Adaptive Selective Ensemble-Independent Component Analysis Models for Process Monitoring

Abstract

Independent component analysis (ICA) has been widely used in non-Gaussian industrial process monitoring. However, the stability of performance and determination of dominant ICs are still the main problems for ICA. Constructing a monitoring model to achieve the best performance for different faults is a great challenge owing to the diversity and unknowability of faults. This study develops an adaptive selective ensemble ICA models method to improve the monitoring performance. Ensemble learning based on the bagging algorithm is adopted to enhance the stability of ICA. According to the difference of ICs selected for ICA modeling and hierarchical clustering, corresponding model sets are constructed for each sample subset. To ensure the accuracy of each submodel, an adaptive method based on just-in-time learning is proposed to model selection. Bayesian inference is applied to determine the final monitoring index. The validity of the proposed approach is attested through a numerical example, TE benchmark process, and wastewater treatment plants.