Development of an Online Fine-Tuning Soft Sensor With Ordinary Differential Equation Network and Elastic Weight Consolidation

For data-driven modeling of industrial processes, unknown time delays, data anomalies, and concept drifts often undermine the performance of offline-trained soft sensors. To mitigate these problems, this study presents a fine-tuning-based online learning recurrent neural network (RNN) for industrial soft sensors. Initially, a robust Lipschitz ordinary differential equation RNN (LRNN) was developed for offline training of soft sensors, integrating a generalized M-estimator with a Huber loss function to handle anomalies in both the X-spaces and Y-spaces. Next, an error-triggered update mechanism was introduced to ensure timely model updates while minimizing unnecessary revisions. This mechanism employs sliding window (SW) techniques to continuously collect data and automatically initiate updates when error discrepancies exceed a predefined threshold. Furthermore, to mitigate catastrophic forgetting during continuous online learning, an elastic weight consolidation (EWC) algorithm was proposed for the robust LRNN (RLRNN). Finally, the proposed approach was validated through simulations on an artificial dataset and an industrial application to predict steam production in a coke dry quenching (CDQ) process. Comparative results with state-of-the-art algorithms highlighted the superiority of the proposed method. The online fine-tuning soft sensors effectively address model degradation due to concept drifts in industrial processes.