Plant wide fault identification using One-class Support Vector Machines

Abstract

In this work, a new approach for doing fault identification based on One-class Support Vector Machine (1-class SVM) is proposed. Just as in the contribution plots in Principal Component Analysis (PCA), appropriate contribution metric has been developed for this method. This contribution metric is based on a sensitivity measure of the kernel distance with respect to the process variable. The main difference between the conventionally used PCA contribution factors and this method is that the latter can handle non-linear cases as well. The performance of this fault diagnosis method is compared to that of the PCA contribution algorithm. The efficacy of this method is demonstrated by applying it to the benchmark Tennessee Eastman process simulation problem and its performance is compared to that of the classical PCA.