Objective Weld tracking based on laser vision is widely used in automatic welding due to its noncontact, high precision, and other advantages. It is critical to obtain precise key position information such as the weld’s centerline, width, and groove edge. However, in the field, there will be welds with different groove forms, and the weld images will be disturbed to varying degrees by noise such as arc light, splash, and smoke. The traditional image processing methods cannot fully adapt to weld tracking in various complex environments. To overcome the noise interference in a complex welding environment and improve the accuracy and adaptability of weld tracking, a feature point extraction network based on a convolution neural network to locate weld feature points is proposed. To ensure accuracy and robustness, the network makes full use of its strong learning ability. It is not necessary to use the proposed convolution neural network to extract weld feature points in each weld image to improve welding efficiency in actual welding. The stable and predictable change of weld position can be used for weld tracking. Therefore, a reliable and fast automatic weld tracking can be realized by using the network to locate the feature point and fusing a kernel correlation filter (KCF) algorithm.

Methods To overcome noise interferences and accurately locate the groove edge of the weld, a weld feature point extraction network with the powerful feature extraction ability and learning ability of the convolution network is proposed. The network’s convolution and pooling operations can extract the position and edge contour of the laser line in the weld image. A prior region generation module is used in the network to divide the input weld image into several prior regions. It transfers the key position detection of the weld from the entire welding image to the prior regions, reducing the difficulty of extracting the weld’s key positions and improving positioning accuracy. The recognition and location module in the network can combine the prediction of the position with the prediction of the confidence of the feature point, which effectively suppresses the interference of noise and improves the anti-interference ability of the network. The training weld data set is expanded to include multiple groove weld types, which improves the network’s generalization ability and adaptability to different groove weld types. To track the weld feature points and improve welding efficiency in actual welding, the proposed network and a KCF are fused. Because the shape and position of the laser line of the weld image of adjacent frames change little, which is stable and predictable, the cyclic shift method is used in the KCF to obtain enough training samples to ensure weld tracking accuracy. Simultaneously, the fast Fourier transform is used to reduce the time complexity of the algorithm, ensuring weld tracking speed.

Results and Discussions The location results of feature points that were interfered by noises such as smoke and splash demonstrate that the weld feature point extraction network has a strong anti-interference ability (Fig.4). The weld feature points are located more accurately because the network combines the predictions of the position and confidence, which can suppress the noise interference. The location results of feature points lying in various groove types demonstrate that the network has strong adaptability and generalization ability in actual welding scenes (Fig.5). The training data set contains a variety of weld groove types, which improves the network’s robustness and generalization. Therefore, the network learns the welding characteristics of different groove types to improve adaptability. The tracking results under various noise interferences demonstrate that the proposed method can improve weld tracking accuracy (Fig.6). Furthermore, tracking results for various groove types show that the proposed method in this paper is widely applicable to multigroove welds with good generalization (Fig.8).

Conclusions In this paper, an automatic weld tracking method that combines a convolutional neural network and a correlation filter are proposed. Various degrees of noise interference during welding pose significant challenges to the accurate positioning of weld feature points. The prior region generation module in the network transfers the feature point location to the prior region, which ensures the accuracy of the feature point location. The network’s identification and location module combine position prediction and confidence prediction to suppress noise interference and improve the network’s anti-interference ability. The proposed method overcomes noise interference in complex welding environments and avoids feature point misjudgment. The fusion of a correlation filter and a network enables the automatic tracking of weld feature points. Furthermore, the correlation filter employs a fast Fourier transform to reduce the time complexity of the algorithm, ensuring welding speed. In addition, for different groove types of welds, this method which has strong adaptability can locate feature points more accurately. To summarize, the proposed method has a certain anti-interference and generalization ability that meets the actual welding requirements.