Toward Multimode Process Monitoring: A Scheme Based on Kernel Entropy Component Analysis

Modern industries are facing diverse market demands and exhibit typical multimode characteristics. In this context, not only common process characteristics such as nonlinearity, non-Gaussianity, and multiscale become more prominent and complex but also faults, especially small ones, are easily masked by normal process behaviors such as transition operations, which pose great challenges to traditional process monitoring methods. To address these challenges, this article makes full use of the capability of kernel entropy component analysis (KECA) in capturing cluster structure and extracting data features and proposes a monitoring scheme applicable to multimode processes. The scheme is formulated under a multimodel framework with Bayesian fusion. First, an improved fuzzy $C$ -means (IFCM) clustering method based on KECA is presented for mode division and identification. Then, multiscale KECA (MSKECA) is developed for multivariate statistical modeling, and angle variance index (AVI) is designed for global-scale monitoring. Finally, a Bayesian inference probability (BIP) index that integrates the results of each unsupervised MSKECA-AVI monitoring model is constructed to determine the process status. The case study on the Tennessee Eastman process (TEP) with operational variability shows that the proposed scheme is effective, especially for small-magnitude faults and transition operations, and performs better overall than some state-of-the-art methods.