Fault Diagnosis of the Chiller Under Imbalanced Data Environment via Combining an Improved GAN and the Random Forest Algorithm

Fault diagnosis is crucial for chiller's safe operation. However, current methods neglect the imbalanced data issue of chiller, which results in reduced the precision of the minority fault patterns. In order to address this problem, an enhanced fault diagnosis method of chiller via an improved GAN method and random forest algorithm (IGAN-RF) is developed in our work. Firstly, a multi-head attention method is incorporated into the GAN method. The multi-head attention mechanism enables the generator to focus more on the key features of minority class samples during the generation process, which makes the generator generate high quality rebalanced fault pattern datasets. Then, the rebalanced training dataset obtained by the IGAN is used to train multiple decision trees using the random forest method. Finally, the chiller's fault pattern is determined using the majority voting method. The effectiveness of the IGAN-RF method under imbalanced data environments is confirmed through specific diagnostic results obtained from the ASHRAE RP-1043 datasets.