Вы находитесь на странице: 1из 6

Proceedings of the 30th Chinese Control Conference

July 22-24, 2011, Yantai, China

A New Fault Isolation Method Based on Unified Contribution Plots


DENG Xiaogang, TIAN Xuemin
College of Information and Control Engineering, China University of Petroleum, Dongying 257061, P. R.China
E-mail: dengxiaogang@upc.edu.cn, tianxm@upc.edu.cn

Abstract: Contribution plots are important to identify the unusual process variables for fault isolation. In this paper a new fault
isolation method based on unified contribution plots was presented for linear and nonlinear principal component analysis.
Principal component analysis based process monitoring was first analyzed and variable contributions to monitoring statistic were
modified to bring a new unified framework for contribution calculation. Then this method was generalized to nonlinear kernel
principal component analysis. Some issues were further discussed including negative contribution and confidence limit, and then
relative contribution was introduced to give a more reasonable explanation on variable contribution. Simulations on a continuous
stirred tank reactor system were performed to show the proposed method is effective to isolate fault variables.
Key Words: Contribution Plots, Fault Isolation, Kernel Principal Component Analysis

contribution calculating framework for linear and nonlinear


1 Introduction methods.
Computer control systems of modern industrial processes The rest of this paper is organized as follows. In Section 2,
allow large amounts of plant data to be collected and stored. a brief review of contribution plots in PCA-based process
This motivates the development of the data-driven fault monitoring is described and then new contribution plots are
diagnosis method, which is also referred to as the proposed to modify the traditional contribution plots.
Multivariate Statistical Process Monitoring (MSPM) Section 3 will generalize the modified contribution plots to
technique. One of well-known MSPM methods is principal nonlinear PCA as KPCA. In Section 4 a new unified
component analysis (PCA) and recently some nonlinear contribution calculating framework is summarized and
principal component analysis methods have emerged in this some issues are further discussed. Case studies on a
field such as kernel principal component analysis (KPCA) continuous stirred tank reactor (CSTR) system are given in
and neural PCA. Both linear and nonlinear methods are Section 5. Finally we present our conclusions in Section 6.
effective in fault detection but have great difficulty in
diagnosing the potential root causes. The main reason for 2 Contribution Plots for PCA
this is that the monitoring charts based on detection statistics
indicate the deviation from normal operating conditions but 2.1 Traditional contribution plots
do not give information on what is wrong with the process, In PCA, the data matrix X is decomposed into principal
or which process variable causes the process out of control. component (PC) subspace and residual subspace.
As a challenging research topic, the task of fault variable X TP T  E (1)
nu m
isolation has attracted much attention from researchers. This where X  R represents m variables with n samples, T is
was firstly studied by Miller et al.[1] and MacGregor et al.[2] score matrix in which each row is score vector
which proposed the contribution plots to isolate fault t [t1 , t2 ,..., tk ] , P is loading matrix, E is residual matrix.
variables by comparing the contribution of each process
When PCA is used in fault detection, two monitoring
variable to the monitoring statistic. In the contribution plots,
statistics are constructed for monitoring.
the variable with the greatest contribution value usually
indicates fault sources. Based on Miller’s study, Westerhuis T 2 = t 1t T = xP 1PT x T (2)
2
et al.[3], Alcala et al.[4], Alawi et al[5], Alvarez et al.[6] and ˆ T = x(I - PPT )x T
ˆ - x)
Q = e = (x - x)(x xPr Pr T x T (3)
Li et al.[7] made further discussion on the application and where x is a measurement vector with m variables and x̂ is
development of contribution plots. However most of studies a reconstructed vector,  is the diagonal matrix of the
on contribution plots were focused on calculating variable eigenvalues associated with score vectors, Pr is an
contribution to fault in a constructive way in linear PCA and
these methods can not be generalized to nonlinear PCA. Cho orthonormal matrix which meets Pr Pr T  PP T I .
et al.[8] indicates the contribution plot for PCA cannot be Westerhuis et al.[3] discussed contribution plots theory
directly applied to the case of nonlinear kernel PCA. For the in batch process and continuous process and extended it
purpose of building a general contribution plots theory for from PCA models to latent variable models with correlated
linear and nonlinear PCA, this paper will analyze the scores. In their work, the contribution to the Q-statistic for
contribution plots in PCA-based methods and present a new process variable j is calculated as Eq.(4).
cQj (enew, j )2 ( xnew, j  xˆnew, j )2 (4)
where xˆnew, j is reconstruction variable corresponding to a
*
This work is supported by the Fundamental Research Funds for the new measurement xnew, j .
Central Universities (No. 10CX04046A) and the Natural Science
Foundation of Shandong Province, China (Grant No. Y2007G49).

