Abstract: Aiming at the low detection rate and slow efficiency of defects (especially minor defects) in train axle ultrasonic detection, a method based on deep learning was proposed. On the basis of YOLO v5s network, the feature extraction layer structure was improved and SE attention mechanism was added. The dataset was constructed using real axle detection data, CIVA simulation data and GAN generated data and validation experiments were conducted. The experimental results showed that by adding simulation data and GAN generated data samples, this proposed method can effectively improve the detection rate of the actual axle ultrasonic detection defects, and the detection rate reached 99.25%, which showed a high application value and prospect.