A Unified Framework With Collaborative Reasoning Capacity for Nonuniform Industrial Processes Root Cause Identification

Abstract

Capturing the root cause is crucial for ensuring the safety and efficiency of industrial processes. Nevertheless, most of the existing methods are unavailable for resultful causality mining, particularly those exhibiting nonuniform characteristics. Besides, insufficient utilization of process knowledge and data leads to suboptimal model performance. To fill this gap, a unified root cause identification framework with collaborative reasoning capacity is proposed for nonuniform processes. Initially, the nonuniform dynamic time warping (NDTW) and maximal relevance minimal redundancy (mRMR) methods are designed to acquire highly uniform data with less computational burden. Subsequently, the semantic-aided path search algorithm is proposed for gaining the refined knowledge causal graph (RKCG), enhancing the interpretability and providing a preliminary reference for subsequent causality identification. After that, integrating the highly uniform data and RKCG was into the advantageous attention-causal-aware gated recurrent unit (ACGRU) model for a more reliable causality matrix. Then, the comprehensive causal graph is established with the aid of knowledge-data collaborative reasoning. Finally, the proposed method is verified by ASHRAE RP-1043 centrifugal chiller and real-world coal mill fault datasets.