Improved Kernel Fisher Discriminant Analysis for Nonlinear Process Fault Pattern Recognition

Abstract

Kernel Fisher discriminant analysis (KFDA) has emerged as an well-known nonlinear fault pattern recognition method. However, traditional KFDA method does not consider the utilization of the high order statistical information of process variables, and ignores the mining of the local data structure characteristic. To achieve better fault pattern recognition performance, this paper proposes an improved KFDA method, called statistics local KFDA(SLKFDA). In the proposed method, two techniques, including statistics pattern analysis (SPA) and local structure analysis (LSA), are combined to enhance the basic KFDA method. Firstly, SPA is applied to extract the original process variables’ statistics with different orders. Then the KFDA optimization objective is modified by considering the local structure preserving. Lastly, a fault classifier is developed to recognize fault pattern. Simulations on one benchmark process demonstrate that the proposed SLKFDA method has a superior fault pattern recognition performance.