4280
Contribution to D-statistic (T2-statistic) for continuous 1
processes is described in Eq.(5). In Eq.(12) T can be proved easily for PCA-based
2
c Dj t Tnew / 1[ xnew, j p j (PT P)1 ]T (5) process monitoring, so the variable contribution can be
where P T P I for PCA method, t new is scores calculated as
1 w
corresponding to new measurement xnew ,  is the diagonal C_T 2 ( x j ) new xj (xP 1P T xT )
matrix of the eigenvalues associated with score vectors 2 wx j (13)
t new , p j is the jth row of P. xP 1p j T x j t 1 ( x j p j T )
The existing work has made a lot of discussions on 1 w
contribution plots and displayed their effectiveness in some C_Q( x j ) new xj (xPr Pr T xT )
2 wx j (14)
fault scenarios. However, deep theory explanations and
generalization to nonlinear PCA are still needed. So this x j xPr p r j T
x j ( x j  xˆ j ) x je j
paper is to make further study of contribution calculation in where p j represents the jth row of Pˈ p rj represents the jth
the next section.
row of Pr .
2.2 New contribution plots Compared to traditional methods in Eq.(4)(5),
For simplification the statistics T2 and Q are replaced Eq.(13)(14) are called new contribution plots methods. It is
with a variable symbol f as Eq. (6). noted that in both methods contribution to T2 is essentially
same but contribution to Q in the new method makes
f = x6xT (6) important modification.
1
where  P P for T statistic and  I  PP for Q
T 2 T
3 Contribution Plots for Nonlinear PCA
statistic.
Variable f is a function of measured variables Recently nonlinear PCA has become a hot research topic.
x [ x1 , x2 ,..., xm ] expressed using Eq.(7) Kernel PCA(KPCA) was proved to be simple and effective
f f ( x1 , x2 ,..., xm ) (7) as a nonlinear PCA method whose nonlinear mapping is
accomplished by the kernel function. Originally Kernel
Where x j is the jth variable. PCA was developed by Schölkopf et al. [9] and has been
From Lagrange’s mean value theorem, Eq. (8) is applied in fault diagnosis and process monitoring by Lee et
obtained. al.[10] and Sun et al.[11]. The studies on KPCA show that it
f ( x1 ,..., xm )  f ( x10 ,..., xm 0 ) can detect fault effectively but it is hard to compute the
§ contributions of original process variables in KPCA-based
w w ·
¨ ( x1  x10 )  ...  ( xm  xm 0 ) ¸ f ( x10 (8) monitoring because it is difficult or even impossible to find
© wx1 wxm ¹ an inverse mapping function from the feature space to the
 T ( x1  x10 ),..., xm 0  T ( xm  xm 0 )) (0  T  1) original space. In this paper, we can give contribution plots
Given x10 0,..., xm 0 0 , the above equation can be for KPCA method using the new contribution framework.
KPCA is to map the input space into a feature space by
formulated as nonlinear transformation and extract the principal
f ( x1 ,..., xm ) components in feature space. It is assumed that ) (.) is a
§ w w · (9) nonlinear mapping function that projects the data in input
f (0,..., 0)  ¨ x1  ...  xm ¸ f (T x1 ,..., T xm )
© wx1 wxm ¹ space to feature space and the covariance matrix in feature
From Eq.(9), it is seen that each statistic can be space can be expressed as:
1 n
decomposed as a sum of each variable’s derivative which
can be viewed as a variable contribution. So jth variable
cov() ( X)) ¦
n 1 j 1
)(x j )T )(x j ) (15)
contribution to PCA-based monitoring statistic is defined by where xj (j=1,…,n) is the jth row of X. In the feature space
Eq.(10). the eigenvalue problem must be solved for KPCA
w decomposition.
C_f ( x j ) x j f (T x1 ,...,T xm ) (10)
wx j cov() ( X))p i Oi p i (16)
PCA-based monitoring statistics T2 and Q have some It should be pointed out that it is difficult to solve the
similar interesting characteristics: statistics are zero when problem of Eq. (16) as ) (.) can not be obtained in most
the normalized variable x equals vector zero and two cases. However, there exist coefficients D ij (j=1,2,…n)
statistics have similar structure. According to these property,
Equation (11)(12) are given as such that
n
f (0,...,0) 0 (11) pi ¦D ij )(x j )T (17)
w w j 1
f (T x1 ,...,T xm ) T 2 f ( x1 ,..., xm ) (12) Combining Eqs.(15), (16) and (17), the following
wx j wx j
equation can be obtained
According to Eq.(9), it can be concluded that the (n  1)Oi  i K i (18)
monitoring statistic is the sum of all variables’ contribution.
where K is defined as [K]ij=k(xi,,xj) = ) ( xi ),) ( x j ) .

