A fault detection method using multi-scale kernel principal component analysis

A fault detection method based on multi-scale kernel principal component analysis (MSKPCA) is developed for nonlinear processes monitoring. It integrates wavelet analysis and nonlinear transformation using kernel principal component analysis. Wavelet analysis can decompose measured signal into approximation part and detail part at multiple scales to capture time-frequency information, while kernel principal component analysis is performed for nonlinear principal components at each scale by kernel functions. The combined method can simultaneously extract cross correlation, auto correlation, and nonlinearities from the data. Furthermore, a multi-scale principal component analysis similarity factor is proposed for identifying fault pattern. Simulation of a TE benchmark process shows that the proposed method has a better performance compared with the traditional PCA method in fault detection and diagnosis.