Anti-interference visual detection method and system for edge position of welding groove of double-YOLO model

Abstract

The invention discloses an anti-interference visual detection method and system for a welding groove edge position of a double-YOLO model, and the method comprises the steps: collecting n frames of welding region global images when an electric arc is located at different positions of a groove, building a YOLO-I model which outputs the welding region global images containing a plurality of welding feature position frames, the method comprises the following steps: acquiring m feature images which comprise different sizes, shapes and numbers of welding interferences and comprise groove edges, establishing a YOLO-II model which outputs the feature images of the groove edges which comprise interference area position information, generating groove left and right edge ROI images through the YOLO-I model, and identifying the groove left and right edge ROI images through the self-adaptive feature identification capability of the YOLO-II model. Welding features in a global image of a welding area or welding interference information such as welding spatter, welding fume and welding arc light in a groove edge ROI image are effectively recognized, and a groove edge line is reconstructed through a linear principal component analysis method based on Hough transform. And the influence of interference factors in the welding process on the detection precision of the edge position of the groove is obviously reduced.

Description

Technical Field

[0001] The present invention belongs to the technical field of intelligent welding, and in particular relates to a method and system for detecting the edge position of a welding groove using a dual YOLO model. Background Art

[0002] In the field of intelligent welding technology, seam tracking control accuracy is a key factor in determining welding quality and production efficiency. Accurate detection of the weld groove edge position, a core technical component of the weld seam tracking system, plays a decisive role in ensuring welding accuracy and joint quality. However, the current weld groove edge is affected by multiple factors, including groove machining accuracy, workpiece assembly errors, variable gap conditions, and dynamic gap changes during welding. This can cause the groove morphology to continuously change. This nonlinear deformation poses a challenge to accurately detecting the groove edge position. Inaccurate edge position detection will exacerbate the deviation between the welding gun and the ideal groove center position, causing fluctuations in the arc heat source's heating efficiency on the groove sidewalls, affecting the flow behavior of the molten pool, and threatening joint penetration control and weld quality stability. These technical difficulties make traditional detection methods difficult to meet the process requirements of high-precision welding. Therefore, developing precise groove edge detection technology with strong anti-interference capabilities has become a key breakthrough to improve weld tracking accuracy and promote the intelligent upgrading of welding processes.

[0003] To accurately detect the groove edge position during welding, the choice of sensing technology is crucial. Visual sensing methods, with their advantages of high detection accuracy, non-contact nature, and strong anti-interference capabilities, have attracted widespread attention. Visual sensing methods include active and passive methods. Passive visual sensing, in particular, has become an important technical tool for welding process detection and monitoring due to its simple structure, low cost, lack of advanced detection, the need for additional auxiliary light sources, good adaptability to groove configurations, and ability to synchronize arc position detection. However, during visual welding inspection, strong arc radiation can cause image sensor saturation, leading to loss of key feature information and significantly increasing the complexity of feature extraction algorithms. Randomly distributed weld spatter, which adheres irregularly to the groove edge, disrupts the geometric continuity of the target area, making groove edge detection more difficult. Dynamically changing welding smoke causes nonlinear distortion of the local grayscale distribution, increasing the accuracy of groove edge recognition. These factors severely restrict the reliability and accuracy of traditional welding groove edge detection algorithms and limit the application of narrow-gap welding processes in high-end equipment manufacturing.

[0004] In recent years, computer vision technology based on deep learning has provided new solutions for intelligent perception of the welding process. In particular, real-time target detection models such as YOLO (You Only Look Once) have demonstrated significant advantages in target recognition tasks in complex industrial scenarios due to their powerful feature extraction capabilities and efficient detection speed. Compared with traditional image processing methods, the YOLO model can better adapt to the strong interference environment of the welding process and effectively identify interference features such as spatter, smoke, and arc light in welding images. However, although the YOLO model can effectively identify interference data, it has certain limitations in data feature reconstruction and data calculation capabilities, making it difficult to fully recover and reconstruct the complete feature information of the target object.

