BiGraphormer: A Bidirectional Graphormer on Directed Causal Graph for Fault Detection in Complex Systems

Effective fault detection of complex systems can significantly enhance safety, availability, and maintainability. Recently, graph neural networks (GNNs), which leverage spatial structures among variables, have gained substantial attention due to advancements over previous data-driven methods. However, existing GNN-based fault detection models primarily focus on adjacent neighborhood for feature fusion, neglecting the long-range dependencies in graphs. Moreover, these models often utilize undirected correlational graphs, potentially limiting their applicability and modeling efficiency for target systems. In this work, a causal graph-based bidirectional Graphormer (BiGraphormer) is proposed for complex systems fault detection. The causal relationships among monitoring variables are mined and represented as a directed acyclic causal graph, in which nodes denote variables and directed edges indicate causal influence. Then, the dependencies including the global spatial structure, the node and edge information are encoded and fused using the proposed BiGraphormer. By incorporating both ancestral cause and descendant effect nodes along the directed causal graphs, comprehensive representations are constructed for system fault detection. To validate the effectiveness of the proposed framework, a case study concerning real monitoring data of high-speed train braking systems is conducted, with the results showing efficacy of the proposed method.