Enhancing Root Cause Diagnosis for Industrial Process Faults: Asymmetric Sparsemax-Driven Predictive Modeling

In industrial processes, the accurate pinpointing of root causes is pivotal to ensure timely and precise interventions, ultimately culminating in the mitigation of underlying disruptions. While root cause diagnosis (RCD) methods endeavor to transcend the limitations of traditional fault diagnosis approaches by tracing origins, they often grapple with the labyrinthine causal interdependencies interspersed among multivariate time series. The presence of redundant, bidirectional, or even cyclic relationships further convolutes the diagnostic landscape. To address these challenges, we present an innovative RCD methodology designed specifically for intricate industrial processes. Our unique RCD framework integrates node state surveillance, grounded in a pre-fault causality graph, circumventing the customary delays associated with amassing sufficient post-fault samples in extant RCD techniques. Importantly, our method adeptly handles cyclic causality graphs, a recurrent challenge in the feedback controls of complex systems. For pre-fault causality construction, we propose asymmetric sparsemax-driven predictive modeling (ASPM), which refines the sparsity and asymmetry of causality graphs, bolstering the reliability and clarity of RCD outcomes. Central to our method is a pioneering graph structure learning module, skilled in removing unnecessary connections while preserving vital predictive interrelationships in multivariate time series. Empirical validation with the continuous stirred-tank reactor (CSTR) process and blast furnace (BF) ironmaking process confirms the enhanced efficacy of our proposed RCD approach in industrial settings.