4281
One can avoid nonlinear mapping by defining kernel Q (k (xt , xt )  tt T ) (k (xt , xt )  K t T  T K t ) (21)
function k(x, y) =  ) ( x )ˈ) ( y ) ! . With the use of kernel where K t is normalized kernel matrix,  is transformation
function, the score vectors are expressed as: matrix,  is weighted matrix.
n
t i  p i , ) (x) ! ¦D j 1
ij  ) (x j ), ) (x) ! (19) According to Eq.(10), variable contribution to statistics
can be calculated by Eq.(22)(23). In this study, radial basis
2
Two statistics used in KPCA-based process monitoring kernel function k(x, y) = exp˄ x  y c ) is chosen, so
are T2 and Q: Eqs.(26)-(28) can be available.
T 2 t/ 1t T K t T  -1 T K t (20)

w wK t T
C _ T 2 ( xt j ) xtj (K t T  -1 T K t ) 2 xt j  -1 T K t (22)
wxtj wxt j
T xt T xt

w § wk (xt , xt ) wK t T T ·
C _ Q ( xt j ) xtj (k (x t , x t )  K t T  T K t ) xt j ¨ 2  K t ¸ (23)
wxtj T xt ¨ wx wxtj ¸
© tj ¹ T xt
where
wK t T wk (T xt , x1 ) wk (T xt , x 2 ) wk (T xt , x n ) (24)
[ ... ]
wxtj wxtj wxtj wxtj
T xt

2 n 1 n n
k (T xt , T xt ) 1  ¦ exp( T xt  xl ¦¦ exp( x
2 2
/V) l  xk /V) (25)
nl1 n2 l 1 k 1

wk (T xt , T xt ) 2 n 1
¦ 2T (T xtj  xlj ) exp( T xt  xl /V)
2
(26)
wxtj n l 1V
1 n 1 n 1 n n

¦ exp( T xt  xl ¦ exp( xi  xl ¦¦ exp( x


2 2 2 2
k (T xt , xi ) exp(  T xt  xi /V)  /V) /V)  l  xk / V ) (27)
nl1 nl1 n2 l 1 k 1

wk (T xt , xi ) 1 1 n 1
¦ 2T (T xtj  xlj ) exp( T xt  xl
2 2
 2T (T xtj  xij ) exp( T xt  xi /V) /V) (28)
wxtj V n l 1V

In Eq.(22)(23), T is required for contribution graphic. Another issue is a contribution’s confidence limit.
calculation. There are two calculating methods: one is to Currently the contribution plots are a graphical
calculate T according to Eq.(9), by which a precise value of representation of the original absolute contributions and
T can be available. But we think this is not necessary confidence limits are not used to identify the bounds of
because it is complex and time-consuming to compute a normal behavior for individual multivariate statistical
different T for each sample. So another substituted method contributions. Colin et al.[12] has given three methods to
is to use an approximation of T which can be obtained by calculate the derivation in confidence limits which in
some normal behavior data sets. instinct make use of the standard derivation of the
contribution. Considering each variable’s contribution with
4 A Unified Framework and Some Issues a different confidence limit, so relative contribution is built
for more accurate fault isolation. Its expression is shown in
In the above analysis a new method to calculate variable
Eq.(30).
contribution has been discussed for both PCA and KPCA.
Analogously, the proposed method can be generalized to C_ f  mean(C_ f normal )
C_R f (30)
other similar MSPM methods, so a unified framework for std (C_ f normal )
MSPM contribution plots can be built for fault isolation. For
where C_ f normal is variable contribution under normal
any statistics expressed by f, variable contribution to f can
be given using Eq.(29) operation condition.
w
C_f ( x j ) x j f (T x1 ,...,T x m ), (0  T  1) (29) 5 Case Study
wx j
The proposed fault isolation strategy is tested with a
where T can be approximately estimated by some normal simulated process, a non-isothermal continuous stirred tank
behavior data sets for the purpose of fault variable isolation. reactor (CSTR) system. A schematic diagram of the CSTR
with feedback control system is shown in Fig. 1. It is
There are some issues to be discussed about new assumed that the classic first order irreversible reaction
contribution plots. The first topic is about negative happens in CSTR. The flow of solvent and reactant A into a
contributions. The contribution of process variables in reactor produces a single component B as an outlet stream.
Eq.(29) can be positive or negative .We think the sign of the Heat from the exothermic reaction is removed through
contributions is not important and the absolute value cooling flow of jacket. The reactor temperature is controlled
indicates the affections of process variables to statistics. So to set-point by manipulating the coolant flow. The level is
the absolute value of contributions is used in a contribution controlled by manipulating the outlet flow.

