection strategy, which combines the WT 111–120.
and ET statistics together, improves the detection rate of faulty samples MacGregor, J. F., Jaeckle, C., Kiparissides, C., & Koutoudi, M. (1994). Process
significantly. The contribution-based method produces reliable fault monitoring and diagnosis by multiblock PLS methods. AIChE Journal, 40(5),
826–838.
diagnosis results. Mason, R. L., & Young, J. C. (2002). Multivariate statistical process control with industrial
applications. Philadelphia: ASA/SIAM.
Declaration of competing interest Nomikos, P., & MacGregor, J. F. (1995). Multivariate SPC charts for monitoring batch
processes. Technometrics, 37(1), 41–59.
Park, Y. J., Fan, S.-K. S., & Hsu, C.-Y. (2020). A review on fault detection and process
The authors declare that they have no known competing finan-
diagnostics in industrial processes. Processes, 8(9), 1123.
cial interests or personal relationships that could have appeared to Peng, X., Ding, S. X., Du, W. L., Zhong, W. M., & Qian, F. (2020). Distributed process
influence the work reported in this paper. monitoring based on canonical correlation analysis with partly-connected topology.
Control Engineering Practice, 101, Article 104500.
Acknowledgments Qin, S. J. (2012). Survey on data-driven industrial process monitoring and diagnosis.
Annual Reviews in Control, 36, 220–234.
Qin, L., Tong, C., Lan, T., & Chen, Y. (2021). Statistical process monitoring based on
This research was supported by the Zhejiang Provincial Basic Public just-in-time feature analysis. Control Engineering Practice, 115, Article 104889.
Welfare Research Project of China (No. LGG21F030010), the National Shang, C., Yang, F., Huang, B., & Huang, D. (2018). Recursive slow feature analysis
Natural Science Foundation of China (No. 52075489), the Fundamental for adaptive monitoring of industrial processes. IEEE Transactions on Industrial
Electronics, 65(11), 8895–8905.
Research Funds for the Central Universities, China (2021ACOCP04).
Song, B., Shi, H. B., Tan, S., & Tao, Y. (2020). Multi-subspace orthogonal canonical cor-
relation analysis for quality related plant wide process monitoring. IEEE Transactions
References on Industrial Informatics, 17(9), 6368–6378.
Valle, S., Li, W., & Qin, S. J. (1999). Selection of the number of principal components:
Anderson, T. W. (2003). An introduction to multivariate statistical analysis. Huboken, New The variance of the reconstruction error criterion with a comparison to other
Jersey: John Wiley & Suns, Inc. methods. Industrial and Engineering Chemistry Research, 38(11), 4389–4401.
Aparisi, F., Jabaloyes, J., & Carrion, A. (1999). Statistical properties of the |S| Wan, X., Tong, C., & Luo, L. (2020). Distributed statistical process monitoring based
multivariate control chart. Communications in Statistics. Theory and Methods, 28(11), on multiblock canonical correlation analysis. Industrial and Engineering Chemistry
2671–2686. Research, 59(3), 1193–1201.
Box, G. E. P. (1954). Some theorems on quadratic forms applied in the study of Xie, L., Zeng, J., Kruger, U., Wang, X., & Geluk, J. (2015). Fault detection in dynamic
analysis of variance problems. I. Effect of inequality of variance in the one-way systems using the Kullback–Leibler divergence. Control Engineering Practice, 43,
classification. The Annals of Mathematical Statistics, 25, 290–302. 39–48.
Cao, Y., Yuan, X., Wang, Y., & Gui, W. (2021). Hierarchical hybrid distributed PCA Yan, L., Peng, X., Tong, C., & Luo, L. (2020). A multigroup fault detection and diagnosis
for plant-wide monitoring of chemical processes. Control Engineering Practice, 111, scheme for multivariate systems. Industrial and Engineering Chemistry Research,
Article 104784. 59(47), 20767–20778.
Dong, Y., & Qin, S. J. (2018). A novel dynamic PCA algorithm for dynamic data Yin, S., Ding, S. X., Xie, X. C., & Luo, H. (2014). A review on basic data-
modeling and process monitoring. Journal of Process Control, 67, 1–11. driven approaches for industrial process monitoring. IEEE Transactions on Industrial
Down, J. J., & Vogel, E. F. (1993). A plant-wide industrial process control problem. Electronics, 61, 6148–6428.
Computers & Chemical Engineering, 17, 245–255. Yu, J., & Qin, S. J. (2009). Multiway Gaussian mixture model based multiphase
Fezai, R., Mansouri, M., Abodayeh, K., Nounou, H., & Nounou, M. (2020). Online batch process monitoring. Industrial and Engineering Chemistry Research, 48(18),
reduced gaussian process regression based generalized likelihood ratio test for fault 8585–8594.
detection. Journal of Process Control, 85, 30–40. Yu, J., Yoo, J., Jang, J., Park, J. H., & Kim, S. (2018). A novel hybrid of auto-associative
Ge, Z. (2017). Review on data-driven modeling and monitoring for plant-wide industrial kernel regression and dynamic independent component analysis for fault detection
processes. Chemometrics and Intelligent Laboratory Systems, 171, 16–25. in nonlinear multimode processes. Journal of Process Control, 68, 129–144.
He, F., & Zhang, Z. Y. (2020). Nonlinear fault detection of batch processes Zhang, Q., Li, P., Lang, X., & Miao, A. (2020). Improved dynamic kernel principal
using functional local kernel principal component analysis. IEEE Access, 8, component analysis for fault detection. Measurement, 158, Article 107738.
117513–117527. Zhang, C., Yu, J., & Ye, L. (2021). Sparsity and manifold regularized convolutional
Jiang, Q., Yan, X., & Huang, B. (2019). Review and perspectives of data-driven auto-encoders-based feature learning for fault detection of multivariate processes.
distributed monitoring for industrial plant-wide processes. Industrial and Engineering Control Engineering Practice, 111, Article 104811.
Chemistry Research, 58(29), 12899–12912.
Luo, L., Bao, S., & Mao, J. (2019). Adaptive selection of latent variables for process
monitoring. Industrial and Engineering Chemistry Research, 58(21), 9075–9086.

12