Improved ICA for process monitoring based on ensemble learning and Bayesian inference

Abstract

Independent component analysis (ICA) has been widely applied to the monitoring of non-Gaussian processes. Despite a number of applications, the selection logic for the dominant independent components (ICs) for constructing the latent subspace has always been empirical. This article presents a new ICA-based monitoring scheme that integrates an ensemble learning strategy with Bayesian inference. Given that there is no direct evidence to conclusively show which dimensionality reduction criterion in ICA is optimal for the purposes of process monitoring, it seems logical to combine different criteria within an ensemble solution. The Bayesian inference can then be used as a local fusion strategy to obtain the final decision. It is shown that the proposed ICA methodology can provide improved performance regardless of how the dominant ICs are determined. The monitoring performance is illustrated through case studies on a numerical example and the Tennessee Eastman benchmark process.