Generalization of Deep Neural Networks for Imbalanced Fault Classification of Machinery Using Generative Adversarial Networks

Mechanical fault datasets are always highly imbalanced with abundant common mechanical fault samples but a paucity of samples from rare fault conditions. To overcome this weakness, the simulation of rare fault signals is proposed in this paper. Specifically, frequency spectra are employed as model signals, then Wasserstein generative adversarial network (WGAN) is implemented to generate simulated signals based on a labeled dataset. Finally, the real and artificial signals are combined to train stacked autoencoders (SAE) to detect mechanical health conditions. To validate the effectiveness of the proposed WGAN-SAE method, two specially designed experiments are carried out and some traditional methods are adopted for comparison. The diagnosis results show that the proposed method can deal with imbalanced fault classification problem much more effectively. The improved performance is mainly due to the artificial fault signals generated from the WGAN to balance the dataset, where the signals that are lacking in training dataset are effectively augmented. Furthermore, the learned features in each layer of the generator network are also analyzed via visualization, which may help us understand the working process of the WGAN.