Research On Weld Inspection Technology Based On Deep Learning

With the popularization of robots in Chinese industrial manufacturing field and the increasing maturity of digital image processing technology,machine vision technology has been more and more applied in industrial welding.Robot based intelligent welding technology has been widely used in many key fields such as aerospace,machinery manufacturing and petrochemical industry,and has become one of the key core technologies in the field of automatic welding.Considering the factors such as production quality,production cost and welding site working environment,the traditional manual welding method can not meet the market demand.With the help of weld detection technology based on deep learning,the detection accuracy and efficiency can be improved,which has great research and application value.Through the research on weld detection and related robot calibration technology,this paper finally determines the target detection algorithm used in weld detection technology and the structure of depth neural network used in robot calibration.The specific work is as follows:(1)Aiming at the problem of detecting the position of weld feature points,through the analysis and comparison of R-CNN series algorithms based on region proposal and YOLO series algorithms based on regression,this paper finally determines that YOLO based target detection algorithm is adopted for weld detection,and proposes a weld target detection technology based on improved YOLOv4-tiny algorithm,Thus,the position of weld feature points on the image collected by the camera can be extracted quickly and accurately.Firstly,the image annotation software is used to manually label the collected weld image,and the self-made weld image data set is made.Then the self-made data set is used to train the improved YOLOv4-tiny model offline to obtain the weight parameters.Finally,the weld image obtained by the camera in real time is transmitted to the trained neural network to realize the real-time detection of the position information of weld feature points.(2)This paper studies the robot vision calibration technology in the weld tracking system,proposes a calibration method of weld detection robot system based on depth neural network,establishes the coordinate mapping method of weld feature points based on depth neural network,and accurately maps the image coordinates of weld image feature points to the three-dimensional space coordinates in the world coordinate system.The experimental results show that this method can realize the simple and fast calibration of the robot,avoid the complex nonlinear operation in the traditional calibration method,reduce the cumulative error between coordinate conversion,and the positioning accuracy meets the welding requirements.(3)In this paper,the function test experiment of weld tracking system based on deep learning is carried out.Experiments are carried out on the modules of weld feature point detection and robot vision system calibration,and the experimental results are compared and analyzed.The experimental results show that the average accuracy of the improved YOLOv4-tiny algorithm is 95.3%,which is significantly improved compared with the YOLOv4-tiny algorithm.The experimental results show the effectiveness and practicability of the weld detection technology based on deep learning.It also shows that based on the traditional image processing technology and calibration algorithm,the deep learning algorithm with data science as the core can provide a new method for robot weld tracking technology.