Proactive Anomaly Mitigation in Industrial Processes using a Deep Learning-Driven Predictive Controller with Integrated Sensor Fusion and Temporal Forecasting

Abstract

Contemporary manufacturing environments are subject to multifaceted process disturbances that culminate in quality deterioration, unscheduled downtime, and elevated operational costs. Reactive control architectures, which respond to anomalies after their manifestation, are fundamentally inadequate for processes demanding stringent quality assurance. This paper introduces the Deep Learning-based Anomaly-driven Predictive Controller (DL-APC), a novel framework that synergistically integrates convolutional-autoencoder-based anomaly detection, bidirectional long short-term memory (Bi-LSTM) temporal forecasting, and deep Q-network (DQN) reinforcement learning to enable proactive process parameter adjustment. By continuously fusing heterogeneous sensor streams, constructing a quantitative risk score from predicted anomaly trajectories, and translating this score into real-time actuator commands, the DL-APC preemptively eliminates root-cause deviations before they propagate to product quality metrics. Extensive experiments on a simulated semiconductor chemical mechanical planarization (CMP) plant demonstrate that the proposed system achieves an anomaly detection F1-score of 0.981, a prediction RMSE of 0.024, and a defect rate reduction of 34.7% compared to the next-best baseline. These results confirm that the DL-APC substantially outperforms five state-of-the-art algorithms across all evaluation dimensions.