A study on source domain selection for cross-domain fault diagnosis in HVAC systems

Cross-domain fault diagnosis methods facilitate fault diagnosis modeling for heating, ventilation, and air conditioning (HVAC) systems in scarce-labeled-data scenarios by transferring diagnostic knowledge from systems with abundant labeled data (source domain) to inform model development for systems with scarce labeled data (target domain). However, the performance of this approach depends on the similarity between the source and target domains—higher similarity typically yields greater diagnostic accuracy, making source domain selection crucial. Despite its importance, systematic strategies for source domain selection remain underexplored. To address this issue, this study investigates source domain selection for HVAC cross-domain fault diagnosis. Specifically, diverse cross-domain fault diagnosis scenarios are first created. The similarity between source-target domain pairs and the accuracy of the cross-domain fault diagnosis model in each scenario are then quantified. Ultimately, this study extracts quantitative correlations between source-target domain similarity and diagnostic accuracy. Based on these correlations, the source domain selection strategy is derived, specifying the requisite similarity thresholds—in terms of temperature, flow rate, and power consumption variables—to achieve distinct tiers of diagnostic accuracy. The proposed method is validated using an experimental HVAC system, yielding an actionable source domain selection strategy for three diagnostic accuracy levels. Multiple real-world cross-domain fault diagnosis scenarios further confirm the reliability of the derived strategy. The findings provide a methodical guide for selecting appropriate source domains, ensuring that the resulting cross-domain fault diagnosis models achieve satisfactory accuracy.