Escolar Documentos
Profissional Documentos
Cultura Documentos
Neurocomputing
journal homepage: www.elsevier.com/locate/neucom
art ic l e i nf o a b s t r a c t
Article history: With the increase of scale and complexity of modern chemical process, fault diagnosis and detection are
Received 11 October 2015 playing crucial roles in process monitoring. Accidents can be avoided if faults can be detected and
Received in revised form excluded in time. In this paper, Principal Components Analysis (PCA) and Recursive Feature Elimination
7 December 2015
(RFE) are combined with Support Vector Machine (SVM) for fault diagnosis and detection. Specifically,
Accepted 6 January 2016
the original SVM, PCA-SVM and SVM-RFE are respectively utilized to identify three faults from the
Communicated by Xudong Zhao
Available online 2 February 2016 simulation of Fed-Batch Fermentation Penicillin (FBFP) process. Experimental results show that PCA-SVM
and SVM-RFE perform better than the original SVM, and the fault detection schemes based on PCA-SVM
Keywords: and SVM-RFE generate satisfactory results.
Fault detection
& 2016 Elsevier B.V. All rights reserved.
Fault diagnosis
Support vector machine
Principal components analysis
Fed-batch fermentation penicillin
http://dx.doi.org/10.1016/j.neucom.2016.01.027
0925-2312/& 2016 Elsevier B.V. All rights reserved.
118 C. Yang, J. Hou / Neurocomputing 190 (2016) 117–123
independent of sample dimension. The SRM theory has been The separating interval can be calculated as J w2 J . The objective
proposed to minimize the structural risk, i.e., the sum of empirical function to obtain the separating hyperplane can be written as
risk and confidence risk. By means of SRM, SVM possesses
minϕðwÞ ¼ 12 J w J 2 ð3Þ
remarkable generalization ability.
The original SVM can be applied to two-class classification In order to solve Eq. (3), Lagrange function is introduced
problems. The 1-v-r SVMs, 1-v-1 SVMs and H-SVMs methods can Lðw; d; aÞ ¼ 12 J w J aðyððw xÞ þ dÞ 1Þ ð4Þ
be used to solve multi-class classification generally.
In the 1-v-r SVMs method, each class of samples are regarded In Eq. (4), aðai 40Þ is the Lagrange multiplier. Then the problem
as in a category, and the rest are treated as in another category. In is converted to the corresponding dual problem
total c classification models will be constructed for classification X
l
1Xl Xl
problem with c class of samples. The advantages of 1-v-r SVMs lie max Q ðaÞ ¼ aj a a y y ðx x Þ
j¼1
2i¼1j¼1 i j i j i j
in the simplicity and computation efficiency. However, the short-
coming is that the classification may be inseparable. If there are X
l
many classes of samples, and one class of the samples is con- s:t: aj yj ¼ 0; j ¼ 1; 2; …; l; aj Z 0; j ¼ 1; 2; …; l ð5Þ
j¼1
siderably less than the sum of the others, this imbalance will affect
the classification accuracy. The optimal solution calculated by Eq. (5) is an ¼ ðan1 ; an2 ; …; anl ÞT ,
n
In the 1-v-1 SVMs method, the classification models are trained and the optimal weight vector wn and d are written as
with every two classes of samples. In total cðc 2 1Þ classification
X
l
models are trained for c classes of training samples. The training wn ¼ anj yj xj ð6Þ
samples of each classifiers are relevant, and the voting method is j¼1
used for classification.
