Fault diagnosis method for data imbalance in chiller systems based on CWGAN-GP and DS evidence theory fusion

Fault diagnosis of chillers in real-world scenarios is frequently hindered by data imbalance, leading to degraded diagnostic accuracy. To address this challenge, this study systematically investigates the application of traditional generative adversarial networks (GANs) and their variants, including conditional GAN (CGAN) and Wasserstein GAN with gradient penalty (CWGAN-GP), for generating fault-specific synthetic data. By comparing the diagnostic performance of balanced datasets constructed using different generative approaches, the superiority of CWGAN-GP is experimentally validated. In addition, due to the certain randomness of the generated samples, directly using them may reduce the learning ability of the model. Therefore, a strict screening method is needed to select higher-quality samples for dataset expansion. In this study, a high-quality sample screening method integrated with the DS evidence theory is proposed, and the Multi-Layer Perceptron (MLP) and Light Gradient Boosting Machine (LightGBM) are used as evidence sources to obtain the highest accuracy and F1 scores. The experimental results show that under the condition of limited real fault data, the comprehensive performance of fault diagnosis is improved through the data generation and screening process of CWGAN-GP-DS. Specifically, in the comparative experiment of the generative model, the diagnostic accuracy of CWGAN-GP after achieving data balance is 1.40 %-4.31 % higher than that of GAN and CGAN. Meanwhile, when the number of real fault samples is 20, the accuracy and F1 coefficient of the diagnostic model reach 81.58 % and 81.33 % respectively, which is an improvement of 7.56 % and 7.68 % compared with the original unbalanced situation. When the number of real fault samples is 50, they reach 91.89 % and 91.86 % respectively, which is an improvement of 2.34 % and 2.24 % compared with the original unbalanced situation. Finally, the screening method integrated with the DS evidence theory also highlights its advantages in QCP, ELQCP and SSL. The improvement in accuracy reaches 1.32 %-9.3 %, and the improvement in the F1 coefficient reaches 1.11 %-9.5 %. Although with the increase in the number of real samples, the training of the diagnostic model gradually matures and the improvement effect of the proposed method decreases, it can still effectively improve the diagnostic accuracy problem in scenarios with imbalanced and few samples in the overall situation. Experimental results from the ASHRAE Research Project RP-1043 validate the robustness of the CWGAN-GP-DS method in imbalanced data environments. This systematic combination of synthetic data generation and evidence-based quality control provides a reliable solution for chiller fault diagnosis under challenging conditions.