Volume 60

Dual-Weight Adversarial Learning for Partial Domain Adaptation in Cross-Domain HVAC Fault Inference Yutian Lei, Cheng Fan1, Yonghang Xie, Jiena Cai

Abstract

Data-driven fault inference in building HVAC systems faces a critical challenge due to the absence of labeled operational data in the target working condition. To address this, transfer learning-based methods are proposed to leverage inference knowledge from labeled datasets acquired in distinct operational contexts, defined as the source domain. However, these methods typically assume that fault categories between source and target domains are the same. This assumption is invalid in real applications where the target domain often contains fewer fault types than the source. Such scenarios may introduce negative transfer caused by irrelevant source domain categories. To solve this problem, a novel dual-weight partial domain adaptation (DW-PDA) method is proposed, which implements selective knowledge transfer by the class-wise adversarial learning for category-aware feature alignment, and a dual-weight mechanism for dynamic sample selection. Experimental results on both HVAC air-side and water-side datasets demonstrate the effectiveness of the DW-PDA in partial domain adaptation scenarios. Compared to the global domain adaptation methods, the average accuracy improvement of the DW-PDA is 30.87% and 25.26% in AHU and chiller tasks, respectively.

Keywords partial domain adaptation, transfer learning, fault inference, HVAC, building energy system