[0005] Therefore, based on the efficient recognition capabilities of the YOLO model and the integration of a principal component analysis method based on the Hough transform, a dual-YOLO model-based weld groove edge position anti-interference detection method and system was developed. The dual-YOLO-I model generates a groove edge ROI image, and the YOLO-II model locates the interference area in the groove edge ROI image. Principal component analysis based on the Hough transform is then used to reconstruct the weld groove edge line and accurately detect the groove edge position based on the retained valid data. Applying this fused visual algorithm to weld groove edge detection is expected to overcome the technical bottlenecks of traditional methods in terms of anti-interference, detection accuracy, and real-time performance, providing more reliable visual perception capabilities for intelligent welding systems, thereby further promoting the development of automated and intelligent welding technologies. Summary of the Invention

[0006] The purpose of the present invention is to solve the key technical bottlenecks of the existing narrow gap welding groove edge position detection method, including poor environmental adaptability, insufficient anti-interference ability and other problems, and innovatively propose a welding groove edge position detection method and system based on a dual YOLO model with high detection accuracy and strong anti-interference ability.

[0007] The technical solution adopted by the anti-interference visual detection method of the welding groove edge position of the dual YOLO model of the present invention includes the following steps:

[0008] 1) Collect n frames of global images of the welding area when the arc is located at different positions on the groove, annotate each frame of the global image of the welding area and the corresponding data set of multiple welding feature position frame coordinate texts to form n data sets, and establish a YOLO-I model based on the n data sets. The output of the YOLO-I model is a global image of the welding area containing multiple welding feature position frames;

[0009] Collect m feature images containing the groove edge of welding interferences of different sizes, shapes, and quantities, and annotate them to obtain a dataset containing the feature images of the groove edge and the coordinate text of the corresponding interference area position information, forming m datasets. A YOLO-II model is established based on the m datasets. The output of the YOLO-II model is the feature image of the groove edge containing the position information of the interference area.

[0010] 2) A global image of the welding area of ​​the groove to be inspected is collected and input into the YOLO-I model to generate a ROI window containing the groove edge. The ROI images of the right and left edges of the groove are intercepted and generated through this window.

[0011] 3) Extract the feature image containing the groove edge through the groove edge ROI image processing method;

[0012] 4) The extracted feature image of the groove edge is input into the YOLO-II model to obtain the position information of the interference area. The confidence of the interference area position information is compared with the confidence threshold. When the confidence threshold is greater than the confidence threshold, the interference area position information is the interference area of ​​the welding groove edge. The original groove edge line is extracted through the edge operator, and the interference area position is mapped in situ onto the original groove edge line. The interference area data is eliminated and the groove edge line is reconstructed.

[0013] Furthermore, in step 4), when reconstructing the groove edge line, a linear principal component analysis method based on Hough transform is used to reconstruct the groove edge data groove edge line according to the retained groove edge data.

[0014] Furthermore, in step 4), if there is no interference data on the groove edge line, in step 4), the edge operator is used to extract the original groove edge data, and the extracted original groove edge data is directly reconstructed by the principal component analysis method based on Hough transform.

[0015] Furthermore, step 2) is replaced by: collecting a frame of global image of the welding area of ​​the groove to be detected, obtaining the arc contour and feature point coordinate values ​​through the arc contour extraction combination method, comparing the horizontal coordinates of the arc feature point with the image width coefficient, and obtaining the ROI images of the left and right edges of the groove.

