Welding defect detection based on phased array images and two-stage segmentation strategy

The rail transit vehicle body is composed of numerous welded structures, and to prevent failures during operation, it is essential that each weld undergoes strict and accurate quality inspection. Integrating segmentation algorithms with phased array ultrasonic testing (PAUT) offers a novel solution for the quality inspection of train welds. However, due to the high sensitivity of the phased array method in detecting weld defects, erroneous signals may be generated in non-welding areas, interfering with the judgment of deep learning algorithms and leading to incorrect detection results. To address the issue of existing algorithms being unable to completely eliminate false signals, this paper proposes a welding defect segmentation network with regional determination capabilities, which leverages both the defects and valid regions in phased array welding images. The concept of the proposed region determination performance is founded on establishing region-type rules for the defect detection task. Specifically, it involves the design of a two-stage network to assist in formulating the rules, along with a determination module to refine them. To assess the rationality and effectiveness of the proposed method, various parameters and modules of the model undergo extensive testing. The experimental results demonstrate that by splitting the defects and the valid regions in phased array welding images, reasonable and necessary determination rules can be constructed. This approach leads to more efficient and accurate weld defect segmentation.