Fault detection of nonlinear systems using an improved KPCA method

Statistical control charts are essential to ensure both safety and efficient operation of many industrial processes. Many dimensionality reduction techniques such as principal component analysis (PCA) and Partial Least Squares (PLS) regression exist, and are often employed for modeling purposes as they are relatively easy to compute. However, these techniques are only effective for modeling and monitoring linear processes. The Kernel Principal Component Analysis (KPCA) method is an extension of PCA that helps deal with any nonlinearities in the process data. However, KPCA-based fault detection methods may result in a higher false alarm rate than the conventional method. In this paper, an improved KPCA method is developed in order to tackle the issue of high false alarm rates, by utilizing a mean filter to smoothen the detection statistics that are obtained from the KPCA method. The advantages presented by the developed method are illustrated using a simulated nonlinear model. The results clearly show that the improved KPCA method provides improved fault detection results with low missed detection and false alarm rates, and smaller ARL1 values compared to the conventional methods.