Dual-Attention Generative Adversarial Networks for Fault Diagnosis Under the Class-Imbalanced Conditions

Deep learning has been widely applied to intelligent fault diagnosis with balanced training set. However, certain available fault data are extremely limited, resulting in an imbalanced training set in most cases inevitably. In general, the performance of the deep learning-based diagnosis methods will deteriorate on the imbalance dataset. To solve the problem, a novel dual-attention generative adversarial network (DAGAN) is proposed for dealing with imbalanced fault diagnosis. Firstly, an attention model is constructed to selectively enhance the features at each position and adaptively fuse the interdependent channel maps. Then, the attention model is embedded into generative adversarial network (GAN) to capture the informative features of inputs and improve feature representations. As such, the DAGAN can learn the fault-related features effectively and generate sufficient fault samples. Finally, the diagnosis model can be trained on the rebalanced dataset to improve its classification performance under class-imbalance conditions. Two different datasets are used to validate the proposed method, and the effects of the multiple imbalance ratios on classification performance are discussed. Results show that the proposed method achieves high diagnosis accuracy and outperforms other methods.