[0016] The technical solution adopted by the dual YOLO model welding groove edge position anti-interference visual detection system of the present invention is: it includes a welding torch, a welding torch control cabinet, an infrared camera, a filtering system, an arc position generator, an image real-time transmission module, a computer image processing module and a camera fixture, the infrared camera is aligned with the center line of the welding groove and is placed at an angle to the horizontal plane, and is connected to the output end of the arc position generator and the input end of the image real-time transmission module through a video cable respectively, the input end of the arc position generator is connected to the welding torch control cabinet through a wire, and the welding torch control cabinet is connected to the arc motion driver of the narrow gap welding torch. The real-time image transmission module includes an image acquisition card and image acquisition software, which is connected to a computer image processing module through a video cable, and a filtering system is coaxially arranged at the front end of the infrared camera; the welding wire passes through the welding torch and extends into the welding groove to generate an arc, the arc position generator collects the position signal of the arc, and the filtering system is coaxially installed at the front end of the infrared camera. After receiving the arc position signal, the infrared camera collects a frame of global image of the welding area where the arc is located on the left wall of the groove, the right wall of the groove, or the center of the groove, as well as a feature image including the groove edge that contains welding interference of different sizes, shapes, and quantities.

[0017] Compared with the prior art, the advantages and beneficial effects of the present invention are mainly:

[0018] 1) The arc feature point position information is fused through the arc contour extraction combination method, or the YOLO-I model is used to generate the left and right edge ROI images of the groove. The left and right edge lines of the groove are extracted simultaneously or in a time-sharing manner through the groove edge ROI image processing method. The adaptive feature recognition capability of the YOLO-II model is used to effectively identify the welding features in the global image of the welding area or the welding interference information such as welding spatter, welding smoke and welding arc in the groove edge ROI image. The groove edge line is reconstructed through the linear principal component analysis method based on Hough transform, which significantly reduces the influence of interference factors in the welding process on the accuracy of groove edge position detection.

[0019] 2) A passive visual image acquisition system based on arc position was established. By developing an arc position generator, a global image of the welding area with the arc located at the center or side wall of the groove was acquired. This achieved synchronous detection of both edges of the groove in a single-frame welding image, increasing the inference speed by 50%, and effectively resolving the technical problem of the traditional visual sensing system being out of sync with the dynamic process of the welding arc.

[0020] 3) Using dual YOLO models, an intelligent recognition system for weld spatter interference was constructed, effectively filtering out interference signals, which account for up to 70% of weld spatter. The YOLO-I model simultaneously generates local feature images within the global image of the weld area. The YOLO-II model identifies and locates interference information within the feature images and, through principal component analysis based on the Hough transform, reconstructs the groove edge line and calculates its position, significantly improving the system's engineering practicality and environmental adaptability.

[0021] 4) The system innovatively integrates arc position information with visual images through multimodal data fusion to establish a groove edge recognition model with strong anti-interference capabilities. Experiments show that even under harsh working conditions such as strong arc light, large spatter, and dense smoke and dust, the system can still accurately identify welding interference data and precisely locate and detect it, significantly improving the system's reliability and stability in complex industrial environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] Figure 1 2 is a block diagram of the anti-interference visual detection system for the welding groove edge position based on the dual YOLO model of the present invention;

[0023] Figure 2 To adopt Figure 1 The schematic diagram of the global image of the welding area where the arc is located on the right side of the groove (taking the I-type groove as an example) collected by the system.

[0024] Figure 3 To adopt Figure 1 Schematic diagram of the global image of the welding area captured by the system with the arc at the midpoint (taking I-groove as an example);

[0025] Figure 4 This is a flowchart of the anti-interference visual detection method for the welding groove edge position using the dual YOLO model of the present invention;

[0026] Figure 5 Figure 1 is an example of the visual sensing detection result of the time-sharing detection of the welding groove edge position based on the YOLO-Ⅱ model; (a) and (b) are the global images of the welding area when the arc is located on the left and right sides of the groove, respectively; (c) and (d) are the adaptively positioned right and left edge ROI windows of the groove, respectively; (e) and (f) are the binary images containing the extracted right and left edge ROIs of the groove, respectively; (g) and (h) are the regional positioning results of the welding spatter interference on the right and left edges of the groove based on the YOLO model, respectively; (i) and (j) are the in-situ mapping of the regional positioning results to the original right and left edge lines of the groove, respectively; (k) and (L) are the right and left edges of the groove after removing abnormal data, respectively; (m) and (n) are the right and left edges of the groove reconstructed using the linear principal component analysis method based on Hough transform, respectively;

