A chiller is a critical component of heating, ventilation, and air conditioning (HVAC) systems. Faults in chillers can lead to energy waste and even safety incidents. Therefore, fault diagnosis for chillers is essential for HVAC systems. Data-driven fault diagnosis methods rely on large amounts of historical data, but labeled fault data is often difficult to collect, resulting in reduced diagnostic accuracy of models. To address this issue, this paper proposes a fault diagnosis method based on a consistency loss generative adversarial network (CLGAN). First, CLGAN is trained with a small number of labeled samples and a large amount of unlabeled data to generate realistic fault samples. Next, a balanced dataset containing multiple fault categories is constructed by combining both generated and historical data. Finally, a fault classifier is trained on this balanced dataset to perform real-time fault diagnosis. By introducing a consistency loss function into the discriminator, CLGAN effectively leverages unlabeled data, increasing data utilization. Meanwhile, the generator is guided at multiple scales to meet the discriminator′s requirements, enabling the model to produce high-quality samples even under various disturbances and thus enhancing diagnostic accuracy and robustness. Experimental results on the ASHRAE and HY-31C datasets demonstrate that, with only five labeled samples per class, CLGAN achieves fault diagnosis accuracies of 92. 8% and 95. 9% , respectively, illustrating its excellent performance. Moreover, in noise and cross-condition experiments, CLGAN shows superior robustness and generalization compared with other methods.