Causal similarity learning with multi-level predictive relation aggregation for grouped root cause diagnosis of industrial faults

Existing root cause diagnosis (RCD) methods infer causal relationships among abnormal variables by decomposing causal graphs into intra-group and inter-group levels, reducing redundancy according to direct causality. However, the indirect causality trigged by wide-range fault propagation may be ignored when inferring within groups, leading to the mismatch between causality distribution and grouping results. To overcome the challenge, we propose a causal similarity learning method with multi-level predictive relation aggregation, which contains a complementary similarity measurement framework covering both single-level and high-level causal relationships. First, an attention mechanism with temporal misalignment is designed, which can convert the undirected correlations of features into directed high-level causal similarity by extracting lagged predictive relations. Further, a graph-cutting penalty term is proposed to promote causality distribution to exhibit intra-group denseness and inter-group sparsity, so that single-level causal similarity can be considered during grouping. Finally, a dual RCD method is proposed to search root causes from the causal graph with intra-group and inter-group causality. In this way, numerous redundant causations caused by complex fault propagation can be succinctly described by inter-group causation, and the search for root cause variables can be limited to subgroups to improve diagnosis efficiency. The validity of the proposed method is illustrated through both the Tennessee Eastman benchmark example and a real industrial process.