[0027] Figure 6 This is an example of the result of simultaneously locating the global image features of the welding area based on the YOLO-Ⅰ model. DETAILED DESCRIPTION

[0028] The anti-interference visual detection system for the welding groove edge position of the dual YOLO model of the present invention is as follows Figure 1 As shown, the apparatus primarily comprises a welding torch, a welding torch control cabinet 14, an infrared camera 8, a filtering system 7, an arc position generator 9, a real-time image transmission module 10, a computer image processing module 11, and a camera fixture 12. The infrared CMOS camera 8 is aligned with the centerline 13 of the welding groove 6 and positioned at an angle θ with the horizontal plane. The infrared CMOS camera 8 is fixed to the welding torch via the fixture 12 and connected to the output of the arc position generator 9 and the input of the real-time image transmission module 10 via video cables. The input of the arc position generator 9 is connected to the welding torch control cabinet 14 via a wire, and the welding torch control cabinet 14 is connected to the arc motion driver 3b of the narrow gap welding torch. The real-time image transmission module 10 includes an image acquisition card and image acquisition software and is connected to the computer image processing module 11 via a video cable. The filtering system 7 comprises a filter, a neutral density filter, and a UV filter.

[0029] The welding wire 3d passes through the welding torch and extends into the welding groove 6 to generate the arc 2. The arc position generator 9 collects the position signal P of the arc 2. arc Filter system 7 is coaxially mounted on the front end of the infrared camera 8, the infrared camera 8 receives the arc position signal P 2 arc After that, a frame of global image 15 of the welding area where the arc is located at the left wall of the groove, the right wall of the groove, or the center of the groove can be collected. Figure 2 , and sent to the computer image processing module 11 through the image real-time transmission module 10.

[0030] The height and width of the collected global image 15 of the welding area are h1 and w1 respectively, and it includes the welding arc 2, welding wire 3d, groove left edge 4, and groove right edge 5. Among them, the arc position signal P generated by the arc position generator 9 is arc The arc 2 position in the global image 15 of the welding area is determined to be at the right wall or midpoint of the groove.

[0031] See also Figure 4 The steps of the anti-interference visual detection method for the welding groove edge position of the dual YOLO model of the present invention are as follows:

[0032] 1. Establish dual YOLO models, namely YOLO-Ⅰ and YOLO-Ⅱ models.

[0033] use Figure 1The system shown in FIG. 1 collects n frames of global images of the welding area when the arc is located at different positions of the groove through the infrared camera 8, such as Figure 2 and Figure 3 The global image 15 of the welding area shown is inputted, and n frames of the global image 15 of the welding area are inputted into the computer image processing module 11. Then, in the computer image processing module 11, the inputted n frames of the global image 15 of the welding area, with the arc at different positions on the groove, are inputted into the annotation software. The annotation software completes the annotation of the corresponding multiple welding feature position frames Z1 in each frame of the global image 15 of the welding area, and obtains a dataset of each frame of the global image 15 of the welding area and the corresponding multiple welding feature position frames Z1 coordinate text, generating a total of n datasets. The datasets of the n datasets, accounting for 70%, 20%, and 10%, are divided into three parts: training sample set, validation sample set, and test sample set. These are inputted into the YOLO program, and the output is the global image of the welding area containing multiple welding feature position frames Z1, obtaining the established YOLO-I.

