Intelligent welding robot system based on deep learning

• Published in: 2019 Chinese Automation Congress (CAC) pp. 2944 - 2949
• Main Authors: Ma, Xufeng, Pan, Shuwen, Li, Yanjun, Feng, Chao, Wang, Aoran
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01-11-2019
• Subjects: Automatic welding; Binocular vision; Calibration; Cameras; Deep learning; Intelligent recognition; Matrix decomposition; Robot kinematics; Robot vision systems; Welding
• ISSN: 2688-0938
• DOI: 10.1109/CAC48633.2019.8997310

An intelligent welding robot system based on deep learning is designed. Firstly, the binocular vision system needs to be fixed and calibrated, and correct the left and right cameras to make the polar lines of the left and right images collinear. Then, the left camera is used to take 1000 photos of arbitrary arranging fence. The position of the weld is marked by LabelImg software, and the training set and validation set are determined in a ratio of 3:1, and then used to train the convolutional neural network. After the loss function tends to converge, the training stops. The model optimizer converts the TensorFlow model into an IR model through the model optimizer in openvino, which is then deployed to the Raspberry Pi 3B+, which performs forward propagation through Intel's NCS2 edge computing stick to identify the location of the weld in the image. The exact position of the weld is found by machine vision, and the parallax of all the welds is calculated by matching with the right image, and the depth of the weld is obtained by calculation. Finally, transfer the UR3 robot through Ethernet and control it for automatic welding.