Self-Supervised Learning via Domain Adaptive Adversarial Clustering for Cross-Domain Chiller Fault Diagnosis

The chiller is the core component of the heating, ventilation, and air conditioning (HVAC) system, such that the intelligent fault diagnosis of the chillers is of great significance for equipment safety and energy conservation. The working conditions of the chiller are complex and changeable, resulting in considerable differences in the distribution of their operating data. In addition, collecting fault data under various working conditions is also very costly. To address fault diagnosis challenges without labeled data for changing working conditions, this article proposes a domain adaptive adversarial clustering (DAAC) algorithm, which achieves fault diagnosis under dynamic distributions in evolving working conditions. Specifically, the residual model based on autoencoders is used to decouple features. Then, the feature representation can be obtained through the data encoding network and mapped to a unit sphere space. Simultaneously, a single-layer unbiased neural network classifier’s weight matrix is initialized as the prototype. In addition, the target domain features are stored in a memory buffer (MB), and a nonparametric softmax classifier is adopted to calculate the similarity entropy between the target and prototypes. On this basis, the iterative adversarial optimization is carried out to achieve dynamic clustering of prototypes and target domain features. Experimental results on the American Society of Heating, Refrigerating, and Air Conditioning Engineers (ASHRAE) RP-1043 dataset and HY-31C dataset show that the self-supervised adaptive clustering method exhibits excellent feature alignment, enabling high-precision fault diagnosis without labeled data for target conditions.