[0034] Similarly, using Figure 1 The illustrated system uses an infrared camera 8 to capture m feature images I1 containing weld interferences of varying sizes, shapes, and quantities. Feature images I1 containing weld interferences can also be obtained by simulating different weld spatter test methods. These m feature images I1 containing weld interferences are then input into a computer image processing module 11. In this module, the m feature images I1 containing weld interferences are then input into annotation software, where the weld interference locations are annotated within each feature image I1. After each feature image I1 containing weld interferences is processed through standard software, a dataset containing the feature image I1 containing weld interferences and the corresponding Z2 coordinate information is obtained, resulting in a total of m datasets. The m datasets, representing 70%, 20%, and 10% of the total, are divided into a training set, a validation set, and a test set. These datasets are then input into a YOLO program, which outputs a feature image I2 containing the weld interference location information, thus establishing YOLO-II.

[0035] See also Figure 6 , is a YOLO-Ⅰ model established using n=40 544×544 pixel global images of the welding area 15 as a data set. Figure 6 The global image 15 of the welding area with the arc located at the center of the groove is input into the established YOLO-Ⅰ model, and the output is the 1# red frame containing the arc, the 3# orange frame containing the welding wire 3d, and the two 4# yellow frames containing the left edge and the right edge of the groove. The confidence levels of these four welding feature position frames Z1 are 0.85, 0.84, 0.75 and 0.79 respectively. They are compared with the confidence threshold Td For comparison, when the confidence of the detection results is greater than the confidence threshold T d When , the identified welding feature position frame Z1 is considered to be the valid area. Figure 6 In the above example, the confidence threshold T d Set to 0.7, Figure 6 The confidence levels of the detection results in are all greater than the confidence threshold of 0.7, so the located welding feature position box Z1 is a valid area.

[0036] The established YOLO-Ⅰ and YOLO-Ⅱ models are embedded in the computer image processing module 11 to form a detection program based on the YOLO model. When the global area welding image 15 or the feature image I1 containing the groove edge is input into the detection program, multiple welding features in the global area welding image 15 or the feature image I2 of the groove edge containing the interference area position information can be identified.

[0037] 2. Generate groove edge ROI image. Figure 1 The system shown captures a global image 15 of the welding area of ​​the groove to be inspected, and separates the arc from the background through the arc contour extraction combination method to obtain a global binary image containing the arc contour. The arc contour extraction combination method includes image processing algorithms such as image filtering, image thresholding, and morphological operations. By traversing the pixel points of the arc contour in the global binary image, the coordinates of the highest point in the arc contour height direction are found as the arc feature point. Figure 3 As shown, the width w1 and height h1 of the global image of the welding area are both 544 pixels, for example Figure 5 As shown in Figures (a) and (b). After median filtering and threshold processing with a global threshold of 240, and fusion of morphological opening operations, a global binary image containing the arc contour is formed. By searching the highest point of the arc contour in the global binary image row by row, the coordinates of the highest point of the arc contour when the arc is located on the left and right sides of the groove are (156, 189) pixels and (396, 228) pixels respectively. Figure 5 As shown in Figures (c) and (d).

[0038] Compare the arc feature point's x-coordinate with the image width factor, which is defined as half the global image width of the welding area, i.e., w1 / 2. Then dynamically determine the x-coordinate of the vertex of the groove edge ROI window. When the arc feature point's x-coordinate is less than w1 / 2, it indicates that the arc is located on the left side wall 1a of the welding groove. In this case, it is necessary to set the groove right edge ROI window 16b, as shown in the following example: Figure 5 As shown in Figure (c), the x coordinate of the highest point of the arc is 156, which is less than (w1 / 2) = 274. Therefore, the ROI window is set on the right wall of the groove, as shown in Figure 3. Figure 5As shown in the white box in Figure (c), the height h2 and width w2 of the ROI window are both 100 pixels. When the x-coordinate of the arc feature point is greater than w1 / 2, it means that the arc is located on the right side wall 1b of the welding groove. At this time, the ROI window 16a on the left edge of the groove needs to be set, for example Figure 5 As shown in Figure (d), the x-coordinate of the highest point of the arc is 396, which is greater than (w1 / 2) = 274. Therefore, the ROI window is set on the left wall of the groove. Based on the obtained groove edge ROI window, the right and left edges of the groove are intercepted and generated.

