Continual Learning Assisted Dynamic Modeling for Industrial Multimode Quality Prediction

Conventional dynamic modeling approaches for multimode processes with dynamics can only establish a local model for each mode and have difficulty increasing their prediction range due to distribution discrepancies. When a new mode emerges, the re-trained prediction model inevitably forgets important information from the historical modes, resulting in performance degradation. To address this limitation, we propose a continual learning-based dynamic modeling method called continual echo state network (CESN) for multimode processes. The method uses elastic weight consolidation (EWC) to alleviate catastrophic forgetting of the historical modes. Finally, we present the experimental results of the three-phase flow process to demonstrate the effectiveness of the proposed CESN method.