Improved Equipment Fault Prediction for Smart Building IoT Systems Based on Topology-Aware GNN

Abstract

HVAC equipment, energy meters, controllers, gateways, and environmental sensors are among the many connected devices in smart building IoT systems. Due to the high dependency of device states between physical spaces, control loops, and communication links, failures are rarely isolated. Topology-driven fault propagation cannot effectively consider independent series fault prediction models. This paper proposes a topology-aware GNN technique to enhance the capability of equipment fault prediction in smart building IoT systems. To predict device-level failure probabilities, create heterogeneous device graphs, encode sensor state sequences, and perform topology-aware graph aggregation. The tests include normal operation, sensor degradation, HVAC failures, controller failures, and communication errors. The accuracy and other metrics (macro F1) are 96.4% and 95.2% respectively, with an AUC of 0.973, and the average early warning time is 38.5 minutes. Therefore, through topological perception graph learning, the performance of predictive maintenance for multiple devices in smart buildings can be improved.