Wang 2016

Available online at www.sciencedirect.
com
Journal of the Franklin Institute ] (]]]]) ]]]–]]]

www.elsevier.com/locate/jfranklin
Fault isolation based on residual evaluation

and contribution analysis$
Jing Wangn, Wenshuang Ge, Jinglin Zhou, Haiyan Wu, Qibing Jin
College of Information Science and Technology, Beijing University of Chemical Technology, Beijing 100029, China
Received 9 November 2015; received in revised form 12 July 2016; accepted 2 September 2016
Abstract
A new fault isolation strategy for industrial processes is presented based on residual evaluation and contribution
plot analysis. Based on the space projection, the residual evaluation and contribution plot are unified into a
framework. First, parity space and subspace identification methods are used to generate residuals for fault
detection. Then, the optimal residuals are utilized to obtain a process fault isolation scheme. A new contribution
index is calculated according to the average value of the current and previous residuals. The smearing effect can
be eliminated, and the fault evolution can be acquired based on this index. This would be helpful for engineers to
find the fault roots and then eliminate them. A numerical model and the Tennessee Eastman process are
considered to assess the isolation performance of the proposed approach. The results demonstrate that the
smearing effect is improved and the primary faulty variable can be located accurately when several different faults
occur simultaneously. A superior isolation performance is obtained compared with the PCA-based isolation
method. The reasonability of the isolation results is analyzed using fault propagation.
& 2016 The Franklin Institute. Published by Elsevier Ltd. All rights reserved.
Keywords: Fault isolation; Residual evaluation; Contribution plots; Fault evolution; Smearing effect
☆
The work is supported by the National Natural Science Foundation of China, China (61473025, 61403017,
61573050), State Key Laboratory of Management and Control for Complex Systems, Institute of Automation, Chinese
Academy of Sciences, (20160107).
n
Corresponding author.
E-mail address: jwang@mail.buct.edu.cn (J. Wang).
http://dx.doi.org/10.1016/j.jfranklin.2016.09.002
0016-0032/& 2016 The Franklin Institute. Published by Elsevier Ltd. All rights reserved.
Please cite this article as: J. Wang, et al., Fault isolation based on residual evaluation and contribution analysis,
Journal of the Franklin Institute. (2016), http://dx.doi.org/10.1016/j.jfranklin.2016.09.002
2 J. Wang et al. / Journal of the Franklin Institute ] (]]]]) ]]]–]]]
1. Introduction
Fault detection and isolation (FDI) have attracted much attention for complex industrial
processes during the last decades. FDI monitoring systems are very useful and effective to detect
faults in a timely fashion and accurately isolate the faulty equipment. This can avoid shutting the
plant down or accidents. Some researchers have reviewed and surveyed the development of
different FDI methods and schemes [1–3]. These approaches can be categorized into two classes:
model-based and data-driven methods. The analytical redundancy approach [2] is a typical
model-based method. Statistical process monitoring (SPM) or latent space approaches, such as
principal component analysis (PCA), partial least squares (PLS) and independent component
analysis (ICA) [4,5], are classified as data-driven methods. A unified and composite scheme of
model-based and data-driven methods in FDI systems is given in [6].
Recently, [7,8] proposed a new data-driven design based on observer or parity space
monitoring. A fault estimation strategy based on parity space is designed for linear time-varying
systems [9]. A two-step incipient fault detection strategy is proposed based on fault extraction
and residual evaluation [10]. This scheme combines the subspace identification methods (SIM)
with the residual evaluation, which is different from the typical data-driven techniques. Its
advantage is to address process dynamics and obtain the process model identified from process
data. Similarly, this paper focuses on the method of fault detection and isolation based on
residual evaluation in parity space.
Fault isolation is performed to determine the fault roots when a fault alarm is triggered by the
fault detection system. The structured residual approach based on residual evaluation is a popular
tool for fault isolation. A residual that is insensitive to one subset of faults is generally sensitive
to other fault subsets. Li and Shah [11] presented a structured residual vector (SRV) approach to
identify sensor faults. Some expansible works have been presented in [12]. Partial principal
component models are introduced to optimize the structured residuals [13], which produce some
results in the integration of residual evaluation and data-driven methods. These methods,
however, need many prior knowledge of the process fault to identify the potential fault direction.
According to the classification of fault isolation methods [14], another fault isolation strategy is
based on contribution plots, in which only process data are used to find faulty variables without any
prior fault knowledge. A contribution plot is an unsupervised method that is commonly used to isolate
faults in multivariate statistical process monitoring. [15,16] summarized and reviewed the contribution
plot methods, such as complete decomposition contributions (CDC), partial decomposition
contributions (PDC) and reconstruction-based contributions (RBC), and they also analyzed their
diagnosable characteristics. A successful contribution plot should follow these properties: The
contribution of each variable has the same mean in normal conditions, and the fault variable has an
extremely large contribution value compared to other normal variables when a fault is presented. For
example, it can be proven that the RBC of a normal variable is less than or equal to that of a faulty
variable according to the Cauchy–Schwarz inequality analysis.
Sometimes the contribution plot suffers from the smearing effect, i.e., a non-faulty variable
exhibits a larger contribution value while the contribution of the faulty one is smaller. This
variable will be smeared out to the other variables in performing the PCA process [17]. Kerkhof
mathematically analyzed the smearing effect of three contributions: CDC, RBC and PDC [18].
The smearing effect is caused by the compression and expansion operation when process data are
mutually transformed between the measurement space and the latent space. A numerical case is
used to explain the smearing effect, in which contribution analysis cannot guarantee the correct
fault isolation in multiple sensor fault conditions. A data-driven fault isolation method is
J. Wang et al. / Journal of the Franklin Institute ] (]]]]) ]]]–]]] 3
proposed based on Bayesian decision theory that evaluates the current fault decision based on the
previous fault decision when fault evolution and propagation occur [19].
Some new approaches to fault isolation have been proposed recently. Liu demonstrated an
extended maximum variance unfolding algorithm [20], which aims to avoid the weakness of
kernel principal component analysis in nonlinear process monitoring. A new transfer entropy
method is used to isolate process faults [21] and can generate distinct patterns of information
flow and recognize correlations among process variables. Zhang proposed a multiscale nonlinear
fault diagnosis method, and derived the nonlinear scores of each scale [22]. State observers
[23,24] or a bank of isolation estimators [25] are designed for generating residuals that are only
sensitive to the corresponding fault subset. Yin proposed a new fault diagnosis approach for a
vehicle suspension system [26] and complex industrial system [27].
All of these fault isolation methods are based on either a data-driven method or a residual
evaluation. For example, the structured residual method is based on residual evaluation, and the
contribution plot technique is a data-driven approach. Though they are popular tools for fault
isolation, some weaknesses still exist such as the necessity of a prior fault set for the residual
evaluation method. Additionally, the smearing effect can cause mistaken isolation for the data-
driven method during the fault evolution.
A new fault isolation method integrating the residual evaluation and contribution plots is
proposed that can address the situation of fault evolution. The residual evaluation and data-driven
(REDD) techniques [7,8,10] are extended in this work, and a different isolation scheme is
considered. The fault isolation based on the REDD method should collect different fault sets to
obtain accurate fault isolation. To avoid this bottleneck, the contribution plot is introduced to
analyze more fault types. Meanwhile, the isolation statistic calculated from the REDD approach
can be used to eliminate the smearing effect. Furthermore, the current assessment of faulty
variables is based on previous samples, which are taken into account for obtaining information
on the fault evolution. Based on this theory, the process fault isolation strategy is demonstrated.
This article is organized as follows: Section 2 introduces fault isolation based on the traditional
PCA method and RBC. Section 3 gives the novel fault isolation, an integration of the residual
evaluation and contribution plot technique. A numerical example and a typical TE process are
applied to analyze the isolation performance of the proposed method in Section 4. Finally,
conclusions are given in Section 5.
2. PCA-based fault isolation
PCA is a statistical analysis theory based on process data. Consider the process data
X A ℜnm , where n is the number of samples and m is the number of sensors or variables. The
data matrix should first be normalized for a zero mean and unit variance for pre-processing.
Based on the singular valued decomposition (SVD) algorithm, data matrix X can be decomposed
as

