Domain Adversarial Transfer Learning for Bearing Fault Diagnosis based on Data Fusion and Attention

In the research of equipment health status estimation, transfer learning has been greatly used and developed. However, during the application of transfer learning, it is found that when the data quality is poor or the data distribution between the training and testing objects is significantly different, the transfer learning exhibits a good performance in the source domain, but a poor performance in The target domain. This means that the network does not learn the essential relationship between signal and fault characteristics. To overcome the above problems, the data fusion and attention mechanism are introduced in this research. By fusing the original data with the manually extracted feature data at the feature level, certain guidance can be introduced under the condition of ensuring the integrity of information. At the same time, the attention mechanism gives more weight to the signal fault location, which enables the network to reduce the interference of unimportant signals and learn more generalized fault feature recognition patterns. Then it's introduced the above method into the Domain Adversarial Neural Network. The results validate that compared with original neural network, the method of introducing the data fusion and attention mechanism has a significant improvement effect.