Dalarna University's logo and link to the university's website

Accurate fault diagnosis in chillers is essential for maintaining optimal energy efficiency and operational reliability in building Heating, Ventilation, and Air Conditioning (HVAC) systems. However, chillers often operate under diverse conditions, which can cause data-driven models trained on specific operational conditions to suffer significant performance deterioration when deployed in new conditions. Transfer learning offers a promising solution by leveraging knowledge from source domains, but its "black-box" nature raises concerns about model interpretability, hindering practical application. To address this challenge, this study developed an evaluation framework integrating Layer-wise Relevance Propagation (LRP), Integrated Gradients (IG), and SHapley Additive exPlanations (SHAP) to assess feature-level interpretability in transfer learning-based FDD model. We modified the 1D CNN architecture specifically for compatibility with interpretation methods. The model achieves an overall accuracy of 97.01 %, with component-level faults showing higher accuracy than system-level faults. All interpretation methods consistently identify physically meaningful discriminative features, and the robustness of the methods is validated through 10 randomized trials. Meanwhile, data volume critically impacts the clarity of interpretation results—larger datasets yield higher feature importance scores, though even 1 % of training data is capable of identifying discriminative features. To simulate real-world chiller operation, multiple cross-condition transfer learning tasks were designed, covering a wide range of operating scenarios. Results demonstrate that the Domain-Adversarial Neural Network (DANN) and Fine-Tuning (FT) improve target-domain accuracy by 25 % over baseline models while preserving physically meaningful discriminative features from the source domain. © 2025 Elsevier B.V., All rights reserved.