Improved neural component analysis for monitoring nonlinear and Non-Gaussian processes

Abstract

Neural component analysis (NCA) is one of the latest nonlinear multivariate statistical process control (MSPC) methods, which consists in combining artificial neural networks (ANNs) with principal component analysis (PCA). However, NCA cannot handle the non-Gaussian feature and the extracted principal components (PCs) in NCA may not be the key information in the process data. Herein, we propose an improved NCA (INCA) which introduces a new cost function based on kurtosis to restrict the Gaussianity of PCs. We also propose a novel PC selection mechanism based on the information of PCs in the original data space rather than in the PC data space. INCA achieves almost 100% detection rates in three different types of faults in a simulation model test, and it can detect the fault in the thermal power plant process more than 1 min ahead of orthogonal nonlinear PCA (O-NLPCA) and NCA.