The regular detection of weld seams in large-scale special equipment is crucial for improving safety and efficiency, and this can be achieved effectively through the use of weld seam tracking and detection robots. In this study, a wall-climbing robot with integrated seam tracking and detection was designed, and the wall climbing function was realized via a permanent magnet array and a Mecanum wheel. The function of weld seam tracking and detection was realized using a DeepLabv3+ semantic segmentation model. Several optimizations were implemented to enhance the deployment of the DeepLabv3+ semantic segmentation model on embedded devices. Mobilenetv2 was used to replace the feature extraction network of the original model, and the convolutional block attention module attention mechanism was introduced into the encoder module. All traditional 3×3 convolutions were substituted with depthwise separable dilated convolutions. Subsequently, the welding path was fitted using the least squares method based on the segmentation results. The experimental results showed that the volume of the improved model was reduced by 92.9%, only being 21.8 Mb. The average precision reached 98.5%, surpassing the original model by 1.4%. The reasoning speed was accelerated to 21 frames/s, satisfying the real-time requirements of industrial detection. The detection robot successfully realizes the autonomous identification and tracking of weld seams. This study remarkably contributes to the development of automatic and intelligent weld seam detection technologies.
Keywords: DeepLabv3+; inspection robot; wall-climbing robot; weld seam identification.