Graph-based unsupervised domain adaptation for fault diagnosis of HVAC systems

While data-driven methods have shown strong potential for fault diagnosis in heating, ventilation, and air conditioning (HVAC) systems, their effectiveness is often limited by the scarcity of well-labeled data in real buildings. Unsupervised domain adaptation (UDA) mitigates this challenge by transferring knowledge from labeled source domains to diagnose faults in unlabeled target domains. Most existing UDA approaches for HVAC fault diagnosis employ convolutional neural networks (CNNs), which overlook the underlying system topology. Graph neural networks (GNNs) offer a powerful alternative for capturing complex interactions among system components. However, their potential for UDA remains unexplored due to the absence of predefined graph structures for HVAC systems. To address this gap, this study proposes a GNN-based UDA framework that transforms HVAC tabular data into fault-specific causal graph representations. The proposed method achieves over 74.5% accuracy across multiple scenarios, demonstrating its effectiveness and superior performance compared to benchmark models.