EICD: Causal Structure Recovery of Bearing Failure Data Containing Latent Confounding

Deep learning is widely used in the field of intelligent fault diagnosis. However, data-driven deep learning models are with unsatisfying interpretability and their performance cannot be maintained when a sample distribution shift occurs. Causal discovery can enhance the interpretability of deep learning models by establishing causal graphs to uncover causal invariance in the data. Nevertheless, causal discovery suffers from the curse of dimensionality due to the large conditional sets used in conditional independence (CI) tests, and the Markov equivalence class (MEC) obtained contains many undirected edges due to some causal structures having the same CI. In response, we incorporate the causal direction criterion (CDC) into iterative causal discovery (ICD) to construct the extended ICD (EICD) algorithm. EICD can obtain a detailed causal graph with lower complexity, enhancing the interpretability of the diagnostic framework. First, EICD reduces the number of required CI tests by limiting the size of the conditioning set and its distance from the target node. Second, based on CDC, EICD uses causal asymmetry to infer the direction of causal edges between pairs of nodes, thus accelerating ICD and adding more directional causal edges compared to MEC. In the experimental section, we compare the number of CI tests and the improvement in directional accuracy with other algorithms and use the EICD algorithm to reconstruct the causal graph of the faulty system. The experiments demonstrate that the EICD algorithm can reduce the required number of CI tests and recover the underlying causal structure more accurately.