Class-imbalanced domain generalization fault diagnosis for chiller based on simulation data

Domain generalization for chiller fault diagnosis requires balanced data, but automatic shutdowns cause fault data scarcity and class imbalance. We propose a simulation-based method that integrates simulated and real data to address this issue. First, a chiller simulation model is designed, and its heat transfer parameters are fine-tuned using operational data to reduce the discrepancy between the simulation model’s response and the actual unit’s response. Then, a mapping model is developed, employing an encoder-decoder network to further improve the similarity between simulated and real data. These two refinement steps ensure that the simulation model’s response closely approximates the behavior of the real system. Finally, a domain-invariant feature extraction model for chiller is developed. This model uses domain adversarial training to extract domain-invariant representations and incorporates a local domain discriminator to perform class-level domain differentiation, enhancing the feature extraction capability. The proposed method is validated on a laboratory chiller unit system. Comparative analysis with other domain generalization methods demonstrates that the proposed approach achieves higher diagnostic accuracy, with an average diagnostic rate reaching 88.97% across three tasks, while offering enhanced data interpretability.