(32q) Causal Discovery and Hybrid Deep Learning for Process Fault Detection and Diagnosis

Fault detection and diagnosis (FDD) in complex manufacturing processes is vital for operational efficiency, quality control, and safety. However, current FDD research lacks methods that combine the interpretability needed for complex fault diagnosis with the scalability required for large-scale industrial applications, hindering the adoption of advanced techniques in modern manufacturing environments. This paper addresses this research gap by presenting a novel end-to-end FDD framework that integrates domain expertise with advanced data-driven techniques. We introduce a hybrid causal graph construction algorithm that enhances data-driven Causal Discovery Algorithms with manufacturing-specific knowledge, underpinning a knowledge-enhanced machine learning approach. Our framework achieves FDD by combining this causal knowledge with a graph-based deep learning architecture, integrating a Disentangled Graph Convolutional Gated Recurrent Network (DisenGRU) and a Bidirectional Long Short-Term Memory (BiLSTM) network for comprehensive 'relatiotemporal' analysis. Evaluated on five fluid catalytic cracking unit (FCCU) fault scenarios, the framework achieves a high average F1 score of 0.952 for unsupervised fault detection. We demonstrate a 32% reduction in forecasting error by incorporating domain knowledge into causal discovery. Crucially, this integration of knowledge enables accurate root cause diagnosis of unseen or novel faults by delineating fault propagation pathways from the causal graph. Furthermore, we show how causal knowledge helps the model differentiate between true fault cases and control loop responses to process disturbances. This advancement offers both rapid detection and insightful diagnosis, with the hybrid approach facilitating scalability in accurate and interpretable process monitoring systems.