Autoencoder-Based Dempster-Shafer Decision-Making Framework for Intelligent Process Monitoring: Applications to Nonlinear Chemical Processes

Abstract

Industrial chemical processes are nonlinear, large-scale, and strongly coupled, making them susceptible to disturbances that threaten operational safety. Conventional model-based and deep learning approaches often struggle with nonlinear dynamics, limited labeled fault data, detection latency, and diagnostic uncertainty. This study presents an intelligent process monitoring framework integrating autoencoder (AE)-based feature learning using denoising autoencoder (DAE), sparse autoencoder (SAE), and variational autoencoder (VAE) models with Dempster–Shafer theory (DST) for uncertainty-aware decision fusion. The framework was evaluated on a benchmark continuous stirred-tank reactor data set containing 12 fault scenarios and normal operating conditions. Baseline accuracies of 87% (DAE), 79% (SAE), and 74% (VAE) improved to 93%, 92%, and 90%, respectively, after DST integration. Following this integration, DAE-DST reduced the false alarm rate (FAR) to 1.41%, increased the G-mean to 0.911, and shortened the mean detection delay (MDD) to 0.56 min. The proposed framework provides a robust and computationally efficient solution for reliable real-time process monitoring in nonlinear chemical processes.