Research On Intelligent Welding Technology Based On Deep Learning

Intelligent welding technology has broad application prospects in the field of industrial production,however there are still some problems to be solved urgently.For example,the harsh industrial production environment greatly affects the accuracy of weld recognition,the system calibration is difficult,the feature matching accuracy of the binocular system is low,there is no efficient welding path planning algorithm,and its efficiency is low.Based on the deep learning of traditional welding technology,this article has made innovations,improved the defects of traditional welding technology,and proposed an intelligent welding system based on deep learning.After the calibration of each part of the system,the image collected was de-noised.Then,the CNN was used to identify the weld,and the 3D coordinates of the weld were obtained by combining binocular correction,binocular matching and parallax-matching.Finally,the path planning of the weld was carried out to control the robot welding.The specific work content is as follows:(1)Organize,summarize and study in-depth intelligent welding technology at home and abroad,analyze its advantages and disadvantages,and conduct key research on several unresolved problems.(2)Design and build an intelligent welding system,and analyze the camera's internal model,binocular vision model,hand-eye relationship model,and robotic arm model at the same time,calculate the coordinate conversion relationship in each model,and give specific calibration methods accordingly.In order to facilitate the calibration calculation,this paper designs an automatic calibration software written in C++,which can achieve single camera calibration,binocular system calibration,and hand-eye calibration.(3)In order to improve the accuracy of subsequent intelligent recognition,the image taken by the binocular camera is denoised.After the performance research and experimental analysis of various image filtering algorithms,the denoising effects of different image filtering algorithms are compared.Finally,the NL-means denoising algorithm is used to strengthen the characteristics of the weld.(4)Determine the CNN model suitable for the system through experimental analysis and comparison,pre-train the model,collect and make training samples and test samples,use the sample set to train and convert the model,and deploy it to the tree equipped with edge computing sticks on the Raspberry Pi,the trained model will finally be tested to test its performance.(5)Make all the polar lines of the binocular camera parallel to each other through binocular correction,and use the NCC algorithm for feature matching,use the edge detection algorithm to extract the weld information,and distinguish the weld and interference factors according to the specific matching algorithm to obtain the weld in the 2D coordinates in the left and right images,finally the 3D position information of the weld is calculated by the binocular parallax matching algorithm.(6)Improve the traditional ACO and apply it to weld path planning.Through experimental analysis,it is found that the improved ACO has a qualitative improvement in solving TSP questions and welding path planning questions.The average distance and minimum distance of the planned trajectory become smaller,and the optimal path can be found faster.The system is transmitted to the UR3 robot via Ethernet and controlled to perform intelligent welding.