XT UX Λ 0 T
S ¼ V UΛ UVT ¼ P P~ ~ P P~ ð1Þ
n1 0 Λ
where loading matrix P AℜmA consists of the A eigenvectors of S corresponding to the A larger
eigenvalues of diagonal matrix ΛA ℜAA . P~ Aℜmm A contains the remaining eigenvectors in
accordance with the diagonal matrix Λ~ A ℜm Am A . A is the number of principal components
(PCs), which can be determined based on the cumulative percent variance (CPV) method.
The new sample vector x A ℜm is used to calculate the fault detection statistics. The general
description based on the quadratic form [15] is
StatisticðxÞ ¼ ‖x‖2M ¼ xT Mx ð2Þ
where M is the representative of several specific fault detection statistics. For example, the M
representative for the squared prediction error (SPE) and Hotelling's T2 statistic is given as
~ ¼ P~ P~ T ; M 2 ¼ D ¼ PΛ1 PT
MSPE ¼ C ð3Þ
T
where MSPE and MT2 are symmetric and positive definite or semi-definite matrices.
The threshold of the SPE statistic is
0 qffiffiffiffiffiffiffiffiffiffiffiffi 11=h0
Cα 2θ2 h20 θ2 h0 ðh0 1ÞA
δ2α ¼ θ1 @ þ1þ ð4Þ
θ1 θ1 2
P
m
where δ2α stands for the threshold with a confidence level 1 α, θi ¼ λij , i ¼ 1; 2; 3,
j ¼ Aþ1
h0 ¼ 1 2θ1 θ3 =3θ1 2 . λij is the eigenvalue of S, and C α is the threshold of the standard normal
distribution corresponding to the confidence level 1 α.
The threshold formula of the T2 statistic is
Aðn2 1Þ
T 2α ¼ U F A;n A;α ð5Þ
nðn 1Þ
where T 2α stands for the threshold with confidence level 1 α, and F A;n A;α is the F-distribution
with A and n A degrees of freedom with confidence level 1 α.
The fault detection logic is to decide whether the detection statistic (Eq. (2)) exceeds its
threshold (Eq. (4) or Eq. (5)). If one of these statistics is greater than its threshold, a fault is
detected. If not, the current sample is under a normal condition. When a fault is detected, the
operators should focus on the specific variable or process loop. A contribution plot, which shows
the contribution value of each process variable independently, is popularly used to analyze fault
roots. The variable with the largest contribution value is considered to be the faulty variable.
A general decomposition of contributions [16] is given as follows:
GDCIndex
i ¼ xT M1 β ξi ξTi Mβ x; 0r β r1 ð6Þ
CDC and PDC are the special forms of contribution GDC, i. e.,
PDC Index
i ¼ xT Mξi ξTi x; β ¼ 0 ð7Þ
1
CDC Index
i ¼ xT M1=2 ξi ξTi M1=2 x; β ¼ ð8Þ
2
where i ¼ 1; 2; ⋯; m; M is determined by Eq. (3) and ξi is the i column of the identity matrix. It
should be noted that CDC is always positive.
The reconstruction-based contribution (RBC) was proposed in [15]. The main idea of
reconstruction is that if one variable is outside of its process data, its value can be estimated by
the remaining variables. Assuming that f i is the estimation of a real fault value, the
reconstruction-based contribution of variable xi along direction ξi is defined as
RBC i ¼ ‖ξi f i ‖2M ð9Þ
To find the proper f i , the difference between the reconstructed vector and variable sample
should be minimized:
Indexðx ξi f i Þ ¼ ‖x ξi f i ‖2M ð10Þ
This minimization is solved by
∂Indexðx ξi f i Þ 1 T
¼ 0 ) f i ¼ ξTi Mξi ξi Mx ð11Þ
∂f i
Substituting Eq. (11) into Eq. (9) yields
2
1 T ξTi Mx
RBC Index
i ¼x T
Mξi ξTi Mξi ξi Mx ¼ ð12Þ
ξTi Mξi
The i variable at the k time with the largest contribution is considered as the fault source,

CPIndex
k ¼ max GDCIndex i orRBC Index
i ð13Þ
i
where i ¼ 1; 2; ⋯; m.
The relative contribution is introduced to analyze the fault isolation results easily,
FCPIndex
i;k
RFCPIndex
i;k ¼ P
m ð14Þ
FCPIndex
i;k
i¼1
where FCPIndex
i;k ¼ GDC Index
i or RBC Index
i . The fault source can be written as
n o
RFCPIndex
k ¼ max RFCPIndex i;k ð15Þ
i
where i ¼ 1; 2; ⋯; m.
The RBC isolation method suffers from the smearing effect and it does not guarantee a correct
isolation for multiple sensor faults, although it is the most popular tool for fault isolation. The
isolation performance of RBC is thus insufficient [18]. Thus, a new isolation strategy is presented
to eliminate the smearing effect.
3. REDD-based fault isolation
Residual evaluation and residual generation have different types of technique [2]. Parity space
and subspace identification methods are used to generate residuals in this work. An optimized
parity space approach for the actuator FDI is used in [28]. Then, residual evaluation is combined
with a data-driven technique for application in practice [7,8]. This approach was also used for the
optimal fault detection of linear time-varying systems [29,30]. The current results of REDD are
simply explained in Section 3.1. The combination of the residual evaluation and contribution plot
analysis for fault isolation is presented in Section 3.2.
3.1. Residual evaluation based on parity space
Consider a process system,

