An Improved YOLOv8n-Based Method for Ultrasonic Phased Array S-Scan Defect Detection in LPG Spherical Tank Welds

Abstract

To address the high missed-detection rate of traditional object detection methods caused by small defect sizes, irregular shapes, and low contrast in ultrasonic phased array S-scan images of LPG spherical tank welds, an improved YOLOv8n-based weld defect detection method is proposed. Building upon the lightweight architecture of YOLOv8n, a Multi-Level Channel Attention (MLCA) mechanism is introduced to enhance the model’s capability in representing low-contrast and small-scale defect features. In addition, the C2f module is integrated with Deformable Convolution (DCNv3) during the feature extraction stage to construct a C2f-DCNv3 feature extraction module, thereby improving the model’s ability to capture irregular weld defect characteristics. Experiments are conducted on an LPG spherical tank weld ultrasonic phased array S-scan image dataset. The results demonstrate that, compared with the original YOLOv8n and several mainstream object detection methods, the proposed approach achieves significant improvements in Precision, Recall, and mAP@0.5. Moreover, it satisfies the real-time requirements of engineering applications while maintaining high detection accuracy. The findings indicate that the proposed method provides an effective technical solution for intelligent ultrasonic inspection of LPG spherical tank welds.