All classes of samples are divided into two sub-categories in H- n
X
l
d ¼ yi yj anj ðxj xi Þ ð7Þ
SVMs method, and the sub-categories will be divided into another
j¼1
two categories after tested by SVM. After some iterations, the
single category can be obtained. Therefore H-SVMs method avoids Therefore, the optimal separating hyperplane is written as
n
the case that samples cannot be divided, and it has a stable pro- ðwn xÞ þ d ¼ 0, and the optimal separating function is
80 1 9
motion performance [31]. < X l =
n n
f ðxÞ ¼ sgnfðw xÞ þ d g ¼ sgn @ n A
aj yj ðxj xi Þ þ d
n
ð8Þ
: ;
2.2. Principal components analysis j¼1
X ¼ TP ¼ TP T þ T~ P~ ¼ TP T þ E
T T
ð11Þ
T ¼ XP ð12Þ
E ¼ T~ P~ ¼ XðI PP T Þ
T
ð13Þ
nd md
T AR and P A R represent scoring matrix and loading
matrix respectively. E A Rnm is residual matrix and d o m is the
number of principal components. According to Eqs. (11) and (12),
the projection is scoring matrix in each direction of loading
matrix, and the loading vector of principal components is actually
the covariance matrix M of the data matrix X:
1 T
M¼ X X ð14Þ Fig. 1. Fed-batch fermentation penicillin process.
n1
Therefore, the correlation between variables can be reflected by
the principal components obtained by projection. When the PCA Table 1
statistics is used for fault diagnosis and detection, the PCA T 2 Simulation conditions of normal and fault operations.
statistics for the training data and the testing data are
Simulation conditions Normal Fault 1 Fault 2 Fault 3 Unit
T 2 ¼ x T P ΛP T x ð15Þ dataset
The PCA SPE statistics for the training data and the testing data Initial conditions
are Substrate concentration 15 15 15 15 g/L
Dissolved oxygen 1.16 1.16 1.16 1.16 g/L
T T
SPE ¼ x ðI PP Þx ð16Þ concentration
Biomass concentration 0.1 0.1 0.1 0.1 g/L
The threshold setting of PCA-SVM is the same as the one of Penicillin concentration 0 0 0 0 g/L
basic PCA, if the values of T 2 and SPE are under the corresponding Culture volume 100 100 100 100 L
thresholds, the data can be categorized as normal, otherwise the Carbon dioxide concentration 0.5 0.5 0.5 0.5 g/L
pH 5 5 5 5
data will be treated as faulty. Fermenter temperature 298 298 298 298 K
Generated heat 0 0 0 0 kcal
Set points
4. Fed‐batch fermentation penicillin process Aeration rate 8.6 8.6 8.6 6 L/h
Agitator power 29.9 21 29.9 29.9 W
Substrate feed flowrate 0.0426 0.0426 0.027 0.0426 L/h
The FBFP process is a typical dynamic, non-linear, time-varying Substrate feed temperature 296 296 296 296 K
and multi-stage batch process. In general, the process can be Temperature set point 298 298 298 298 K
divided into three stages, namely cell growth stage, penicillin pH set point 5 5 5 5
Sampling Interval 1 1 1 1 h
synthesis stage and cell autolysis stage. In the cell growth phase,
Simulation Time 400 400 400 400 h
nutrient substance of the initial substrate is consumed and new
cells are synthesized constantly. After cell growth stage, penicillin
synthesis stage begins and productivity is maximized until the
ability of penicillin synthesis declines. Cell autolysis is the final
Table 2
stage, where pH of the fermentation broth increases and the Variables of FBFP process.
ability to synthesize penicillin also weakens. Throughout the
entire fermentation process, many factors affect the effectiveness Number Process variable Unit
of penicillin fermentation, such as temperature, pH, substrate
1 Aeration rate L/h
concentration, dissolved oxygen concentration and so on. There- 2 Agitator power W
fore it is imperative to conduct effective process monitoring. 3 Substrate feed temperature K
4 Substrate concentration g/L
5 Dissolved oxygen concentration g/L
6 Biomass concentration g/L
5. Fault diagnosis 7 Penicillin concentration g/L
8 Culture volume L
Professor Cinar of Illinois Institute of Technology developed a 9 Carbon dioxide concentration g/L
simulation software pensim2.0 [38] for penicillin fermentation 10 pH
11 Fermenter temperature K
monitoring and process modeling in 2002, in order to simulate the
12 Cooling water flow rate L/h
aeration rate, biomass concentration, substrate feed temperature
and other variables of the fermentation process. This software
provides a benchmark simulation platform for penicillin fermen-
initial phase of about 45 h and the fed-batch stage of around
tation modeling, optimal control and process monitoring, and
many scholars have done a lot of works [39–42] in penicillin fer- 355 h. The simulation conditions of normal and fault operations
mentation process monitoring and fault diagnosis with the plat- are listed in Table 1, and the sample number of each data set is 400
form. In this paper, the data sets were generated from the simu- since the sampling interval is chosen as 1 h. The data sets con-
lator on the web site http://simulator.iit.edu/web/pensim/simul. sisting of 12 process variables are listed in Table 2.