xðk þ 1Þ ¼ AxðkÞ þ BuðkÞ þ wðkÞ
yðkÞ ¼ CxðkÞ þ DuðkÞ þ vðkÞ ð16Þ

where xA ℜ ; u A ℜ ; yA ℜ are the state variable vectors, input signals and measured output
n l m
signals, respectively, and wðkÞA ℜn ; vðkÞA ℜm are the process noise and measurement noise
with zero-mean value and normal distribution. They are independent of the input vector uðkÞ and
initial state vector xð0Þ. System matrices A,B,C,D and the system order n are generally unknown.
First, a data vector is defined as θðkÞ A ℜζ , which can represent xðkÞ; uðkÞ; yðkÞ, where the
corresponding ζ are n; l; m. Some new notations are introduced,
2 3
θðk s þ 1Þ
6 7
θs ðkÞ ¼ 4 ⋮ 5 A ℜs U ζ ; Ξk ¼ θðkÞ ⋯ θðk þ N 1Þ A ℜζN ð17Þ
θðkÞ
2 3
Ξs1
6 7
Ξp ¼ 4 ⋮ 5 ¼ θs ðkÞ ⋯ θs ðk þ N 1Þ A ℜs U ζN
Ξk
2 3
Ξkþ1
6 7
Ξf ¼ 4 ⋮ 5 ¼ θs ðk þ sÞ ⋯ θs ðk þ s þ N 1Þ A ℜs U ζN ð18Þ
Ξkþs
where s and the sample number N are selected by users such that s Z n and N is large enough. In
this description, subscript p means past data, and f means future data.
The standard framework of the process model for subspace identification can be obtained by
iterating Eq. (16),
Yf ¼ Γs Xk þ Hu;s Uf þ Hw;s Wf þ Vf ð19Þ
where Γs is the extended observability matrix and Hu;s ; Hw;s are the lower triangular block
Toeplitz matrixes,
2 3 2 3 2 3
C D 0 ⋯ 0 0 0 ⋯ 0
6 CA 7 6 CB D ⋯ 07 6 ⋯ 07
6 7 6 7 6 C 0 7
Γs ¼ 6 7; Hu;s ¼ 6 7; Hw;s ¼ 6 7
4 ⋮ 5 4 ⋮ ⋱ ⋱ ⋮5 4 ⋮ ⋱ ⋱ ⋮5
CAs 1 CAs 2 B ⋯ CB D CAs 2 ⋯ C 0
ð20Þ
The residual generator based on the parity space is
rs ðkÞ ¼ W0 ðYp ðkÞ Hu;s Up ðkÞÞ; W0 A Γs? ð21Þ
where Γs? is denoted as the left null space of Γs , i.e., Γs? Γs ¼ 0.
A residual can be generated from Eq. (21) with the identification of Γs? and Γs? Hu;s from
process data. A novel data-driven based identification approach was presented
h to obtain
i the
? ? T T
identification Γs and Γs Hu;s [31]. Constructing Hankle matrices Zp ¼ Yp Up ; Zf ¼
T T
h i
YTf UTf with past and future data, Eq. (19) can be rewritten as
Γs? ½I; Hu;s Zf ¼ Hw;s Wf þ Vf ð22Þ
Now, an instrumental variable ZTp is introduced for

1
lim Γ ? I; Hu;s Zf ZTp ¼ 0 ð23Þ
N-1 N s
Therefore, Γs? ½I; Hu;s is in the left null space of lim ð1=NÞZf ZTp . Then, performing
N-1
singular value decomposition on ð1=NÞZf ZTp yields
" #
1 Ez;1 0
Zf Zp ¼ Uz
T
VTz ð24Þ
N 0 Ez;2
where Uz A ℜsðlþmÞsðlþmÞ ; Vz A ℜsðlþmÞsðlþmÞ ; Ez;2 0, and

" # " #
U1 U11 U12
Uz ¼ ¼
U2 U21 U22
where U11 A ℜsmðslþnÞ ; UT12 A ℜðsm nÞsm ; UT22 A ℜðsm nÞsl . The identification results can be
obtained by
Γs? ¼ UT12 ; Γs? Hu;s ¼ UT22 ð25Þ
provided that the input excitation condition is
" # !
1 Xk T
rank Zp ¼ n þ sl ð26Þ
N Uf
However, the residuals generated by Eq. (21) do not have a maximized sensitivity to faults, so
an optimal vector-based approach is used to obtain the enhanced W0 [11].
Generally, the squared weighted residual (SWR) index is used for fault detection, and the
detection index is
( T
r s ðkÞRe 1 r s ðkÞ; if Re is known:
J¼ ð27Þ
r Ts ðkÞr s ðkÞ; if Re is unknown:
where Re is the covariance matrix of the residual, which is unknown in practice. Then, we make
the decision according to the diagnostic logic
(
JoJ th ; fault free
ð28Þ
J4J th ; fault
where J th ¼ χ 2α ðmÞ is the threshold.
3.2. Fault isolation based on REDD and contribution analysis
If a detection index is over the threshold, the fault isolation (FI) strategy is initiated. Fault
evolution usually happens over the monitor procedure. The proposed fault isolation method is
used to evaluate the current decision based on the previous faulty samples when the fault
evolution happened or the fault was propagated. The novel fault isolation approach based on
REDD and contribution analysis is motivated by this approach.
From Eqs. (2) and (3), one can obtain that
SPE ¼ ‖x‖2SPE ¼ xT P~ P~ x ¼ eT e
T
ð29Þ
where
e ¼ P~ x
T
ð30Þ
Eq. (30) demonstrates that the residual is obtained by projecting the new sampling data into
the modeling space. Similarly, the residual generator from Eq. (21) can be rearranged as
" #
h i Yp ðkÞ
rs ðkÞ ¼ W0 W0 Hu;s ¼ Φs Zp ðkÞ ð31Þ
Up ðkÞ
where
" #
h i Yp ðkÞ
Φ s ¼ W0 W0 Hu;s ; Zp ðkÞ ¼ ð32Þ
Up ðkÞ
Eq. (31) shows that the residual based on REDD can also be described as the projection from
the sample data (Zp ðkÞ) onto the modeling space (Φs ). Therefore, there are similar fault detection
methods for PCA and REDD, and it is helpful to explore the connection of fault isolation from
the view of contribution plot analysis.
There is many analytical information in the residual generator by Eq. (31), as the REDD is
based on the parity space approach. Liu noted that the fault isolation can be improved when
current samples and previous samples are used [14]. Thus, the previous samples are used for fault
isolation in this work. The parity space method has an important parameter s, which is called the
order of the parity vectors. The Hankle matrices involve the past or future s samples from
Eqs. (17) and (18). When the state order n is known, the dimension of rs ðkÞ is less than m Us.
Thus, the previous s 1 samples are used for fault isolation.
Furthermore, the residual from Eq. (31) can be divided into two parts,
rs ðkÞ ¼ Φs Zp ðkÞ
¼ ½Φs;y Yp ðkÞ; Φs;u Up ðkÞ ð33Þ
where Φs;y ¼ W0 ; Φs;u ¼ W0 Hu;s . Eq. (33) means that the residual is relevant to the process
output and input, and they provide information on the unknown variations such as faults,
disturbances and uncertainties. Therefore, the proposed fault isolation approach, Process Fault
Isolation (PFI), is presented.
Define the reduced order residual,
2 3
r new;1 ð1 : m; kÞ
6 r new;2 ðm þ 1 : 2m; kÞ 7
6 7
6 7
rnew ðkÞ ¼ Φs Zp ðkÞ ¼ 6
6 ⋮ 7
7 ð34Þ
6r 7
4 new;i ðm Uði 1Þ þ 1 : m Ui; kÞ 5
⋮
where i ¼ 1; ⋯; s 1, rnew ðkÞA ℜmðs 1Þ1 .

