Research On Automated Welding Based On Active Vision Line Structured Light Guidanc

Fusion welding and additive manufacturing is popular in the modern welding field.However,because of harsh welding environment,variable forms and stricter quality requirements,the traditional industrial robot based on manual teaching cannot meet the requirements of large-scale and high-quality industrial production.It is still very difficult to realize the automatic welding of intelligent robot.The extraction of welding feature is the key to realize welding automation.The interference of strong noise such as arc and splash make the traditional morphological image processing technology difficult to obtain ideal results.As a result,it is necessary to propose a welding seam feature extraction algorithm with high robustness,high precision and high efficiency.The thesis studies on the active visual line structure welding seam tracking technology for the complex and various addictive welding industrial scene.The target segmentation and target tracking based on deep learning is introduced to realize the accurate online extraction of welding features.The main contents of this thesis are as follows:(1)A seam tracking system and supporting software based on line structured light is designed and built.Through the complete camera calibration,line structured light plane calibration and robot hand-eye calibration,the two-dimensional image coordinates collected by the camera are directly converted into the motion coordinates that the robot can perform.By this way,the system can directly transfer the welding feature point coordinates to the robot through the weld feature extraction process.It provides the basis for subsequent path planning and online position correction.A variety of welding experiments show that the system has stable and reliable performance.(2)The plate additive welding seam tracking technology based on active vision is proposed.However,the interference of strong noise such as arc light and splash in the welding process brings about the difficulty of the feature extraction.In this thesis,the line structured light vision sensor is designed to obtained the welding images.Besides,the target segmentation in the deep learning is applied to the welding information extraction.The PSPNet network is the basic framework for feature extraction to simultaneously obtain the center line of the laser stripe and welding feature points.The phenomenon of extreme imbalanced of samples is solved by using online hard example mining(OHEM)and Dice Loss as the loss functions.The results show that the error of the feature points extracted by the algorithm and manual mark in each dimension of the weld pass is less than 0.70 mm,the average height error is less than 0.50 mm,and frame rates can reach more than 30 FPS.So,it meets the requirements of high precision,stability and online ability in the field of additive manufacturing.(3)The welding seam feature extraction algorithms based on single frame image such as target detection and target segmentation are useless for a variety of strong noise additive welding scenes.In this thesis,based on the strong noise continuous weld image sequence obtained by the large groove filling additive scene,the target tracking in the deep learning is applied to weld information extraction.Besides,the Siam FC tracking algorithm is used as the tracker of welding feature points.The reliability of the algorithm is verified by comparing the performance of the algorithm with morphological extraction algorithm and kernel correlation filter tracking algorithm.The results show that no model drift exists in the Siam FC tracking algorithm,and the regression absolute error of feature points is within 4 pixels.In the actual welding process,the dimensional error of solder joints obtained by the algorithm and manual mark is within 1.00 mm.Finally,it verifies the accuracy and robustness of the tracking algorithm.