Heterogeneous Sensor Fault Detection for Networked Systems Based on a Graph Transformer

Networked systems, such as fuel, petrochemical, and water supply systems, use various sensors for monitoring and control. Due to harsh environmental conditions, sensors are prone to failure and pose a significant threat to the stability of the system, making the accurate identification of such anomalies crucial. In recent years, data-driven sensor fault detection methods utilized sensor data correlations, spatial and temporal features, but they have not fully harnessed the physical structure relation of sensors in real system. Our study introduces a novel sensor fault detection method that integrates graph neural networks (GNNs) and transformers. The model efficiently uses the interdependence among various sensors for information transmission and aggregation, relying on the physical structure of networked system. To enhance the information exchange among identical sensor, the model divides the graph neural network into subgraphs based on sensor types. Within each subgraph, the attention mechanism of transformers is employed to capture sensor features and distributions, thereby improving the learning of the sensor node representation vector. Finally, a decoder with reconstructed input is utilized for sensor fault detection. The experiments demonstrate better fault detection accuracy in both real networked systems and four heterogeneous sensor datasets. Across all experimental cases, the
${F}1$
score of our method is 95.79%, and the AUC value is 98.83%. Finally, through ablation study, the effectiveness of each component of the model is confirmed.