New Fusion Architectures for Performance Enhancement of a PCA-Based Fault Diagnosis and Isolation System S. Pirooz Azad*, S. Bahrampour*, B. Moshiri*, K. Salahshoor** * Control and Intelligence Processing Center of Excellence, School of ECE, University of Tehran, Iran (e-mail: s.piroozazad@ece.ut.ac.ir, s.bahrampour@ece.ut.ac.ir, moshiri@ut.ac.ir). **Petroleum University of Technology, Tehran, Iran (e-mail: salahshoor@put.ac.ir) Abstract: This paper proposes some new fusion architectures to enhance the diagnostic operation of a principal component analysis (PCA)-based fault diagnosis and isolation (FDI) system. The first approach presents a serial classifier fusion methodology by incorporating a support vector machine (SVM) classifier to diagnose PCA-based fault features. Then, parallel fusion architecture is proposed to fuse the fault diagnostics due to the individual SVM, Bayes and K-nearest neighbor (K-NN) classifiers. The two series and parallel fusion architectures are finally incorporated into a new combined framework to yield a more efficient and powerful diagnostic capabilities. Extensive simulation test experiments are conducted to demonstrate the comparative performances of the new proposed fusion-based FDI systems on the Tennessee Eastman (TE) process plant as a large-scale benchmark problem. 1. INTRODUCTION Large amounts of data are collected in many industrial processes for monitoring and control purposes. The most important parts of condition monitoring are fault detection and isolation (FDI). Fault detection is used to determine when abnormal process behaviour has occurred whereas in fault isolation (FI), one distinguishes between faults. Data driven methods have attracted much attentions in the recent research works. These methods can be divided into statistical or non-statistical ones. Principal component analysis (PCA), fisher discriminate analysis (FDA), partial least squares (PLS) and statistical pattern classifiers form major classes of statistical feature extraction methods (Chiang et al, 2000). Neural network and support vector machines (SVM) (Widodo et al, 2007) are the most powerful methods of non-statistical methods. Good generalization ability is an important characteristic of SVMs. Furthermore, it is very useful for solving problems with small sample set and high dimension. Thus, it is a promising algorithm for application to fault diagnosis. A multiclass SVM has been recently proposed for FDI in (Hui et al, 2005). Nowadays, data fusion problem is widely used in different fields of study. Furthermore, its accomplishments still continue due to the promotion in other majors. The synergistic use of overlapping and complementary data sources provides such a rich database which is not available through individual sources. As a matter of fact, the overlaps of the data sources bring about robustness which is not necessarily attained through separate sources. Moreover, this technique can yield higher accuracy compared to each single sources. The application of information fusion method in multi-classifier, namely multi-classifier fusion, can fuse the output of corresponding well-behaved classifiers by proper fusion methods and reduce the limitations of each single classifier. Therefore, the fused classifier leads to more precise results by using the complementarities among different classifiers. (Ruta et al, 2000). There are generally two types of classifier combination which are known as classifier selection and classifier fusion (Woods et al, 1997). In classifier selection, each classifier makes decisions based on the information which exists in a subspace of the whole data set and then the output of different classifiers are fused together, leading to a final decision (Rastrigin et al, 1981). On the other hand, in classifier fusion, each classifier is trained over the whole feature space and the final decision is derived through an aggregation process (Ng et al, 1992). In this paper, classifier fusion, one of the appealing aspects of data fusion, is mainly discussed. The other approach considers another structure in which the output of one classifier is the input of second classifier. This approach is usually accomplished by pattern recognition algorithms (Duda et al, 1973). The major contribution of this paper is to propose some fusion structures which can be used for FDI. The first algorithm is called Serial Classifier Fusion which utilizes support vector machine classifier and PCA to identify faults in large scale processes. Parallel Classifier Fusion is the second proposed strategy which fuses SVM, Bayes and K Nearest Neighbour (KNN) classifiers. Bayes is a parametric method of classification and KNN is a nonparametric one which has been widely used for classification problems (Duda et al, 2000). The correct classification rates of these two fusion structures far overweigh separate classifiers i.e. SVM, Bayes and KNN and yield into a more proper fault diagnosis. Finally, a Parallel-Serial classifier fusion scheme is proposed which is the combination of Serial and Parallel Proceedings of the 7th IFAC Symposium on Fault Detection, Supervision and Safety of Technical Processes Barcelona, Spain, June 30 - July 3, 2009 978-3-902661-46-3/09/$20.00 © 2009 IFAC 852 10.3182/20090630-4-ES-2003.0306