Fault-Relevant Optimal Ensemble ICA Model for Non-Gaussian Process Monitoring

Independent component analysis (ICA) has been extensively used in non-Gaussian process monitoring. Most ICA-based monitoring methods usually establish a single monitoring model based on a certain dimension reduction criteria. However, the fault information is complex and changeable in actual industrial processes, causing the weak generalization ability of the single monitoring model. From the perspective of ensemble learning, we use the normal and available fault samples to establish several base models and ensemble them to improve the generalization ability of the monitoring model. First, the fastICA algorithm is used to extract all independent components based on the normal data. Then, a multi-objective optimization (NSGA-II) is adopted to optimize the two conditions (accuracy and diversity) of ensemble learning techniques to obtain an optimal set of base models. Finally, the monitoring results of the base models are fused through Bayesian inference. The proposed method can significantly improve the monitoring performance and the generalization ability for process monitoring. Case studies on the TE benchmark process and the wastewater treatment plant illustrate the validity and benefits of the proposed approach. In addition, we also analyzed the effects on the generalization ability of the ensemble model when the number of faults available is different.