Multiblock Independent Component Analysis Integrated with Hellinger Distance and Bayesian Inference for Non-Gaussian Plant-Wide Process Monitoring

Abstract

A novel multiblock plant-wide process monitoring method based on Hellinger distance (HD), Bayesian inference, and independent component analysis (ICA) (HDBICA) is proposed in this paper. Multiblock methods are usually employed for plant-wide process monitoring; however, block division is usually based on prior process knowledge that may not always be available. This paper proposes a totally data-driven multiblock monitoring method that employs HD to divide blocks automatically. Variables with similar probability distributions are divided into the same block on the basis of HD, and sub-ICA models are built for sub-block status monitoring. Finally, the monitoring results from all blocks are combined on the basis of Bayesian inference. HDBICA is exemplified by using a numerical study and the Tennessee–Eastman benchmark process. The monitoring results indicate that the performance of HDBICA is superior to the performances of ICA, kernel ICA, and other state-of-the-art variant-based methods.