Attention-empowered transfer learning method for HVAC sensor fault diagnosis in dynamic building environments

Abstract

The Heating, Ventilation and Air Conditioning (HVAC) system is a key system in buildings for providing energy-saving and occupant-centered indoor services. Sensor fault diagnosis (SFD) is essentially important for HVAC systems since incorrect sensory measurements may destabilize system operations. Most SFD methods for HVAC systems require comprehensive manual knowledge or massive labeled data which is hardly available in diverse buildings. To reduce labor cost, transfer learning method is adopted in diagnosing HVAC sensor faults, however, it still encounters critical challenges, such as imbalanced data distribution and insignificant fault features. In this paper, an attention-empowered transfer learning method is innovatively proposed to enhance the capability of SFD in HVAC systems. Firstly, we introduce the multi attention-based module to characterize sensor faults in spatio-temporal domain. Secondly, we explore the relationship between source domain and target domain, and establish the model that enables useful knowledge mapping. Thirdly, to alleviate the effect of data discrepancy existed in new domains, a joint domain loss function is designed to enhance the domain adaptive ability. Experimental validation is carried out on three datasets from completely different real-world scenarios. Compared with existing methods (i.e., TCA, JAN and DDC), the proposed method achieves the highest average accuracy of 93 % in cross-domain building environments. The ablation studies also demonstrate the effectiveness of our method under diverse parameter settings, such as attention modules, fault types and data volume.