Then, the new statistic index at every sample k is defined as
1 X s1
IndexðkÞ ¼ ST ðkÞ ð35Þ
s 1 i ¼ 1 i
where
Si ðkÞ ¼ r new;i ðm Uði 1Þ þ 1 : m Ui; kÞ ð36Þ
where i ¼ 1; ⋯; s 1 and k represents the sample time. Eq. (35) means that the average residual
of the previous s 1 residuals is used to calculate the contribution.
Similar to the complete decomposition contributions (CDC), the contribution value of each
variable in this work is defined,
RContðj; kÞ ¼ IndexðkÞξTj ξj IndexT ðkÞ ð37Þ
where ξj is the j row of the identity matrix, j ¼ 1; …; m. Then, the fault root can be determined by
FI Index
k ¼ maxfRContðj; kÞg ð38Þ
j
The new isolation strategy of process faults is presented. This approach is motivated by a
contribution analysis, but there are several differences between the proposed method and the
contribution analysis. First, the variables are not compressed during the projection, and the
residual by Eq. (31) can improve the smearing effect. Second, the stack samples are utilized to
Fig. 1. The process fault isolation flow chart.
capture the fault evolution, and this can improve the accuracy of the fault isolation. Third, the
contribution index of Eq. (37) is simpler than the traditional index (i.e., RBC).
The relative contribution is introduced for easily analyzing the isolation results,
RContðj; kÞ
RF Index ¼ ð39Þ
j;k P
m
RContðj; kÞ
j¼1
The fault root of the k sample is

n o
RFI Index
k ¼ max RF Index
j;k ð40Þ
j
It is noteworthy that the contribution value of each variable under normal conditions follows a
random distribution or special probability distribution, so there is no leading variable. Thus, the
normal process and the fault stage can be easily distinguished. The REDD-based isolation
strategy is summarized as follows, and the isolation flow chart is shown in Fig. 1.
(1) Collect process data in normal operation conditions, and calculate the left null space of the
extended observability matrix Γs? . The residual generator can be obtained from Eq. (21) offline.
(2) The detection threshold is calculated according to J th ¼ χ 2α ðmÞ.
(3) Collect new data in each sample and normalize the data, and then the optimal residual by
Eq. (21) and the fault detection index by Eq. (27) can be obtained. The detection logic
Eq. (28) is used for determining whether a fault happened.
(4) If a fault is detected, the PFI scheme is performed. First, the isolation index is calculated by
Eq. (35), and the contribution index is calculated by Eq. (37) and Eq. (39).
(5) Eq. (40) is utilized to locate the primary faulty variable. The primary faulty variable can be
more than two over the fault evolution.
4. Illustrative examples
4.1. Numerical example
A numerical example is employed to demonstrate the fault isolation scheme based on REDD
and contribution analysis when multiple faults exist. The process model is given in [18],
2 3 2 3
x1 0:3441 0:4815 0:6637
6 7 6
6 x2 7 6 0:2313 0:5936 0:3545 7 72 t 3
6 7 6 7 1
6 x3 7 6 0:5060 0:2495 76 7
6 7 6 0:0739
74 t 2 5 þ noise
6 7¼6
6 x4 7 6 0:5552 0:2405 0:1123 7 7 t
6 7 6 7 3
6 x5 7 4 0:3371 0:3822 0:6115 5
4 5
x6 0:3877 0:3868 0:2045
where noise Nð0; 0:2 Þ, and t1, t2 and t3 are uniformly distributed with ranges of [0, 2], [0, 1.6]
2
and [0, 1.2], respectively. The simulated faults are

xf ¼ x þ f ξ i
where x is the fault-free data, and the fault magnitude f has a range of [0.1,4]. The fault direction
ξi is uniformly random among the six variable directions. Here, 480 normal samples and 960
faulty samples are generated from the process model in f¼ 1. The faulty variables are states x2,
x4, and the fault happens in the 480th sample.
The PCA model is built with an 85% cumulative percent variance and 99% confidence level. The
fault detection (FD) and isolation results of two faults that happen simultaneously are shown in Fig. 2.
Fig. 2(a) shows the fault detection results under two faulty variables. It is obvious that the statistics T2
and SPE both exceed their thresholds, and the faults are detected. The results of the fault isolation
calculated from the SPE statistic are shown in Fig. 2(b). The upper graphic of Fig. 2(b) is the faulty
root, and the lower one is the relative contribution of each variable. The X-axis is the sample quantity,
and the Y-axis is the variable number. It is shown that variable x6 designates the fault root and variable
x4 is the underlying faulty variable. The real faulty variables, however, are x2 and x4. Traditional RBC
analysis in the PCA process cannot guarantee correct fault source backtracking for multiple sensors’
faults, and the reason for the smearing effect has been analyzed [18].
The FDI results of the proposed REDD approach are shown in Fig. 3. The parameters of this
method are set as s ¼ 13, n ¼ 6. The proposed method detects the fault, as shown in Fig. 3(a).
Fig. 3(b) demonstrates that the variables x2, x4 and x5 are identified as faulty sources. The REDD
Fig. 2. Fault detection and isolation of two faults based on PCA.
Fig. 3. Fault detection and isolation of two faults based on REDD.
approach is an effective tool in the presence of multiple faults, although the variable x5 is
misdiagnosed. It is found that the smearing effect is weakened compared to the existing
contribution plot method shown in Fig. 2(b) and Fig. 3(b).
The cumulative distribution functions (CDF) of the Weibull distribution and the relative contribution
are given in Fig. 4 to illustrate the statistical property of contributions in the normal operation stage. The
CDF of the relative contributions in variable 4 based on PCA is shown in (a), and the CDF of the relative
contributions in variable 4 based on REDD is given in (b). It is shown that the relative contributions
follow the Weibull distribution, with scale and shape parameters of 0.15, 0.70 (PCA) and 0.21, 1.44
(REDD), respectively. Meanwhile, the mean of the relative contributions of variables 1-6 in the normal
process are shown in Fig. 5. Each variable has a similar mean in normal process operation, which is
approximately 0.13 0.22. These results give a basement level compared to faulty operation.
4.2. Tennessee Eastman process
4.2.1. TE benchmark process

