Journal of Engineering Research and Reports

Steel structure buildings, as an important component of modern infrastructure, have safety and reliability directly related to public safety. Weld seams, as the critical connection points of steel structures, are prone to defects such as cracks, pores, and slag inclusions during long-term service. The accumulation of these defects may lead to structural failure. Traditional manual inspection methods are not only inefficient but also pose significant safety hazards in complex environments such as high altitudes and large structures. Therefore, developing intelligent weld seam inspection robot systems has important engineering application value. Magnetic crawler wall-climbing robots can stably attach to and flexibly move on vertical steel structure surfaces, providing an ideal platform for automated inspection. This study focuses on the application of instance segmentation technology in weld seam recognition, aiming to build an efficient and reliable visual navigation system for wall-climbing robots. The research objective is to establish a weld seam recognition system based on SOLOv2-R18 model to provide reliable visual navigation support for magnetic crawler wall-climbing robots. The main contributions of this research are: first application of SOLOv2-R18 instance segmentation algorithm to industrial weld seam detection, proving its advantages in handling elongated targets; establishment of a weld seam dataset containing 3,070 images, providing a data foundation for related research. Experimental results confirm that the proposed method outperforms mainstream algorithms like Mask R-CNN and YOLACT in terms of speed-accuracy balance, achieving 96.4% AP with 30.4 FPS inference speed.