Hybrid independent component analysis (H-ICA) with simultaneous analysis of high-order and second-order statistics for industrial process monitoring

Abstract

Both the Gaussian and non-Gaussian information exist simultaneously in many real industrial processes. However, the standard independent component analysis (ICA) cannot effectively explore process information when there is more than one independent component (IC) following Gaussian distribution. Besides, process may present an obvious temporal structure due to wide applications of controllers, which has not been investigated by standard ICA. To solve the above problem, the paper proposed a hybrid ICA (H-ICA) algorithm for process monitoring by concurrent analysis of both high-order and second-order statistics. The time-delayed correlation matrices are first diagonalized for robust whitening to remove the effect of the additive white noise. Then, non-Gaussian information is explored by FastICA with the utilization of higher-order statistics, and Gaussian information is extracted by analyzing the time structure information contained in second-order nonzero-delayed covariance matrices. By the hybrid analysis of high-order and second-order statistics, both non-Gaussian-distributed and Gaussian-distributed ICs are extracted. In this way, more comprehensive process information is fully investigated and analyzed for the development of process monitoring strategy. The proposed H-ICA process monitoring algorithm is verified by both a numerical example and a real thermal power plant process which illustrates its feasibility and efficacy.