Causal Graph Attention Network with Disentangled Representations for Complex Systems Fault Detection

Considering the importance of complex systems fault detection, much efforts have been dedicated to fault feature extraction with monitoring data. The graph-based approach has become a trending topic, which exploits the non-Euclidean structure and generates representation based on spatial information. However, most graph-based models are built based on correlation assumption, and disregards the causality which are intrinsic in system and its failure process. In this paper, a causal graph attention network with disentangled representations (Causal -GAT) is proposed for fault detection. High-dimensional variables are first characterized into directed acyclic graphs using data-driven causal discovery combining expertise. The causal graph, which represents variables' cause-effect relations, is fed into the Causal-GAT. Disentangled Causal Attention (DC-Attention) is proposed to adaptively aggregate cause variables for embedding the effect variables. To improve feature extraction efficiency in the multi-head attention, the DC-Attention enforces disentangled node representation by regularizing it with a specified causal condition. To verify the effectiveness of the proposed method, a real case study concerning the high-speed train braking systems is considered. Experimental results with the benchmark methods demonstrate the advantages of the proposed method. Validities of causal graph construction, representation disentanglement, as well as interpretability of the model are also discussed in this work.

Automated summary

This paper proposes a causal graph attention network with disentangled representations (Causal-GAT) for fault detection. High-dimensional variables are transformed into directed acyclic graphs using data-driven causal discovery, and the causal graph is fed into Causal-GAT. The introduced Disentangled Causal Attention (DC-Attention) adaptively aggregates cause variables to embed effect variables. Experimental results demonstrate the advantages of the proposed method, and the validity of causal graph construction, representation disentanglement, and interpretability of the model are discussed.