| Welding technology is developing in the direction of automation and intelligence.During welding,it is important to accurately locate the key position of the weld groove.Different welding scenes will have different groove types of welds,and the welding environment is complex and changeable,the collected weld images are accompanied by different degrees of noise,splash,arc light and other interference,the traditional weld detection tracking algorithm can not fully adapt to complex weld tracking.With the development of artificial intelligence technology,intelligent welding requires weld tracking algorithms to have high precision,certain anti-interference ability and adaptability to meet welding in different scenarios.In this paper,combined with the characteristics of weld image and convolutional neural network,the key position detection scheme and weld tracking scheme of weld are designed,and the effectiveness of the proposed method is verified by experiments.The main research content of this article is as follows:(1)Analyze and summarize the characteristics of weld image in depth.There are different types of weld grooves in different welding scenes,and the shape of weld grooves will be deformed in the welding process.Most of the weld images are background information and often accompanied by noise interference.Useful weld information is distributed near the weld laser bar,and the feature distribution is sparse.According to the welding requirements,the key position detection and seam tracking scheme of the weld is designed by using the powerful feature expression ability of convolutional neural network.(2)In order to improve the positioning accuracy and anti-interference ability of key positions of weld groove,a detection network of key positions of weld seam based on multiscale feature fusion is designed.The network attention mechanism is used to focus on the feature information near the laser stripe of the weld seam from the sparse feature of the weld seam image,which suppresses the interference of noise to a certain extent.The low-level texture,position,grayscale and other information of the network are fused with the highlevel semantic information,and various feature information is integrated to infer the key position of weld groove,which improves the accuracy of weld positioning.(3)In order to further improve the welding efficiency,using the continuity and predictability of the welding process,a seam tracking network based on twin structure is designed.In the first frame,the weld key position detection network locates the weld feature points and determines the template area to be tracked.When tracking,template features are used to quickly match and locate feature points near weld feature points,which ensures the tracking speed.The mixed domain attention mechanism is introduced to strengthen the important weld features from channel and space dimensions.Channel attention makes the network pay attention to the important feature channels,shielding noise and background interference,and space attention emphasizes the important spatial features near the inflection point of the weld laser line to ensure the accuracy of weld tracking.The experimental results show that the network designed in this paper has high precision,the average tracking error is0.155 mm,the maximum tracking error is 0.572 mm,and the network model parameters are only 2.28 M,which basically meets the requirements of the seam tracking system. |
