Research on Fault Diagnosis Method for Switch Machine Based on Improved Generative Adversarial Networks with Adaptive Data Alignment

Aiming at the relatively rare and unbalanced fault data of the switch machines on site, this paper proposed a fault diagnosis method for switch machines based on the improved generative adversarial network. This approach improved the losses of the ACGAN network with the goal of fault diagnosis for on-site data, expanded the distribution of samples in latent space by using the generator, learned the feature distribution patterns of fault curve by using auxiliary classifiers of the discriminator and trained the model with a small amount of unbalanced data to achieve accurate diagnosis of common faults. Given the inconsistent length of the action current data of the switch machine, this paper used adaptive data alignment as pre-process, which avoided the destruction of feature patterns by padding or truncating, further enhancing the fault diagnosis performance. Finally, the on-site monitoring data of the switch machine of Chengdu Metro Line 4 were used for fault diagnosis and compared with related methods. The experiments show that the method proposed in this paper has good diagnosis performance under the scenes of the unbalanced small amount of data set of switch machines, with a fault diagnosis AUC index of 0. 999. With high diagnostic accuracy and good real-time performance, the proposed method has good prospects for field application. © 2023 Science Press. All rights reserved.