Weld defect detection and severity assessment via phased array ultrasonic imaging and unsupervised anomaly segmentation

The assessment of weld defect severity is critical for ensuring the safe operation of subway trains. However, existing weld defect detection methods suffer from low inspection efficiency, high costs, and the lack of a comprehensive defect severity assessment system. To address these challenges, this paper proposes a weld defect severity assessment framework that integrates phased array ultrasonic testing (PAUT) with an unsupervised anomaly segmentation network. The proposed framework enables end-to-end defect detection and localization using only normal samples for training. It employs a feature-space-based anomaly signal synthesis strategy to guide the model in distinguishing normal from abnormal features while suppressing interference from non-defect signals. In addition, a defect severity assessment and classification method based on anomaly maps is developed to quantify defect severity and establish mappings to actual defect types. Furthermore, an efficient unsupervised network is designed to address the challenges of sparse defect samples and high annotation costs in real-world scenarios. Experimental results demonstrate that the proposed method achieves detection accuracies of 99.76% at the image level and 99.94% at the pixel level, significantly outperforming seven mainstream unsupervised methods, and enables precise localization and rapid evaluation of weld defects.