CNN-Transformer based Automated Neural Architecture Search for Ultrasonic Railway Track Defect Classification

This paper introduces a novel Neural Architecture Search (NAS)-based framework for the automated detection of ultrasonic railway track defects. To facilitate model training, a domain-specific dataset of B-scan images was constructed and pre-processed to enhance signal clarity and annotation accuracy. A supernet architecture was developed by integrating MobileNetV2, Transformer, and pooling modules, each with multiple candidate operations. NAS was employed to automatically identify the optimal combination of components for this task. Evaluations on a real-world dataset comprising 2,898 images demonstrated that the proposed model achieved a classification accuracy of 97.12 %, outperforming baseline models such as Vision Transformer (93.82 %) and EfficientNet-B2 (93.55 %). Ablation studies revealed that the integration of all three modules resulted in the best performance over the best partial configuration. These findings confirm that the NAS framework significantly enhances both accuracy and generalization for railway defect detection.