The Tennessee Eastman (TE) process, a realistic simulation program of a typical chemical
plant, is a benchmark for the study of control and monitoring research. Fig. 6 gives the flow
Fig. 4. CDFs of relative contribution based on PCA and REDD.
Fig. 5. Mean of relative contribution based on PCA and REDD in normal operation.
diagram of the TE process. The TE process has five major units, including a reactor, condenser,
compressor, separator and stripper. The reaction involves eight components, which are four
reactants, two products, an inert and a by-product. This paper considers the base operating mode
and uses a decentralized control program [32].
The TE process datasets used in our studies can be downloaded from http://depts.washington.
edu/control/LARRY/TE/download.html, which are provided by N. L. Ricker. This TE process
includes 53 measurements: 41 process variables and 12 manipulated variables. Similar to [33],
16 main process variables are selected for the monitoring system shown in Table 1. 20 types of
process fault are given in the TE process operating, and only the several faults shown in Table 2
are considered.
Fig. 6. Tennessee Eastman process.
Table 1
Monitoring variables in TE process.
No. Measured variables No. Measured variables
1 A feed 9 Product separator temperature

2 D feed 10 Product separator pressure
3 E feed 11 Product separator underflow
4 A and C feed 12 Stripper pressure
5 Recycle flow 13 Stripper temperature
6 Reactor feed rate 14 Stripper steam flow
7 Reactor temperature 15 Reactor cooling water outlet temperature
8 Purge rate 16 Separator cooling water outlet temperature
First, two types of data set are obtained from the TE process, which include training and test
data. Training data are collected in a training set, i.e., the system is simulated under normal
operation for 24 h with a sampling interval of 3 min. Then, the size of the data matrix is
480 16. Correspondingly, eight test data sets are collected (including one in normal operation,
six single-fault operations and one multi-fault operation with fault 2 and fault 7 simultaneously).
Each data set lasts for 48 operation hours, and the size of every data matrix is 960 16. All faults
were introduced at the eighth hour, i.e., at the 161th sampling of the operating time. 11
manipulated variables (except the reactor agitator speed) are used as the input variables of the
REDD approach.
4.2.2. Fault isolation result analysis

The fault isolation results of the proposed method and the traditional contribution plot based
on the PCA method are compared for the process faults in this section.
The PCA model based on the training data was first obtained offline, and 9 principal
components that account for 85% of the total variance were selected according to the cumulative
percent variance. Then, the test data are used to verify the performance of the monitoring system
based on the PCA model with the significance level 99%. A residual generator of the REDD
approach was also built based on the same training data set. The alternative parameters, i.e., the
order of the parity space s and the system order n, are set as s ¼ 13, n ¼ 6. Meanwhile, we definite
the fault isolation rate (FIR) for assessing the performance of the proposed approach and the
Table 2
Operational faults.
FaultNo. Process variables Type
2 B composition, A/C ration constant Step

4 Reactor cooling water inlet temperature Step
7 C header pressure loss-reduced availability Step
8 A, B, and C feed composition Random variation
13 Reaction kinetics Slow drift
19 Unknown Unknown
Fig. 7. Fault detection and isolation of fault 2 based on RBC.
RBC based on PCA.

