Research On Laser Weld Seam Recognition And Tracking System Based On Deep Learning

Robot-based laser welding can perform high-speed,high-precision,and high-quality welding tasks,greatly improving production efficiency.It has broad application prospects in fields such as mechanical manufacturing and automobile manufacturing.However,due to various uncertain factors in the industrial production environment,such as thermal deformation of welded parts and clamping errors,the actual weld trajectory deviates from the preset welding trajectory during the welding process.The robot cannot effectively adapt to changes in the weld trajectory in real-time during the welding process.Therefore,this paper proposes a laser weld recognition and tracking system based on structured light sensing technology and neural network algorithms.Experimental verification was carried out using V-groove welded joints as an example.This paper’s work and research mainly include several aspects:Firstly,the relevant research and development status of the weld tracking system were analyzed.Taking a six-degree-of-freedom industrial robot as the main body,an experimental hardware platform for weld tracking was built,including tracking modules,lasers,PC main control machines and other devices.This provided conditions for collecting image data to train neural networks and conducting tracking experiments.Secondly,in response to the problem that traditional image algorithms cannot accurately extract feature stripes in strongly disturbed images containing arc light and splashing,a convolutional neural network is applied to laser feature stripe extraction.A lightweight network based on ERFNet is proposed for V-shaped feature weld seam stripe segmentation.To improve the computational efficiency of the ERFNet model,reduce the size and number of parameters of the model,and improve the network through relevant strategies such as applying deep separable convolution technology and channel attention mechanism.A real-time semantic segmentation model for weld seam feature stripe extraction was constructed.The pixel coordinates of weld seam feature points can be extracted by using gray center method,least squares method combined with neural network inference from images.Experimental results show that compared with ERFNet,DFA-Net and Seg Net,this constructed network has higher segmentation accuracy and faster segmentation speed.The segmentation speed of processing 640x640x3 pixel size welding images on a mobile device with 4GB RAM GTX 1050 Ti GPU can reach 82.1FPS while MIOU is 0.906 and success rate of extracting feature point pixel coordinates reaches 96%.Thirdly,a visual sensing system based on line structured light plane was constructed.The pinhole camera model was analyzed and the transformation relationship between the pixel coordinate system and the camera coordinate system was obtained by combining with the equation of structured light plane.By establishing a hand-eye model(eye in hand),the transformation relationship between the camera coordinate system and robot end coordinate system was obtained.The camera intrinsic matrix and extrinsic matrix were obtained using Zhang’s chessboard calibration method.A set of sensor calibration experiments were designed to calculate and obtain the equation of structured light plane,and Tsai’s two-step method was used to solve for hand-eye matrix.After calibration calculation and experimental verification,it is found that the average absolute errors in X,Y,Z directions of robot based on line structured light visual sensing system are 0.28 mm,0.20 mm,1.22 mm respectively.Finally,the development of a weld seam recognition and tracking system has been completed.Third-party libraries such as OpenCV,Libtorch,and robot motion control were integrated to provide tracking,image processing,and robot functionality.A user-friendly interface was created using Qt software that is both visually appealing and easy to operate.The model for weld seam tracking based on line structured light was analyzed,with PID controllers used to compensate for tracking errors.Experimental results show that the system has high accuracy in tracking with an average absolute error of 0.537 mm between welding trajectory and actual weld seam,while the maximum error is 1.558mm;furthermore,the welding trajectory is smooth resulting in good surface quality of welded joints.