Railway is a key component driving innovation and sustainability in transportation systems. Aiming at solving the problems of metal reflection, oil contamination and complex background interference in railway wheel tread defect detection, this paper will focus on the railway wheel tread defect detection method, SEN-YOLO, based on the YOLOv5s and the comparison between different generations of YOLO detection. To better adapt the model to actual detection scenarios, multi-stage dynamic data augmentation strategy combining illumination robustness optimization and motion blur simulation is designed to construct a railway wheel dataset that closely mirrors real-world conditions. In terms of model architecture, the YOLOv5s-based approach integrates the Squeeze-and-Excitation Networks (SENets) module to enhance the capture of minor defect features and employs an adaptive feature fusion strategy to mitigate background noise. To further improve detection accuracy and generalization, the YOLOv5s network structure is optimized. A multi-scale feature fusion algorithm in the backbone network improves processing of different scale defect features; Additionally, the feature enhancement module is depoyed in the neck network boosts feature expression and the ability to distinguish; Furthermore, an adaptive loss function in the head network balances detection effects. Trained on the open-source railway wheel tread defect dataset from Roboflow and tested, SEN-YOLO achieves 84.1% mAP on the custom dataset and a detection speed of 35 FPS on an RTX3070 GPU, a 12% improvement over the baseline YOLOv5s model. Comparative experiments with other mainstream object detection algorithms show that SEN-YOLO performs well with high detection accuracy and stability, meeting the practical needs of railway wheel tread defect detection.