Graph Deviation Network for Fault Detection and Diagnosis Using Building Automation System Data

Abstract: Automatic fault detection and diagnosis (FDD) is essential for energy efficiency and indoor air quality. Unsupervised FDD methods address the need for labeled data but struggle to identify root causes. This paper introduces a Graph Deviation Network (GDN)-based method for detecting and diagnosing faults in time-series data. GDN models variable relationships and enhances explainability using attention weights. Applied to FCUs in a building case study, it determines fault extent and uses rule-based diagnosis to classify faults. Results show superior anomaly detection and sensor correlation modeling, providing users with insights into the root causes of detected faults.