Research On Robot Weld Recognition And Tracking System Based On Deep Learning Under Strong Noise

Welding plays an important role in various industrial manufacturing fields,such as construction,automobile,shipbuilding,urban pipeline,etc.However,due to its harsh environment and high risk,it leads to a decline in the enthusiasm of employees and a serious shortage.Therefore,industrial robots to replace manpower to complete welding work has become the main way to achieve intelligent manufacturing and staff reduction gain.There are strong noises such as arc light,a lot of smoke and spark spatter in the welding process,which makes the traditional weld recognition method based on image processing perform poorly in complex environments such as multi-weld target recognition,multi-layer multipass welding and so on.In this paper,on the basis of the research and analysis of the traditional robot welding technology,with the help of the advantages of non-contact and high-precision measurement of the vision sensor,combined with the strong feature learning and expression ability of deep learning,the weld recognition and detection in complex and strong noise environment is realized and the weld tracking software is developed,which lays a technical foundation for more flexible and intelligent robot autonomous welding.The specific work is as follows:(1)A robot seam tracking system based on line structure laser vision guidance is built.The vision module of the system is calibrated,and the mapping relationship between the pixel coordinates in the weld image and the 3D world coordinates of the welding robot is established.(2)Using deep learning technology,an end-to-end image denoising depth model is proposed.Dense network structure,mixed attention block,depth supervision optimization and multi-scale cascade output are designed and applied in the model to improve the segmentation and generalization performance of the network.(3)Configure the hardware and software environment of the model network training,make the model training data set,and design the model training strategy,including the model training flow,the selection of evaluation index and the design of model training loss function.On this basis,carry out in-depth network training.(4)Verify the effectiveness of the trained denoising network.Firstly,the optimal denoising network is selected by ablation analysis,and then the extraction efficiency of the network is compared and analyzed.By comparing the detection results of the algorithm,it is proved that the proposed algorithm can effectively extract laser stripes under the background of strong noise.(5)A welding software platform for line structure laser vision welding seam automatic recognition and tracking is built.it integrates different functional modules,such as serial communication module between robot and industrial computer,vision system calibration module(industrial camera,laser plane,robot hand-eye calibration),welding image acquisition module,weld feature extraction module and so on.Convenient visual management in the welding process,the realization of flexible and convenient humancomputer interaction.Combined with the software platform,the experiments of laser stripe centerline and weld feature point extraction are carried out.The results show that the system can extract weld feature points accurately under the strong noise environment of arc and spatter interference.The maximum accuracy error of welding torch welding coordinate point extraction is 0.0730 mm,the average extraction accuracy error is 0.0219 mm,and the extraction efficiency is 65 ms,which can meet the needs of actual welding production.