Fault diagnosis for cross-building energy systems based on transfer learning and model interpretation

Abstract

Timely and accurate fault diagnosis in building energy systems is crucial for improving the energy efficiency of equipment and realizing energy savings. With the widespread use of building automation systems, data-driven deep-learning fault diagnosis methods have gained popularity due to their flexibility and accuracy. However, they have certain associated challenges, such as limited extrapolation capabilities and their black-box nature. To address these issues, this study focuses on transfer learning to improve the energy efficiency of building energy system equipment, exploring the influencing factors and behaviours of transfer learning fault diagnosis performance and model interpretation results. A transfer-learning strategy is employed, utilizing feature-level SHAP and visual interpretation methods to understand the inner workings and diagnostic mechanisms of transfer-learning models. The study constructs five transfer-learning strategies for two typical building energy systems. The results demonstrate that the proposed fine-tuning approach achieves the highest accuracy, of 94.32%, surpassing non-transferred and shallow approaches. The interpreted methods reflect the joint diagnosis of refrigerant leakage faults by compressor discharge temperature and condenser inlet and outlet water temperature in the target chiller. The findings from both the transfer mechanism and interpretation of the deep-learning model provide valuable guidance for the development of practical deep transfer-learning models and corresponding interpretation methods. This research contributes to improving fault diagnosis accuracy and understanding the diagnostic mechanisms in building energy systems, ultimately leading to enhanced energy efficiency and energy savings.