4282
Table 1 Process faults for CSTR system
CAF QF TF LT LC
Fault Description
QC TC
QC TCF
F1 Step change in feed flow rate.
h
F2 The feed temperature ramps up or down.
FC FT
TC TT F3 The feed concentration ramps up or down.
T F4 The hear transfer coefficient ramps down.
F5 Catalyst deactivation.
TC TT
CA Q F6 The coolant feed temperature ramps up or down.
Fig.1. A diagram of the CSTR system F7 Set point change for the reactor temperature.
The process data sets including normal operating
F8 The feed temperature measurement has a bias
situation and ten kinds of fault pattern are generated by
simulating the CSTR process. The applied fault pattern F9 The reactor temperature measurement has a bias
˄ Fault F1~F10 ˅ can be seen in Table 1. These faults F10 The coolant valve was fixed at the steady value.
contain process change, sensor malfunction and valve faults.
In order to test data-driven process monitoring methods, Variable contributions are plotted to identify fault
some important variables should be measured and stored in variables for early remedy. To avoid accident error at some
simulation procedure. Gaussian noise is added to all specific sample, 20 samples from a fault detection sample
measured variables in simulation procedure. The simulation are used to calculate the average contributions. When
interval is set to 2 seconds but in our study database records traditional contribution plot method is applied, Fig.3
one sample every 12 seconds to store long time period data indicates that contribution plots for T2 are more apparent to
sets. By simulating normal situation and fault patterns, 500 isolate fault while contribution plots to Q still cannot isolate
samples are collected for algorithm study and every fault is accurate variables. But if the new contribution plots method
introduced at the 200th sample. is applied without considering of relative contribution, it is
Fault F3 is illustrated for fault detection and isolation, concluded that contributions to T2 and Q both list variable 7
which means the feed concentration change corresponding as the maximum contribution in Fig.4. However, in Fig.4(b)
to variable 7. PCA is performed for process monitoring and the difference between variable contribution is not obvious.
the monitoring statistics alarm threshold is set to a 99% With the application of relative contribution, the isolation
confidence limit. According to the fault detection results in plots in Fig.5 find the right abnormal variables more clearly.
Fig.2, T2 statistic exceeds the confidence limit clearly at the
287th sample.
300 30

200 20
PCA, Q
PCA, T2

100 10

0 0
0 100 200 300 400 500 0 100 200 300 400 500
Sample Number Sample Number
(a) T 2 statistic (b) Q statistic
Fig.2. PCA monitoring with 99% confidence limit for Fault F3
traditional method traditional method
30 0.8
Cont to Q, PCA
Cont to T2, PCA

0.6
20
0.4
10
0.2

0 0
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
Variable Number Variable Number
(a) T 2 statistic (b) Q statistic
Fig.3. Average contribution plot for Fault F3 (traditional method)

4283
new method new method
30 1

Cont to Q, PCA
Cont to T2, PCA
20
0.5
10

0 0
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
Variable Number Variable Number
(a) T 2statistic (b) Q statistic
Fig.4. Average contribution plot for Fault F3 (new method)
new method new method
20 8

Relative Cont to Q, PCA


Relative Cont to T2, PCA

15 6

10 4

5 2

0 0
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
Variable Number Variable Number
(a) T2 statistic (b) Q statistic
Fig.5. Relative contribution plot for Fault F3 (new method)
600 3

