Diagnosis of Device Exception Based on Causality of Device Indicators

Existing intelligent exception diagnosis methods face challenges in unstable feature extraction and difficulty detecting anomalies in high-dimensional spaces. To address these limitations, we propose CGNN—a novel anomaly detection framework integrating causal reasoning with graph neural networks. The framework operates in three stages: 1) Causality Detection (CD) employs the LASAR algorithm to construct sparse causal graphs from monitoring variables. 2) Prediction of Device Status (PODS) leverages GraphSAGE to extract spatio-temporal features from causal graphs for state prediction. 3) Diagnosis of Exception (DOE) utilizes kernel density estimation (KDE) for distribution-agnostic anomaly scoring. Experimental validation on the extended-TE dataset demonstrates that our CGNN framework outperforms baseline algorithms, achieving faster fault identification and higher prediction accuracy. Crucially, CGNN requires minimal prior causal knowledge and adapts to arbitrary data distributions, significantly enhancing interpretability and reliability in complex industrial systems.