Performance assessment of fault classifier of chemical plant based on support vector machine

Support vector machine (SVM) plays an important part in fault diagnosis of chemical plant, and intelligent optimization algorithms are used to optimize the SVM parameters, including the penalty parameter C and parameter g of different kernel function, to improve performance of its faults classification. To assess SVM faults classification capability based on diverse optimization algorithms and various kernel functions, an evaluation index system that is based upon accuracy, recall and precision was proposed, which comprehensively considers overall accuracy, false alarm probability, missing detection probability and robustness of SVM fault classifiers. Tennessee Eastman (TE) process benchmark was used as simulation platform to evaluate SVM classifying faults ability. The results showed that SVM with radical basic function (RBF) is the most sensitive to the optimization algorithm and that SVM with polynomial kernel optimized by Grid Search Method (GSM-Polynomial-SVM) provides the highest robustness. The suggested evaluation index system is conducive to selecting optimum faults classifier and could be used as a framework for future comparison.