400 2
KPCA,Q
KPCA,T2

200 1

0 0
0 100 200 300 400 500 0 100 200 300 400 500
Sample Number Sample Number
(a) T 2 statistic (b) Q statistic
Fig.6. KPCA monitoring with 99% confidence limit for Fault F3

10 25
Relative Cont to Q, KPCA
Relative Cont to T2, KPCA

8 20

6 15

4 10

2 5

0 0
1 2 3 4 5 6 7 8 9 10 1 2 3 4 5 6 7 8 9 10
Variable Number Variable Number
(a) T 2 statistic (b) Q statistic
Fig.7. Relative contribution plot for Fault F3 when KPCA is applied

If KPCA is applied to monitor the CSTR process, it has indicated as the greatest contribution variables. These
better detection ability as nonlinear method. From Fig.6, conclusions are consistent to the process mechanism. So
fault F3 is detected at 265th sample in KPCA-based KPCA with new contribution method not only can detect
monitoring and then the new contribution method is faults more effectively, but can isolate fault variable
performed to isolate fault variables. The isolation results for correctly.
fault F3 are illustrated in Fig. 7 which show that variable 7 is

4284
6. Conclusions [2] J. F. MacGregor, T. Kourti, Statistical process control of
multivariate processes, Control Engineering Practice, 3:
403-414, 1995.
In this paper a unified contribution calculating framework [3] J. A. Westerhuis, S. P. Gurden, and A. K. Smilde,
was built for linear PCA and nonlinear KPCA which Gerneralized contribution plots in multivariate statistical
provides a clear and straightforward solution for fault process monitoring, Chemometrics and Intelligent
variable location. The proposed method can also be Laboratory Systems, 51: 95-114, 2000.
generalized to some classical MSPM methods such as [4] C. F. Alcala, S. J. Qin. Reconstruction-based contribution for
canonical variate analysis and independent component process monitoring. Automatic, 45(7):1593-1600, 2009.
analysis. In fact, many other linear and nonlinear statistical [5] A. Alawi, S. W. Choi, E. Martin, et al. Sensor fault
monitoring methods can be also discussed in this framework identification using weighted combined contribution plots.
which is a topic for future study. The issues of negative Fault Detection, Supervision and Safety of Technical
Processes 2006, pages:908-913
contributions and confidence limits provide a more
[6] C. R. Alvarez, A. Brandolin, M. C. Sanchez, On the variable
reasonable interpretation of contribution plots and lead to contributions to the D-statistic, Chemometrics and Intelligent
relative contribution. The simulation results on the CSTR Laboratory Systems, 88:189-196, 2007.
system reveal that the application of the new method can [7] G. Li, S. Z. Qin, Y. D. Ji, et al. Total PLS based contribution
give more accurate contribution plots than traditional plots for fault diagnosis. Acta Automation Sincia,
method. 35(6):759-765, 2009.
However, it should not be overestimated that contribution [8] J.-H. Cho, J.-M. Lee, S. W. Choi, et al, Fault identification for
plots can make a perfect solution for fault source process monitoring using kernel principal component analysis,
identification in data-driven process monitoring. Chemical Engineering Science, 60: 279-288, 2005.
[9] B. Schölkopf, A. J. Smola , K. Muller . Nonlinear component
Contribution plots can only be viewed as first step of fault
analysis as a kernel eigenvalue problem, Neural Computation,
identification and provide some important help on initial 10 : 1299-1319,1998.
fault location, and it can not diagnose complex faults which [10] J.-M. Lee, C. K. Yoo, S. W. Choi, et al. Nonlinear process
involve many fault variables. More deep study about fault monitoring using kernel principal component analysis,
isolation methods and fault propagation process analysis Chemical Engineering Science, 59:223-234, 2004.
should be developed. [11] R. Sun, F. Tsung, L. Qu. Evolving kernel principal component
analysis for fault diagnosis, Computers & Industrial
Engineering, 53(2):361-371, 2007.
References [12] A. K. Colin, E. B. Martin, A. J. Morris. Confidence limits for
contribution plots, Journal of Chemometrics, 14: 725-736,
[1] P. Miller, R. Swanson, and C. Heckler, Contribution plots: a
2000.
missing link in multivariate quality control, 37th Annual Fall
Technical Conference of ASQC, Rochester, NY.1993.

4285

Вам также может понравиться