html. Fermentation Process Flow Chart of the benchmark is shown Fault detection of test samples is based on SVM. The classifi-
in Fig. 1. The whole duration of each batch is 400 h, comprising cation of the test samples can be obtained with the model trained
120 C. Yang, J. Hou / Neurocomputing 190 (2016) 117–123
by the training data sets. The test sample is normal if the predicted 1.5
label of the sample is 0. Otherwise, the test sample is faulty.
The feature ranking list is obtained based on SVM-RFE and the
most relevant variables are shown in Table 3. The cause of failure
1
can be analyzed by the relevant variables.
As shown in Fig. 2, the threshold of PCA T 2 is 18.4753 and the
Table 3 20
0 50 100 150 200 250 300 350 400
The most relevant variable for each fault.
Samples
Fault 1 2
Fault 2 11 1000
Fault 3 1
800
600
T2
40
T
20 3
SPE
0 2
50 100 150 200 250 300 350 400
0.04 Fig. 5. Fault detection based on the PCA statistics for Fault 1.
0.02
5.2. Fault 2
0
50 100 150 200 250 300 350 400 As shown in Fig. 6, the majority of the samples are classified
samples
correctly. The accuracy of fault detection for fault 2 based on SVM
Fig. 2. The PCA statistics under normal conditions. is 92%, lower than that of PCA.
C. Yang, J. Hou / Neurocomputing 190 (2016) 117–123 121
1.5 100
80
60
T2
1 40
The class of the data
20
0
50 100 150 200 250 300 350 400
0.5
0.5
0.4
0 0.3
SPE
0.2
0.1
−0.5 0
50 100 150 200 250 300 350 400 50 100 150 200 250 300 350 400
Test dataset sample
Fig. 6. Fault detection based on SVM for Fault 2. Fig. 8. Fault detection based on the PCA statistics for Fault 2.
90
normal data set
1.5
training data set
80 testing data set
70
1
60
Variable 11
50
40 0.5
30
20
0
10
0
0 50 100 150 200 250 300 350 400
Samples −0.5
50 100 150 200 250 300 350 400
Fig. 7. Plot of variable 11 in Fault 2. Test dataset
As shown in Fig. 11, PCA statistics are sensitive to fault 3 and Compared with the original SVM algorithm, SVM-RFE and PCA-
the samples of testing data set are all above the threshold, that SVM improve the classification accuracy significantly. While the
means all the data are faulty. The accuracy of fault detection based classification accuracy of PCA-SVM is slightly lower than that of
on PCA statistics is 99.5%, much higher than that of SVM. Com- SVM-RFE, the efficiency of PCA-SVM is obviously better than that
pared with SVM classification method, PCA statistics can be used of SVM-RFE.
to simplify the classification model and improve classification
accuracy in fault detection.
