A dual-parameter optimization KPCA method for process fault diagnosis

Kernel principal-component analysis (KPCA)-based methods have recently shown to be very effective for process fault diagnosis, and significant research has been done in recent years. However, the known fault datasets are not used while modeling and analyzing with KPCA, which results in low efficient fault detection. Another deficiency of traditional KPCA is lack of suitable selection criteria for choice of kernel parameters and the number of principal components, and some of fault detection rate can not meet the requirements or even some kind of fault can not be detected. In this paper, a kernel principal-component analysis (KPCA) method based on Dualparameter optimization is developed. Firstly, the effect of kernel parameters and the number of principal components to the performance of fault detection are analyzed based on traditional KPCA. Secondly, a Dual-parameter optimization for KPCA is introduced. For each type of fault, the effect of kernel parameters on the extraction of nonlinear features and effect of number of principal components on dimensionality reduction are both taken into account. A case study of a Tennessee Eastman (TE) system is provided to illustrate our work. The study results show that Dual-parameter optimization KPCA method can greatly improve the performance of fault detection over the popular KPCA-based methods.