SVDD Based Data-Driven Fault Detection

Abstract

Conventional data driven process monitoring algorithms are limited to Gaussian process data for principal component analysis (PCA) algorithm and non-Gaussian process data for independent component analysis (ICA) algorithm. This paper provides a comparison study between the conventional data driven methods and support vector data description (SVDD) algorithm for fault detection (FD). Different from the traditional methods, SVDD algorithm has no Gausssian assumption. Thus the distribution of process data is not important for SVDD method. In order to compare their FD performances of the proposed methods from the application viewpoint, Tennessee Eastman (TE) benchmark process is utilized to compare the results of all the discussed methods. Simulation results on TE process show that ICA and SVDD methods perform better for false faults than the PCA method.