Feature extraction and identification of multi-layer multi-pass welds based on region growing segmentation combined with deep learning

To address the significant challenges associated with the multi-layer, multi-pass automated welding of medium-thick plate, including uneven laser stripe brightness, local high reflection, serious noise, and complex and variable weld contour features, we propose an adaptive thresholding chunked-area growth segmentation combined with the YOLOv8 deep learning algorithm for weld seam feature extraction. A single-stripe laser and high-resolution camera are utilized to collect multi-layer, multi-pass weld images of fine wire submerged arc. Adaptive thresholding chunked area growth segmentation, convolutional filtering, and segment fitting are then applied to each layer of weld images to extract the weld feature points of each layer. The YOLOv8 algorithm is employed to ascertain the pixel coordinates of the feature points within the designated area of the multi-layer multi-pass weld image, thereby facilitating the construction of the planning robot weld tracking path. The experimental results demonstrate that the adaptive thresholding chunked region growth segmentation algorithm is capable of fully extracting the laser streak image and multi-layer weld seam feature points. After training, the deep learning model achieved a root mean square error (RMSE) of 0.0551 mm and a 98.41% feature point extraction accuracy, The proposed method demonstrates high feature extraction accuracy and robustness, effectively satisfying the requirements of bead identification. It is demonstrated that a well-formed multi-layer and multi-pass weld is obtained by this proposed algorithm, and provides an innovative solution for weld bead identification in multi-layer multi-pass welding of medium-thick plates.