Causal analysis enhanced stacked denoising autoencoder for closed-loop industrial process monitoring

As a crucial tool to ensure the safety of industrial processes, process monitoring technology has recently attracted more and more attention. Due to the complexity between industrial process variables and the presence of diverse noise in the operating environment, improving fault detection has become a hot topic. To tackle this challenge, we proposed a detection strategy using a causal analysis-enhanced stacked denoising autoencoder (CSDAE) with dynamic threshold for the monitoring of industrial closed-loop control systems. Initially, a causality-weighted stacked denoising autoencoder model serves as the detection model. In particular, the causal relationships between input variables and reconstruction errors are explicitly incorporated to dynamically adjust the training weights and enhance detection performance. Subsequently, a dynamic threshold generator based on a regressor is developed to improve the accuracy of fault detection in dynamic industrial processes. Finally, the proposed method is validated through a case study on a closed-loop continuous stirred tank reactor benchmark.