Fault detection and diagnosis based on new ensemble kernel principal component analysis

Kernel principal component analysis is a technique applied for monitoring nonlinear processes. However, compute control limit based on Gaussian distribution can deteriorate its performance. Kernel density estimation is applied to solve the aforementioned issue. In conventional KPCA, a kernel based model depends on a single Gaussian kernel function selected empirically, which means a single model corresponds to a single Gaussian kernel function. It may be effective for certain kinds of fault but not for others which leads to a poor detection performance. Different Gaussian kernel functions may be needed for each kind of fault. To solve these issue, in this work, a novel ensemble kernel principal component analysis-Bayes (EKPCA-Bayes) is proposed. The ensemble learning with Bayesian inference strategy were applied into conventional KPCA. At last, the fault diagnosis performance is tested for the first time through contribution plot to find out the root cause variables. The proposed method was tested in the Tennessee Eastman (TE) benchmark process for fault detection and fault diagnosis as well.