Robust and Interpretable Deep Learning Fault Diagnosis in Complex Chemical Processes: Performance Enhancement via Dempster–Shafer Theory and Feature Attention

Monitoring and diagnosing faults in chemical processes is essential for ensuring operational safety and product quality, particularly as industrial systems become increasingly complex. Deep learning has emerged as an effective data-driven approach for handling high-dimensional and nonlinear process data. Although effective at feature extraction, deep learning models face challenges related to sensor uncertainty, conflicting diagnostic signals, and limited interpretability. In response to these limitations, this study presents a novel fault diagnosis framework that integrates autoencoder (AE) and long short-term memory (LSTM) models with Dempster-Shafer theory (DST) and feature attention mechanisms. The framework combines deep learning's representation learning capability with DST's ability to manage uncertainty, while feature attention enhances interpretability. Validation was conducted using the Tennessee Eastman process (TEP) dataset, during which four key hyperparameters were carefully tuned. The standalone AE and LSTM models attained accuracies of 85 and 70%, respectively, which improved to 91 and 75% after DST integration. The AE-DST approach achieved a precision of 92%, outperforming conventional techniques. In addition, the attention-enhanced AE model achieved strong performance (accuracy: 90%, precision: 91%, F1-score: 90%), while feature attention identified critical process variables and improved diagnostic transparency.