Nonlinear process fault diagnosis using kernel ICA and improved FDA

Abstract

Fisher discriminant analysis (FDA) is a linear technique which is non-optimal on minimizing the overall misclassification rate. In this paper, a new approach called KICA-IFDA is proposed for nonlinear process fault diagnosis. In the KICA-IFDA, the kernel independent component analysis (KICA) is first adopted to extract fault feature data from original fault data. Then an improved FDA (IFDA) criterion which is optimal on classifying the fault feature data is constructed based on the particle swarm optimization. The simulation studies on the Tennessee Eastman process demonstrate that the proposed KICA-IFDA outperforms both the IFDA and FDA.