the number of samples that variable is considered as the fault source
FIR ¼ 100%
the number of fault samples
The isolation results of fault 2, fault 4 and fault 2þfault 7 are discussed to illustrate the
effectiveness of the proposed method in this experiment.
4.2.2.1. Case I. Fault 2 occurs in the 161th sampling. Fig. 7 demonstrates the results of the fault
detection and isolation for fault 2 based on the RBC approach. Fig. 7(a) shows that a process fault can
be detected when the T2 and SPE statistics are both over their thresholds. Then, a fault isolation
strategy should be carried out. The upper part of Fig. 7(b) only shows the variables with the maximal
relative contribution, while the lower part shows the relative contributions for all variables at any time.
Variable 8 is the primary fault source, with FIR¼ 81.00%, and variable 3 is the minor faulty root, with
FIR¼ 10.38%. Furthermore, many variables are affected by the smearing effect according to the
relative contribution analysis.
Fig. 8. Fault detection and isolation of fault 2 based on REDD.
Fig. 9. CDFs of relative contribution based on RBC (a) and REDD (b).
Let us consider the results of fault detection and isolation for fault 2 based on REDD analysis, as
shown in Fig. 8. The fault can be detected by residual assessment, but the fault source is determined
differently from the RBC isolation. The upper part of Fig. 8(b) shows that variable 1 is the primary
faulty variable, with FIR¼ 100%. Thus, a higher FIR demonstrates that the proposed approach has a
better isolation performance, and the smearing effect is thereby clearly eliminated.
The reasonability of the isolation results can be analyzed over the underlying process
operation according to original paper [34]. Fault 2 is a step change in the B composition and the
A/C ratio constant (stream 4). B is the inert component, which will cause a change in the purge
rate, i.e., variable No. 8 in stream 9. Furthermore, the A/C ratio constant will be reflected in
stream 4, and the action of some control loops will also change process variables No.1 and No.3.
Therefore, the isolation results of both the RBC-based and REDD-based approaches are
reasonable according to the physical analysis of the TE plant.
Figs. 9 and 10 show the statistical properties of the relative contributions in the normal stage of
the TE process, in which the results of RBC are given in (a) and those of REDD are given in (b).
The CDFs of the relative contributions via variable 10 in the normal stage are compared in Fig. 9.
It is shown that the relative contribution of variable 5 follows the Weibull distribution, with the
0.1 0.14
0.09
0.12
0.08
The mean of relative contribution
The mean of relative contribution
0.07 0.1
0.06
0.08
0.05
0.04 0.06
0.03
0.04
0.02
0.01 0.02
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16
Variables Variables
Fig. 10. Mean of relative contribution based on RBC (a) and REDD (b) in the normal operation.
Fig. 11. Fault detection and isolation for fault 4 based on RBC.
scale and shape parameters being 0.01, 0.67 (PCA) and 0.06, 1.12 (REDD), respectively. The
contributions of different variables have similar means in the normal condition, are shown in
Fig. 10. Most means of the relative contributions are under the value of 0.14.
4.2.2.2. Case II. Fault 4 occurs in the 161th sampling. Fault 4 is a step change in the reactor
cooling water inlet temperature. When fault 4 happens, the reactor temperature and the separator feed
also vary. The separator cooling water outlet temperature, i.e., variable No.16, must change upon
adapting the variation of the separator unit. When the fault is propagated from steam 8 to stream 6, the
feeds of all compositions should be changed, i.e., variables No. 3 and No. 4 are affected.
Fig. 11 shows the isolation results of the RBC-based method, where the relative contribution is
calculated according to the SPE statistic. It is clear that the smearing effect is very serious, as
shown in Fig. 11(b), where no variables can be identified as the faulty root. The relative
contribution of each variable is similar, and no obvious difference exists between the normal
operation and the faulty condition (after the 161th sample). However, the REDD-based approach
is more valid than the RBC-based method, whose isolation results are given in Fig. 12. Variables
Fig. 12. Fault detection and isolation for fault 4 based on REDD.
Fig. 13. Fault detection and isolation for fault 2 and 7 based on RBC.
3000 20
Thershold
Residual 15
Variables
2500 the moment of fault added
10
the max contribution
5
2000
0
Residual
0 100 200 300 400 500 600 700 800 900 1000
1500 Samples
15
1000 0.4
Variables
10 0.3
500 0.2
5
0.1
0
0 100 200 300 400 500 600 700 800 900 1000 100 200 300 400 500 600 700 800 900
Samples Samples
Fig. 14. Fault detection and isolation for fault 2 and 7 based on REDD.
No. 3 and No. 16 are identified as the primary faults, as shown in Fig. 12(b), and variable No.
4 is diagnosed as the minor faulty variable. Their corresponding FIRs are 73.13%, 28.75% and
6.25%, respectively. The compared results show that the REDD-based method is reliable,
although the smearing effect still exists. A minor faulty variable set is still helpful for operators or
engineers to locate a faulty root.
4.2.2.3. Case III. Multiple faults 2 and 7 occur in the 161th sampling. To illustrate the isolation
performance of multiple faults, fault 2 and fault 7 are considered simultaneously. The FDI results
based on the RBC and REDD approaches are shown in Fig. 13 and Fig. 14, respectively. Fig. 13
(b) shows that only one primary faulty variable is located, which is almost same as in Fig. 7(b). The
FIR of variable No. 8 is 82.13%, which is 81.00% in Case I, so only fault 2 occurs. Similarly, the FIR
of variable No. 3 is 9.875%, and the corresponding value is 10.38% in Case I. Thus, the traditional
RBC method does not identify the faulty root in the presence of multiple faults.
The REDD approach can locate the primary faulty variables, as shown in Fig. 14(b). The upper part
of Fig. 14(b) demonstrates that variables No. 11 and No. 16 are the major faulty roots, and the faulty
variable is evolved under 400–500 samples. This result illustrates that fault 2 and fault 7 are highly
correlated with the product separator underflow (No.11) and separator cooling water outlet
temperature (No.16). The fault evolution shows that the two faults have an important impact on the
process system. The color map Fig. 14(b) shows that the smearing effect is mostly removed.
Fault evolution is difficult to handle in an industrial process because the wrong warning or variable
indicator can lead to incorrect operation, which can cause the decline of product quality or a serious
accident. Fault 7 is the step change of the C header pressure loss, and fault 2 is the B composition
variation. The two faults mainly affect stream 4 in the TE process. Thus, the stripper is affected first,
and then the change is transmitted through the recycle flow. The feed of the vap/liq separator is
changed, so the cooling water outlet temperature (No. 16) is influenced significantly. The separator
underflow (No.11) gradually becomes the main faulty variable as the influence of No. 16 increases.
This is explained that the temperature of the separator changes significantly, so the vapor–liquid
equilibrium varies. Finally, the underflow varies in a large range.
4.2.3. Comparison
The performances of the three different methods using the fault isolation rate (FIR) on the six
faults are compared in Tables 3–5. A Bayesian theory based data-driven fault isolation approach
[19] is used for comparison beside the RBC-based and REDD-based approaches discussed
above. The FIR of each variable that is considered a faulty root is used to provide a quantitative
basis for the fault analysis or technical improvement.
Table 3 illustrates that the RBC-based method has a very serious smearing effect. Because the
SPE statistic has better isolation than the T2 statistic, the former is used for FIR analysis. It is
shown that 8 variables’ FIRs are not zero when fault 2 occurs, which means that half of the 16
Table 3
FIRs (%) based on the RBC (SPE statistic).
Variable no. Fault 2 Fault 4 Fault 7 Fault 8 Fault 13 Fault 19
1 3.75 7.88 7.63 13.1 2.63 4.00

2 5.00 9.75 8.00 12.1 10.8 6.25
3 10.4 10.0 9.88 15.5 4.63 18.1
4 0 12.4 11.6 12.0 2.63 8.00
5 0 15.8 15.3 13.4 2.13 7.00
6 3.88 7.13 8.88 6.88 4.00 8.00
7 0 12.9 11.8 7.63 2.13 7.75
8 81.0 5.25 7.00 4.25 2.38 12.3
9 0 1.38 1.63 1.38 3.63 1.25
10 0 0.125 0.125 0 0 0
11 3.00 11.4 11.3 7.75 3.88 17.8
12 0 0.250 0.250 0 0 0
13 0 2.88 3.50 4.13 4.88 6.38
14 0 10.5 10.9 3.38 2.75 7.50
15 4.13 8.25 9.00 16.8 72.0 11.5
16 0.875 4.25 3.38 1.75 1.63 4.25
Table 4
FIRs (%) based on the Bayesian inference (sigmoid¼35).
1 100 16.4 26.0 98.6 0.250 30.4

2 0 12.3 12.1 0 98.3 24.8
3 0 9.00 9.63 98.4 0 7.25
4 0 12.5 25.3 0 0.125 7.75
5 0 9.38 8.25 0 0 34.0
6 0 12.1 18.8 0 0 7.00
7 0 9.50 21.0 0.250 0.250 6.13
8 0 13.8 17.0 0 0 33.8
9 100 19.8 26.0 0.250 98.5 38.5
10 100 26.6 32.9 0.250 0.250 53.5
11 0 8.75 6.88 0 0 9.75
12 100 26.4 33.4 0.250 0.250 36.4
13 0.125 16.3 26.3 0.125 0.125 24.8
14 0 8.38 7.38 98.3 0 23.6
15 0 7.63 16.4 98.3 0 19.4
16 0 14.4 18.9 0 0 28.4
Table 5
FIRs (%) based on the REDD-based method.
1 100 2.63 3.00 20.8 22.5 13.9

