Unsupervised Domain Adaptation: A Residual Attention Network for Gear Pump Fault Diagnosis

To ensure the reliability of products, deep learning-based fault diagnosis methods have received extensive attention in academic research and industry. These methods usually have two default assumptions: (1) a large number of labeled training samples, and (2) training and test data share similar distributions under constant operating conditions. Therefore, we propose a multi-kernel domain adaptation method based on residual network with transfer attention mechanism (ResNetTAM-MKDA) to solve the problem that the distribution difference between target domain data and source domain data in unknown conditions affects the model performance, and there is no labeled data in the target domain. First, we utilize to extract shallow features in the source and target domains and feed them into a stacked residual convolution module to learn higher-order generic fault information. Then, the higher-order fault features with higher transferability are weighted using the Migratable Attention Module, and the domain alignment is achieved using a multi-core domain adaptation method. Finally, the fault diagnosis classification module constructed using fully connected networks is used to map the higher-order fault features into fault labels. In order to verify the effectiveness of the proposed method, we simulate the gear pump contamination wear experiment to get the failure data under different working conditions. The effectiveness of the proposed method is verified in the six diagnostic tasks set up.