<p>Acoustic emission has become one of the most commonly used non-destructive testing techniques in the track crack detection industry due to its advantages in dynamic monitoring and real-time online detection. Transformer models construct global dependencies through self-attention layers, bringing more possibilities for feature extraction, but they are limited in the ability to extract local features. In order to further improve recognition accuracy and robustness, this paper designs a Transformer based on multi-scale local perception and contrastive learning for train axle fatigue crack acoustic emission detection. The core of this method is the collaborative design of its multi-scale local perception, local-global coupling architecture, and contrastive learning optimization, which breaks through the inherent limitations of traditional Transformer in acoustic emission signal processing and provides a highly robust solution for fatigue crack detection under complex working conditions. Specifically, after capturing global dependencies through the multi-head self-attention module, the convolutional module captures the local features of the sequence to provide more contextual information. By simultaneously incorporating multi-scale convolutional layers to enhance the generalization ability of the model. To eliminate the uncertainty of model predictions, this study also designed an optimization task that combines cross-entropy loss and supervised contrastive learning to enhance fine-grained feature representation capabilities. Finally, the proposed method was evaluated on the collected dataset. The experimental results show that the accuracy of the classification method reached 99.19%, achieving accurate identification and classification of fatigue crack signals and providing a novel and highly promising solution for the diagnosis of fatigue crack faults in train axles.</p>