A modified transformer and adapter-based transfer learning for fault detection and diagnosis in HVAC systems

Abstract

Fault detection and diagnosis (FDD) of heating, ventilation, and air conditioning (HVAC) systems can help to improve the energy saving in building energy systems. However, most data-driven trained FDD models have limited generalizability and can only be applied to specific systems. The diversity of HVAC systems and the high cost of data acquisition present challenges for the practical application of FDD. Transfer learning technology can be employed to mitigate this problem by training a model on systems with sufficient data and then transfer it to other systems with limited data. In this study, a novel transfer learning approach for HVAC FDD is proposed. First, the transformer model is modified to incorporate one encoder and two decoders connected, enabling two outputs. This modified transformer model accommodates absent features in the target domain and serves as a robust foundation for transfer learning. It has effective performance in complex systems and achieves an accuracy of 91.38% for a system with 16 faults and multiple fault severity levels. Second, the adapter-based parameter-efficient transfer learning method, facilitating the transfer of trained models simply by inserting small adapter modules, is investigated as the transfer learning strategy. Results demonstrate that this adapter-based transfer learning approach achieves satisfactory performance similar to full fine-tuning with fewer trainable parameters. It works well with limited data amount in target domain. Furthermore, the findings highlight the significance of adapters positioned near the bottom and top layers, emphasizing their critical role in facilitating successful transfer learning.