Performance-driven ensemble ICA chemical process monitoring based on fault-relevant models

Abstract

Independent component analysis has been widely used in non-Gaussian chemical process monitoring. However, using the normal operation data or adopting a certain independent component selection criteria alone to construct the monitoring model cannot achieve satisfactory monitoring performance. To address the problem effectively, a fault-relevant model selection combined with ensemble learning and the Bayesian inference method is developed in this study. First, numerous models are constructed on the basis of the randomly selected ICs. Second, the models with the highest fault detection rates for each fault are selected. Then, a screening algorithm based on the difference in detection results is adopted to reduce the redundancy of the selected models and thus improve the monitoring performance. Finally, Bayesian inference is adopted to combine the testing results of the retained models. Case studies containing numerical simulations and the Tennessee Eastman benchmark process illustrate the validity of the proposed approach.