Fine-tuning transfer learning based on DCGAN integrated with self-attention and spectral normalization for bearing fault diagnosis

In the current big-data context of Industry 4.0, insufficient training data has become a major bottleneck in developing data-driven diagnosis approaches, restricting the accuracy of deep networks. Targeting this issue, this study proposes a novel fault diagnosis method incorporating data augmentation and transfer learning, which is branded as SA-SN-DCGAN-TL. The SA-SN-DCGAN method is used to generate sufficient synthetic images to meet the training requirement, which integrates the deep convolutional generative adversarial network (DCGAN) with the self-attention (SA) module and spectral normalization (SN). Besides, fine-tuning transfer learning (TL) is proposed to combine the synthetized and original data to train the deep network. The well-trained deep network is divided into two parts, wherein the weight parameter in the upper layers is trained on synthetic images and transferred into the counterpart of the target network. Only a small amount of original data is required to fine-tune the target network's bottom layers for the target task. Ablation studies confirm the importance and effec-tiveness of each component in the SA-SN-DCGAN. Experimental results on two bearing datasets demonstrate that the proposed method can relieve the growing demand for a large amount of original data in deep networks by utilizing SA-SN-DCGAN and the fine-tuning TL and achieve better fault diagnosis accuracy than the existing approaches.