2 3.25 3.13 93.4 9.50 58.4 42.6
3 0 28.8 3.00 51.3 4.00 10.3
4 0 6.25 0 1.63 1.00 2.13
5 1.00 0 0 1.88 2.13 8.38
6 0 0 0 0 0.250 1.38
7 0.375 0 0 1.63 8.38 2.75
8 0.375 0.375 0 6.88 1.00 0.500
9 0 0 0 0.500 0.625 0
10 0.875 0 0 0 1.25 2.13
11 0 0 1.25 2.13 1.38 6.88
12 0 0 0 0.875 1.00 0.750
13 0 0 0 0 0.125 0
14 0 0 1.25 0 1.25 0
15 2.50 1.63 0 2.00 9.00 7.00
16 5.25 73.1 13.4 19.6 7.75 20.4
variables can be considered as fault sources. The other faults of different types, such as step type,
random variation, slow drift and unknown type, have more serious results, i.e., there is nearly no
zero value in FIR. Many variables have similar FIRs, and the faulty root is ambiguous for
engineers.
Table 4 shows the effectiveness of the Bayesian inference method, especially for fault 13. The
higher FIR is powerful information for engineers to locate faulty roots, whereas a better isolation
result can be obtained only when the fault detection strategy is valid, such as in the cases of fault
2, fault 8 and fault 13. Unapparent detection results may lead to invalid isolation, such as in the
case of fault 4, fault 7 and fault 19.
Table 5 shows that a better isolation performance can be obtained based on the proposed
REDD approach. First, the primary faulty variable can be located according to the higher FIR
and the difference in FIRs among different variables. Second, the REDD-based method can
generally remove the smearing effect. There are no more than two variables that have higher
FIRs in this isolation scheme. It is easy for operators to diagnose the real fault source from this
minor fault set. Third, the proposed method is effective for different types of faults.
5. Conclusions
A new fault isolation strategy for industrial processes is presented based on the combination of
residual evaluation and contribution plot analysis. The proposed method uses the parity space to
generate residuals for the fault detection and isolation. Hence, this proposed method can be used
to obtain a novel fault isolation scheme. A new contribution index was obtained based on the
average values of previous several samples, which aims to eliminate the smearing effect. This is
because the previous residuals would influence the fault isolation results in the next sample.
Then, the fault evolution for an actual process can be acquired, and the fault information can be
provided in a timely fashion for the engineers.
The process fault isolation strategy is examined through a numerical example and TE process.
The results demonstrate that a single fault or multiple faults can be isolated successfully even
under conditions of fault propagation. The smearing effect, which is the weakness of the fault
isolation schemes based on PCA, is removed mostly by the proposed isolation scheme.
A comparison among different approaches was conducted using the fault isolation rate (FIR),
which is used to quantitatively analyze the fault isolation performance. The proposed method
shows a better isolation performance, and the primary faulty variable can be located accurately
for all types of faults, especially for unknown types.
Acknowledgments
We are deeply indebted to China Scholarship Council (CSC) for sponsoring the research in
University of Delaware. Special thanks should be given to Prof. Babatunde A. Ogunnaike for his
instructive advice and useful suggestions on this work.
References
[1] A. Varga, New computational paradigms in solving fault detection and isolation problems, Annu. Rev. Control 37
(1) (2013) 25–42, http://dx.doi.org/10.1016/j.arcontrol.2013.03.001.
[2] I. Hwang, S. Kim, Y. Kim, C. Seah, A survey of fault detection, isolation, and reconfiguration methods, IEEE Trans.
Control Syst. Technol. 18 (3) (2010) 636–653, http://dx.doi.org/10.1109/TCST.2009.2026285.
[3] S. Yin, S.X. Ding, X. Xie, H. Luo, A review on basic data-driven approaches for industrial process monitoring,
IEEE Trans. Ind. Electron. 61 (11) (2014) 6418–6428, http://dx.doi.org/10.1109/TIE.2014.2301773.
[4] Y.W. Zhang, J. An, C. Ma, Fault detection of non-gaussian processes based on model migration, IEEE Trans.
Control Syst. Technol. 21 (5) (2013) 1517–1526, http://dx.doi.org/10.1109/TCST.2012.2217966.
[5] Y.W. Zhang, S.J. Qin, Fault detection of nonlinear processes using multiway kernel independent component
analysis, Ind. Eng. Chem. Res. 46 (23) (2007) 7780–7787, http://dx.doi.org/10.1021/ie070381q.
[6] U. Schubert, U. Kruger, Harvey Arellano-Garcia, Thiago de Sa´ Feital, Gunter Wozny, Unified model-based fault
diagnosis for three industrial application studies, Control Eng. Pract. 19 (5) (2011) 479–490, http://dx.doi.org/10.1016
/ j.conengprac. 2011.01.009.
[7] S.X. Ding, Data-driven design of monitoring and diagnosis systems for dynamic processes: A review of subspace
technique based schemes and some recent results, J. Process Control 24 (2) (2014) 431–449, http://dx.doi.org/
10.1016/j.jprocont.2013.08.011.
[8] S.X. Ding, P. Zhang, A. Naik, E.L. Ding, B. Huang, Subspace method aided data-driven design of fault detection and
isolation systems, J. Process Control 19 (9) (2009) 1496–1510, http://dx.doi.org/10.1016/j.jprocont.2009.07.005.
[9] M. Zhong, S.X. Ding, Q. Han, Q. Ding, Parity space-based fault estimation for linear discrete time-varying systems,
IEEE Trans. Autom. Control 55 (7) (2010) 1726–1731, http://dx.doi.org/10.1109/TAC.2010.2047672.
[10] W.S. Ge, J. Wang, J.L. Zhou, H. Wu, Q.B. Jin, Incipient fault detection based on fault extraction and residual
evaluation, Ind. Eng. Chem. Res. 54 (14) (2015) 3664–3677, http://dx.doi.org/10.1021/acs.iecr.5b00567.
[11] W. Li, S. Shah, Structured residual vector-based approach to sensor fault detection and isolation, J. Process Control
12 (3) (2002) 429–443, http://dx.doi.org/10.1016/ S0959-1524(01)00046-4.
[12] W. Lin, S.J. Qin, Optimal structured residual approach for improved faulty sensor diagnosis, Ind. Eng. Chem. Res.
44 (7) (2005) 2117–2124, http://dx.doi.org/10.1021/ ie049213d.
[13] J. Gertler, J. Cao, Design of optimal structured residuals from partial principal component models for fault diagnosis
in linear systems, J. Process Control 15 (5) (2005) 585–603, http://dx.doi.org/10.1016/j.jprocont.2004.10.005.
[14] J. Liu, Fault diagnosis using contribution plots without smearing effect on non-faulty variables, J. Process Control
22 (9) (2012) 1609–1623, http://dx.doi.org/10.1016/ j.jprocont.2012.06.016.
[15] C.F. Alcala, S.J. Qin, Reconstruction-based contribution for process monitoring, Automatica 45 (7) (2009)
1593–1600, http://dx.doi.org/10.1016/j.automatica.2009.02.027.
[16] C.F. Alcala, S.J. Qin, Analysis and generalization of fault diagnosis methods for process monitoring, J. Process
Control 21 (3) (2011) 322–330, http://dx.doi.org/10.1016/ j.jprocont.2010.10.005.
[17] J. Westerhuis, S. Gurden, A. Smilde, Generalized contribution plots in multivariate statistical process monitoring,
Chemom. Intell. Lab. Syst. 51 (1) (2000) 95–114, http://dx.doi.org/10.1016/S0169-7439(00)00062-9.
[18] P.V. Kerkhof, J. Vanlaer, G. Gins, J.F.M. Van Impe, Analysis of smearing-out in contribution plot based fault isolation for
statistical process control, Chem. Eng. Sci. 104 (2013) 285–293, http://dx.doi.org/10.1016/j.ces.2013.08.007.
[19] J. Liu, D. Wong, D. Chen, Bayesian filtering of the smearing effect: fault isolation in chemical process monitoring,
J. Process Control 24 (3) (2014) 1–21, http://dx.doi.org/10.1016/j.jprocont.2013.12.018.
[20] Y. Liu, T. Chen, Y. Yao, Nonlinear process monitoring and fault isolation using extended maximum variance
unfolding, J. Process Control 24 (6) (2014) 880–891, http://dx.doi.org/10.1016/j.jprocont.2014.04.004.
[21] P. Hajihosseini, K. Salahshoor, B. Moshiri, Process fault isolation based on transfer entropy algorithm, ISA Trans.
53 (2) (2014) 230–240, http://dx.doi.org/10.1016/ j.isatra.2013.11.007.
[22] Y.W. Zhang, C. Ma, Fault diagnosis of nonlinear processes using multiscale KPCA and multiscale KPLS, Chem.
Eng. Sci. 66 (1) (2011) 64–72, http://dx.doi.org/10.1016/j.ces.2010.10.008.
[23] M. Du, J. Scott, P. Mhaskar, Actuator and sensor fault isolation of nonlinear process systems, Chem. Eng. Sci. 104
(2013) 294–303, http://dx.doi.org/10.1016/ j.ces.2013.08.009.
[24] S. Yin, X. Zhu, J. Qiu, H. Gao, State estimation in nonlinear system using sequential evolutionary filter, IEEE Trans.
Ind. Electron. 63 (6) (2016) 3786–3794, http://dx.doi.org/10.1109/TIE.2016.2522382.
[25] X. Zhang, M.M. Polycarpou, T. Parisini, Design and analysis of a fault isolation scheme for a class of uncertain
nonlinear systems, Annu. Rev. Control 32 (1) (2008) 107–121, http://dx.doi.org/10.1016/j.arcontrol.2008.03.007.
[26] Shen Yin, Z. Huang, Performance monitoring for vehicle suspension system via fuzzy positivistic C-means
clustering based on accelerometer measurements, IEEE/ASME Trans. Mechatron. 20 (5) (2015)
2613–2620, http://dx.doi.org/10.1109/ TMECH.2014.2358674.
[27] Shen Yin, X. Zhu, O. Kaynak, Improved PLS focused on key-performance-indicator-related fault diagnosis, IEEE
Trans. Ind. Electron. 62 (3) (2015) 1651–1658, http://dx.doi.org/10.1109/TIE.2014.2345331.
[28] H.M. Odendaal, T. Jones, Actuator fault detection and isolation: an optimised parity space approach, Control Eng.
Pract. 26 (2014) 222–232, http://dx.doi.org/10.1016/ j.conengprac.2014.01.013.
[29] M. Zhong, S.X. Ding, E.L. Ding, Optimal fault detection for linear discrete time-varying systems, Automatica 46 (8)
(2010) 1395–1400, http://dx.doi.org/10.1016/ j.automatica.2010.05.022.
[30] Y. Li, M. Zhong, Optimal fault detection for a class of discrete-time switched linear systems, J. Syst. Eng. Electron.
24 (3) (2013) 512–518, http://dx.doi.org/10.1109/ JSEE.2013.00059.
[31] J. Wang, S.J. Qin, Closed-loop subspace identification using the parity space, Automatica 42 (2) (2006)
315–320, http://dx.doi.org/10.1016/j.automatica.2005.09.012.
[32] N.L. Ricker, Decentralized control of the Tennessee Eastman challenge process, J. Process Control 6 (4) (1996)
205–221, http://dx.doi.org/10.1016/0959-1524(96)00031-5.
[33] G. Chen, T.J. McAvoy, Predictive on-line monitoring of continuous processes, J. Process Control 8 (5-6) (1998)
409–420, http://dx.doi.org/10.1016/S0959-1524(98)00023-7.
[34] J.J. Downs, E.F. Vogel, A plant-wide industrial process control problem, Comput. Chem. Eng. 17 (3) (1993)
245–255, http://dx.doi.org/10.1016/0098-1354(93)80018-I.

