Online detection and evaluation of weld surface defects based on lightweight network VGG16-UNet and laser scanning

Online inspection of welding defects is of great help for welding quality assessment, rapid process optimization, and production efficiency improvement. However, in-situ detection is difficult due to the noisy environments, real-time changes in light, diverse weld forms, and requirements for rapid detection. Traditional defect detection methods face many challenges, such as complex processes, slow speed, insufficient information, and weak adaptability. In this study, an online fast detection method based on laser scanning and a lightweight network is proposed. Firstly, high-precision point clouds of weld surfaces are acquired using a laser displacement sensor. Secondly, equal-interval splitting divides the point cloud of a long-seam weld into equal-resolution subframes for online detection. An image processing algorithm combining radius filtering, Sobel convolution, normalization, and Gaussian enhancement is proposed to convert the split point clouds into high-contrast gray-scale images. With the gray-scale images as input and pixel-level classification of the background and three types of defects as output, a semantic segmentation model, VGG16-UNet, is trained through transfer learning, achieving 88.77 % precision and 92.03 % recall. Based on the prediction result and its mapping point cloud, 3D reconstruction and visualization of defects are realized using Delaunay triangulation. 3D location and geometry dimensions of defects are calculated for defect assessment. Compared with the alternative segmentation models, VGG16-UNet performs better while having a smaller weight (94.95 MB) and a faster inference speed (26.8fps), making it suitable for online industrial applications. The work provides strong support for achieving online automated evaluation of weld surface quality.