Small sample size problem of fault diagnosis for process industry

Abstract

Fisher Discriminant analysis is one of the most common used fault diagnosis methods of process industry. But it is not satisfactory in practice. In recent years, kernel methods draw much attention as excellent ability for nonlinear problem. Unfortunately, more severe small sample size (3S) problem will be brought. In this paper, regularized method is used for 3S problem of kernel Fisher Discriminant analysis. The reason why regularization can improve arithmetic stability is proved and an index to measure pattern stability is proposed. Simulation results show regularized KFDA can solve 3S problem effectively, and obtain better diagnosis effect than SVM.