摘要：Pipeline transportation serves as the primary method for conveying oil and gas, with welding being the predominant means of connecting pipelines. Nevertheless, weld defects frequently occur at circumferential welds due to various factors including the welding process and environmental conditions. Failures due to these welding defects threaten the safety of pipeline transportation, and X-ray inspection is used to visually detect defects. While manual inspection has traditionally been employed for X-ray defect detection, it suffers from inefficiency and subjectivity. As computer vision techniques have advanced, methods leveraging deep learning have emerged for weld defect detection. However, the practical application of deep learning methods is constrained by their size, limiting their effectiveness in industrial detection tasks. To address these challenges, this paper proposed a lightweight pixel-level automatic detection network for weld defects based on an encoder-decoder structure. Firstly, a lightweight global feature extraction block utilizing self-attention was introduced to enhance detection performance by extracting more discriminative features within defect areas. Subsequently, a high-low level feature fusion block was incorporated to mitigate feature loss resulting from downsampling. Additionally, the depthwise separable convolution was employed to further reduce network size. Our method demonstrated superior detection performance compared to other advanced methods, as evidenced by qualitative comparison and quantitative analysis, with dice coefficient (DICE) and intersection over union (IoU) scores of 0.881 and 0.789, respectively. Moreover, our network exhibited significantly reduced size compared to other advanced detection methods. Thus, the proposed detection network holds considerable promise for effectively detecting weld defects. ? 2024 by ASME.

ISSN号：0277-027X

卷、期、页：卷5

发表日期：2024-01-01

期刊分区（SCI为中科院分区）：无

收录情况：EI（工程索引）,CPCI-S（科技会议录索引）

发表期刊名称：American Society of Mechanical Engineers, Pressure Vessels and Piping Division (Publication) PVP

通讯作者：钱伟超,任庆滢

第一作者：董绍华,陈林

论文类型：会议论文

论文概要：钱伟超,董绍华,任庆滢,陈林，A LIGHTWEIGHT MODEL OF AUTOMATIC PIXEL-LEVEL DETECTION FOR WELD DEFECTS，American Society of Mechanical Engineers, Pressure Vessels and Piping Division (Publication) PVP，2024，卷5

论文题目：A LIGHTWEIGHT MODEL OF AUTOMATIC PIXEL-LEVEL DETECTION FOR WELD DEFECTS