Deep learning-powered heating, ventilation, and air conditioning compressor fault diagnosis facing unseen domains and class imbalances

Abstract

Reliable fault diagnosis of compressors in heating, ventilation, and air conditioning (HVAC) systems is essential for enhancing their service reliability and energy conservation. However, heterogeneous working environments of HVAC compressors pose significant challenges for applying data-driven fault diagnosis methods. Domain generalization techniques have been developed to address data distribution discrepancies in cross-domain fault diagnosis. Yet, most existing methods assume that source domains have equal sizes and balanced class distributions. These assumptions limit their applicability to real-world scenarios that can encounter multiple levels of imbalance in both domain size and health status. Therefore, this work proposes a novel Adaptive Invariant Representation learning-based domain generalization Network (AIRNet) to enable a better HVAC compressor fault diagnosis performance in handling unseen domains and class imbalances. Specifically, AIRNet employs a probabilistic sampling strategy to adaptively extract balanced training samples from source domains, mitigating class imbalances and driving unbiased model learning. Furthermore, AIRNet integrates fault classification, metric learning, and domain adversarial training modules with a tailored data augmentation strategy, jointly enhancing its robustness and generalizability across unseen domains. These components collaborate to establish fault-discriminative and domain-invariant diagnostic boundaries while improving model resistance against unseen data distribution discrepancies. Extensive computational experiments on HVAC compressors demonstrate the superiority of AIRNet over state-of-the-art methods in addressing real-world industrial fault diagnosis challenges. Compared to the best-performing benchmark, AIRNet achieves an average performance gain of 1.11 % in total accuracy and 2.76 % in macro F1 score across all studied tasks. The code is available at https://github.com/ifuturekk/AIRNet.