Dual‐causal analysis enhanced stacked auto‐encoders for closed‐loop industrial process monitoring

Abstract Recently, the imperative to ensure operational safety and optimize production efficiency has inspired significant advancements in industrial process monitoring. However, due to the feedback compensation mechanism of the closed‐loop control system, the effects of incipient abnormalities are often obscured by the presence of the normal control adjustments. To tackle this challenge, we propose a dual‐causal analysis enhanced stacked autoencoders (DCSAE) network, specifically designed for closed‐loop industrial processes monitoring in dynamic. Initially, a causal self‐attention mechanism is integrated into the training process of the detection model to effectively distinguish abnormalities from normal control adjustments. Subsequently, the causality‐weighted stacked autoencoder model serves as the detection model, identifying abnormal operating conditions based on variations in reconstruction errors. To preserve the necessary causal information in the feature extraction step, conditional Granger causality analysis is incorporated into the encoder to establish the causal relationship between the operational data variables and the anomaly detection indicators. Furthermore, an adaptive threshold generator based on temporal regressor is developed to improve the accuracy of fault detection in dynamic industrial processes. Finally, the effectiveness and superiority of the proposed method are thoroughly validated through case studies involving both the vinyl acetate monomer process and a closed‐loop continuous stirred tank reactor benchmark.