6. Conclusion
5.4. Fault detection
In this paper, SVM, PCA and RFE are firstly reviewed. PCA and
The most relevant variables can be obtained by feature extraction RFE are combined with SVM for fault diagnosis and detection.
based on RFE. In order to obtain the maximum classification accuracy, Then, the original SVM, PCA-SVM and SVM-RFE are respectively
different numbers of features are chosen in prediction by SVM-RFE utilized to identify three faults from the simulation of FBFP
and the results are shown in Table 4. Choosing an appropriate number process.
of variables is helpful to improve the classification accuracy of fault As a result, the classification accuracy of SVM is lower than that
detection. In contrast, if the number of the variables is improper, it will of SVM-RFE and PCA-SVM evidently. PCA-SVM statistical threshold
will result in a decrease in classification accuracy. Therefore, SVM-RFE calculated by the normal data set of FBFP process is used to predict
can be used to maximize the classification accuracy of fault detection whether the samples are faulty or not. The classification accuracy
by selecting the best combination of the variables. based on SVM-RFE can be improved by extracting the most rele-
In the PCA-SVM algorithm, we firstly pre-process the raw data vant variables. Experimental results show that the classification
to remove inappropriate variables. Then we use the PCA method accuracy of SVM-RFE is not only higher than that of original SVM
for feature extraction and dimensionality reduction to obtain the obviously, but also slightly higher than that of PCA-SVM. However,
the classification accuracy may be influenced, if the number of
main features of each fault. In this way, the efficiency and accuracy
features is selected inappropriately.
of PCA statistics based classification is improved.
SVM, SVM-RFE and PCA-SVM are used to detect the three kinds
of faults, and the classification accuracy is shown in Table 5.
References
1500
[1] S.J. Qin, Survey on data-driven industrial process monitoring and diagnosis,
Annu. Rev. Control 36 (2) (2012) 220–234.
1000 [2] S. Yin, O. Kaynak, Big data for modern industry: challenges and trends, Proc.
IEEE 103 (2) (2015) 143–146.
2
T
[3] A. Pouliezos, G.S. Stavrakakis, Real Time Fault Monitoring of Industrial Pro-
500 cesses, vol. 12, Springer Science& Business Media, Dordrecht, 2013.
[4] S. Yin, Z. Huang, Performance monitoring for vehicle suspension system via
fuzzy positivistic c-means clustering based on accelerometer measurements,
0 IEEE/ASME Trans. Mechatron. 20 (5) (2014) 2613–2620.
50 100 150 200 250 300 350 400 [5] S. Yin, X. Xie, J. Lam, K. Cheung, H. Gao, An improved incremental learning
approach for kpi prognosis of dynamic fuel cell system, IEEE Trans. Cybern.
30
http://dx.doi.org/10.1109/TCYB.2015.2498194.
[6] T.J. Rato, M.S. Reis, Fault detection in the tennessee eastman benchmark
process using dynamic principal components analysis based on decorrelated
20
residuals (dpca-dr), Chemomet. Intell. Lab. Syst. 125 (2013) 101–108.
SPE
[20] J.V. Kresta, J.F. MacGregor, T.E. Marlin, Multivariate statistical monitoring of [39] C. Ündey, E. Tatara, A. Çınar, Intelligent real-time performance monitoring and
process operating performance, Can. J. Chem. Eng. 69 (1) (1991) 35–47. quality prediction for batch/fed-batch cultivations, J. Biotechnol. 108 (1)
[21] K. Detroja, R. Gudi, S. Patwardhan, Plant-wide detection and diagnosis using (2004) 61–77.
correspondence analysis, Control Eng. Pract. 15 (12) (2007) 1468–1483. [40] Y. LIU, H.-q. WANG, Pensim simulator and its application in penicillin fer-
[22] E. Vigneau, D. Bertrand, E.M. Qannari, Application of latent root regression for mentation process, J. Syst. Simul. 12 (12) (2006) 3524–3527.
calibration in near-infrared spectroscopy. Comparison with principal compo- [41] S.K. Maiti, R.K. Srivastava, M. Bhushan, P.P. Wangikar, Real time phase detec-
nent regression and partial least squares, Chemomet. Intell. Lab. Syst. 35 (2) tion based online monitoring of batch fermentation processes, Process Bio-
(1996) 231–238. chem. 44 (8) (2009) 799–811.
[23] E.L. Russell, L.H. Chiang, R.D. Braatz, Fault detection in industrial processes [42] J.-M. Lee, C.K. Yoo, I.-B. Lee, Enhanced process monitoring of fed-batch peni-
using canonical variate analysis and dynamic principal component analysis, cillin cultivation using time-varying and multivariate statistical analysis, J.
