Harmonising convolutional networks and Transformer for transfer learning in bearing fault diagnosis

Bearings are indispensable components in modern industrial systems, whose operating conditions are commonly monitored using vibration-based non-destructive testing (NDT) techniques. In practical applications, fault diagnosis models face challenges such as the difficulty of simultaneously capturing local and global features, as well as variations across operating conditions and equipment. A novel fault diagnosis approach is proposed, which leverages transfer learning and combines convolution with Transformer architectures. A shallow, learnable aggregator is designed to extract local fault information, where local dot-product operations replace token-based computations to reduce model complexity. A deep aggregator captures global dependencies through token interactions. By combining shallow and deep aggregators, the model achieves comprehensive feature representation of fault signals. Furthermore, the Higher-order Joint Maximum Mean Discrepancy algorithm was proposed by combining first-order and second-order statistics to measure the distance between the source domain and the target domain. The experimental findings indicate that the developed method attains high diagnostic accuracy with low computational complexity across multiple transfer tasks, offering dependable support for vibration-based non-destructive bearing fault diagnosis.