A novel transfer-learning method based on selective normalization for fault diagnosis with limited labeled data

The application of deep learning to fault diagnosis has made encouraging progress in recent years. However, it is hard to obtain sufficient labeled data to ensure the performance of diagnostic models, due to complex and varying working conditions. Over-fitting often occurs when few labeled data are used in training. To address this crucial problem, a novel transfer-learning method called the selective normalized multiscale convolutional adversarial network (SNMCAN) is proposed in this paper. The proposed model introduces multiscale convolutional neural networks (CNNs) to capture rich fault feature information at multiple scales. A batch normalization (BN) module, widely used in CNNs, is reconstructed into a new normalization method called 'selective normalization' to learn diagnostic knowledge from a pre-trained model and avoid over-fitting with limited labeled data. Joint maximum mean discrepancy (JMMD) is applied to minimize the joint distribution discrepancy between different domains and improve the results of domain alignment. An adversarial training strategy is also used in the proposed model to easily distinguish the distributions of the source and target domains. The superiority of the proposed method is demonstrated using two case studies. The case study results demonstrate that the SNMCAN can achieve better performance in fault diagnosis than comparison methods.