Exploring Causal Graph Using Graphormer for Complex Systems Fault Detection

Data-driven fault detection has gained substantial attention in recent times. Graph-based models, which incorporate spatial information for feature extraction, have shown promising results, particularly in the context of increasing system complexity. However, existing models tend to focus solely on adjacent neighbor features, overlooking long-range dependencies. Moreover, many methods assume predefined correlation relationships between variables, potentially limiting their effectiveness. This paper introduces a novel approach: the Causal-Graph-Based Graphormer model for complex system fault detection. It utilizes causal discovery techniques to mine monitoring data, representing causality in graphs. These causal dependencies are then effectively modelled using the Graphormer framework, which encodes both local and long-range features of nodes and edges. To demonstrate its efficacy, a case study involving real high-speed train-braking systems validates the proposed model's effectiveness.