Entropy-enhanced batch sampling and conformal learning in VGAE for physics-informed causal discovery and fault diagnosis

Abstract

Industry 4.0 has increased the demand for advanced fault detection and diagnosis (FDD) in complex industrial processes. This research introduces a novel approach to causal discovery and FDD using Variational Graph Autoencoders (VGAEs) enhanced with physics-informed constraints and conformal learning. Our method addresses limitations in conventional techniques, such as Granger causality, which struggle with high-dimensional, nonlinear systems. By integrating Graph Convolutional Networks (GCNs) and an entropy-based dynamic edge sampling method, the framework focuses on high-uncertainty regions of the causal graph. Conformal learning establishes rigorous thresholds for causal inference. Validated through simulation and case studies, including an Australian refinery and the Tennessee Eastman Process, our approach improves causal discovery accuracy, reduces spurious connections, and enhances fault classification. Integrating domain-specific physics information also led to faster convergence and reduced computational demands. This research provides an efficient, statistically robust approach for causal discovery and FDD in complex industrial systems.