Wang 2016

Enviado por

Dados do documento

Direitos autorais

Formatos disponíveis

Compartilhar este documento

Compartilhar ou incorporar documento

Opções de compartilhamento

Você considera este documento útil?

Este conteúdo é inapropriado?

Direitos autorais:

Formatos disponíveis

Wang 2016

Enviado por

Direitos autorais:

Formatos disponíveis

Available online at www.sciencedirect.

Journal of the Franklin Institute ] (]]]]) ]]]–]]]

Fault isolation based on residual evaluation

2. PCA-based fault isolation

3. REDD-based fault isolation

3.1. Residual evaluation based on parity space

Consider a process system,

yðkÞ ¼ CxðkÞ þ DuðkÞ þ vðkÞ ð16Þ

Now, an instrumental variable ZTp is introduced for

where Uz A ℜsðlþmÞsðlþmÞ ; Vz A ℜsðlþmÞsðlþmÞ ; Ez;2 0, and

where J th ¼ χ 2α ðmÞ is the threshold.

3.2. Fault isolation based on REDD and contribution analysis

where i ¼ 1; ⋯; s 1, rnew ðkÞA ℜmðs 1Þ1 .

Fig. 1. The process fault isolation ﬂow chart.

The fault root of the k sample is

4.1. Numerical example

and [0, 1.2], respectively. The simulated faults are

Fig. 2. Fault detection and isolation of two faults based on PCA.

Fig. 3. Fault detection and isolation of two faults based on REDD.

4.2. Tennessee Eastman process

4.2.1. TE benchmark process

Fig. 4. CDFs of relative contribution based on PCA and REDD.

Fig. 6. Tennessee Eastman process.

No. Measured variables No. Measured variables

1 A feed 9 Product separator temperature

4.2.2. Fault isolation result analysis

FaultNo. Process variables Type

2 B composition, A/C ration constant Step

Fig. 7. Fault detection and isolation of fault 2 based on RBC.

RBC based on PCA.

Fig. 8. Fault detection and isolation of fault 2 based on REDD.

The mean of relative contribution

Variable no. Fault 2 Fault 4 Fault 7 Fault 8 Fault 13 Fault 19

1 3.75 7.88 7.63 13.1 2.63 4.00

Variable no. Fault 2 Fault 4 Fault 7 Fault 8 Fault 13 Fault 19

1 100 16.4 26.0 98.6 0.250 30.4

Variable no. Fault 2 Fault 4 Fault 7 Fault 8 Fault 13 Fault 19

1 100 2.63 3.00 20.8 22.5 13.9

Você também pode gostar