A multi-module generative adversarial network augmented with adaptive decoupling strategy for intelligent fault diagnosis of machines with small sample

摘要

In actual industrial environment, intelligent diagnosis method requires a sufficient number of samples to ensure application effect. However, once industrial system fails, it usually stops immediately, resulting in extremely limited fault signals collected by monitoring system. Lack of fault samples makes the model difficult to fully train and tends to over-fitting, which makes the effect of intelligent diagnosis method poor. To solve this problem, a multi-module generative adversarial network augmented with adaptive decoupling strategy is proposed. Firstly, we creatively use an adaptive learning method to update the latent vector instead of sampling from specific distribution. The digits of the latent vector obtained in this way can come from different distributions, and a better combination effect can be obtained. Then, in order to solve mode collapse, conditional controller obtained through category label is used to directly adjust each intermediate output of generator to prevent influence of conditional control from weakening as the transmission path extends. Finally, in order to provide stronger constraints for the generator, reconstruction module is added to the network to force final generated samples to cover each subdomain of the target dataset. Two experiments are conducted, and average accuracy of 95.32%, 96.86% are achieved by proposed method.