Chemomet. Intell. Lab. Syst. 51 (1) (2000) 81–93. Biotechnol. 110 (2) (2004) 119–136.
[24] J.F. MacGregor, C. Jaeckle, C. Kiparissides, M. Koutoudi, Process monitoring and
diagnosis by multiblock pls methods, AIChE J. 40 (5) (1994) 826–838.
[25] V. Venkatasubramanian, R. Rengaswamy, S.N. Kavuri, K. Yin, A review of
process fault detection and diagnosis: Part iii: process history based methods,
Chengming Yang received the B.E. degree in agri-
Comput. Chem. Eng. 27 (3) (2003) 327–346.
cultural mechanization and automation from Northeast
[26] C.-W. Hsu, C.-J. Lin, A comparison of methods for multiclass support vector
Agricultural University, Harbin, China, in 2010, the M.E.
machines, IEEE Trans. Neural Netw. 13 (2) (2002) 415–425.
degree in agricultural mechanization engineering from
[27] M. Ge, R. Du, G. Zhang, Y. Xu, Fault diagnosis using support vector machine
Kunming University of Science and Technology,
with an application in sheet metal stamping operations, Mech. Syst. Signal
Kunming, China, in 2013. He is currently working
Process. 18 (1) (2004) 143–159.
toward the Ph.D. degree in control science and engi-
[28] H.J. Shin, D.-H. Eom, S.-S. Kim, One-class support vector machines: an appli-
neering at Harbin Institute of Technology.
cation in machine fault detection and classification, Comput. Ind. Eng. 48 (2)
His research interests include fault diagnosis, fault
(2005) 395–408.
tolerant control, process monitoring and their applica-
[29] V.N. Vapnik, V. Vapnik, Statistical Learning Theory, vol. 1, Wiley, New York,
tions to large-scale industrial processes.
1998.
[30] V. Vapnik, The Nature of Statistical Learning Theory, Springer Science& Busi-
ness Media, New York, 2013.
[31] S. Abe, Support Vector Machines for Pattern Classification, vol. 2, Springer,
London, 2005.
[32] J.E. Jackson, Principal components and factor analysis: Part i—principal com- Jian Hou received his Ph.D. degree in 2007 from Harbin
ponents, J. Quality Technol. 12 (4) (1980) 201–213. Institute of Technology, China. From 2007 to 2010, he
[33] P. Nomikos, J.F. MacGregor, Monitoring batch processes using multiway worked as a postdoctoral researcher in National Uni-
principal component analysis, AIChE J. 40 (8) (1994) 1361–1375. versity of Singapore, Singapore and University of
[34] R. Kohavi, G.H. John, Wrappers for feature subset selection, Artif. Intell. 97 (1) Venice, Italy. In 2011, he was with Ningbo Institute of
(1997) 273–324. Materials Technology and Engineering of Chinese
[35] S.S. Keerthi, C.-J. Lin, Asymptotic behaviors of support vector machines with Academy of Sciences and Xuchang University, China. He
Gaussian kernel, Neural Comput. 15 (7) (2003) 1667–1689. joined the School of Information Science and Technol-
[36] S. Yin, G. Wang, X. Yang, Robust pls approach for kpi-related prediction and ogy at Bohai University, China in 2012, where he is
diagnosis against outliers and missing data, Int. J. Syst. Sci. 45 (7) (2014) currently an associate professor. His research interests
1375–1382. include computer vision, pattern recognition and
[37] L. Xie, X. Lin, J. Zeng, Shrinking principal component analysis for enhanced machine learning. He is a member of the IEEE.
process monitoring and fault isolation, Indust. Eng. Chem. Res. 52 (49) (2013)
17475–17486.
[38] G. Birol, C. Ündey, A. Cinar, A modular simulation package for fed-batch fer-
mentation: penicillin production, Comput. Chem. Eng. 26 (11) (2002)
1553–1565.