A physics-informed dynamic graph and evidence fusion method for fault alarm detection in building HVAC systems

Abstract

To address the superposition of normal thermal disturbances and fault characteristics in building heating, ventilation, and air conditioning (HVAC) systems, as well as conflicts among different diagnostic evidence sources, a physics-constrained dynamic graph and multi-source evidence fusion method for fault alarm detection is developed. The method employs the Dynamic Coupling-Informed Graph Attention Network to characterize the time-varying directional coupling relationships among air-side temperature variables, equipment control variables, and operational states. The graph structure is subject to hard constraints derived from known HVAC airflow paths and control logic, while dynamic edge weights are estimated from rolling lagged relationships in normal operational data. The Physics-Informed Neural Network incorporates discrete thermal equilibrium residuals into the training objective and separates normal thermal disturbances from abnormal disturbances using normal-state supervision, smoothness constraints, and sparsity constraints. The Dual-Stream Temporal-Frequency Attention Encoder extracts short-term abrupt-change and slow-varying drift features, respectively. The Adaptive Conflict Redistribution based on Evidential Reasoning reallocates conflicting evidence according to historical consistency, physical consistency, and confidence stability. Experiments are conducted using Fault Detection and Diagnostics data, with the training, validation, and test sets partitioned by complete fault events. The results show that the proposed method achieves an overall alarm accuracy of 98.7%, with a false alarm rate of 1.2% and a missed alarm rate of 1.1%. Stable identification performance is maintained across different fault categories and under perturbation- and noise-corrupted conditions. These results indicate that the proposed method can improve the reliability of fault identification and alarm fusion under complex HVAC operational conditions.