[0039] Alternatively, the YOLO-I model established above can be used to input a global image 15 of the welding area of ​​the groove to be detected into the established YOLO-I model to directly generate a ROI window containing the groove edge, for example: Figure 6 The two yellow windows (#4) are used to capture and generate ROI images of the right and left edges of the groove. You can choose to use either the established YOLO-Ⅰ model to generate ROI images of the right and left edges of the groove or the arc contour extraction combined method described above to generate ROI images of the right and left edges of the groove, and they can be replaced with each other.

[0040] 3. Extract the feature image containing the groove edge. Based on the generated groove left edge ROI image and groove right edge ROI image, extract the feature image containing the groove edge through the groove edge ROI image processing method. The groove edge ROI image processing method includes image filtering for denoising, morphological operation, and Otsu method adaptive threshold processing method. For example Figure 5 In Figures (e) and (f), the groove edge binarization images are obtained after image mean filtering and Otsu threshold segmentation. Due to the presence of weld spatter, the right edge binarization image of the groove is convex to the left, and the left edge binarization image of the groove is convex to the right. In this case, it is necessary to eliminate the influence of weld spatter on the groove edge line extraction accuracy to improve the detection accuracy of the groove edge position.

[0041] 4. Use the established YOLO-Ⅱ model to identify whether there is welding interference in the feature image. The data set of the established YOLO-Ⅱ model is m feature images I1 of the groove edge containing interference of different sizes, shapes, and quantities collected by system 1. It can also be obtained by simulating different forms of welding spatter test methods. The feature image I1 of the groove edge can be a groove edge binary image or a groove edge image. For example Figure 5The characteristic image I1 of the groove edge shown in Figures (e) and (f) is taken as an example where the characteristic image I1 containing the groove edge is a binary image of the groove edge. The binary image of the extracted groove edge is input into the established YOLO-II model to obtain the position information Z2 of the interference region. For example, Figure 5 as shown by the green boxes in Figures (g) and (h), where the green boxes contain 53 and 49 data respectively, and the confidence levels of the position information Z2 of the interference region are 0.75 and 0.8. Compare the confidence level of the obtained position information Z2 of the interference region with the confidence level threshold T d When the confidence level of the position information Z2 of the interference region is greater than the confidence level threshold T d it indicates that the position information Z2 of this interference region is the interference region of the welding groove edge, and the height of this interference region is h3, as Figure 5 shown by the green box in Figure (g). According to the binary image of the groove edge, for example, Figure 5 in Figures (e) and (f), the original groove edge lines in Figures (e) and (f) are extracted through an edge operator. For example, Figure 5 as shown by the white lines in Figures (i) and (j), the position information Z2 of the interference region with a height of h1 is mapped in-situ to the original groove edge line, as shown by the green boxes in Figures (i) and (j) for example, Figure 5 and h3 (1 < h3 < h2) groove edge data included in the position information Z2 of the interference region are excluded. For example, Figure 5 Figures (k) and (L) are the groove edge lines retained after excluding the interference data on the original groove edge line. Otherwise, the data on the original groove edge line are retained.

[0042] 5. Reconstruct the groove edge line. After being recognized by the established YOLO-II model, if there is interference on the original groove edge line, the interference data on the original groove edge line are excluded through the established YOLO-II. According to the retained (h2 - h3) groove edge data, a linear principal component analysis method based on the Hough transform is used to reconstruct the groove edge line containing h2 groove edge data. If there is no interference data on the groove edge line, directly use the principal component analysis method based on the Hough transform to reconstruct the original groove edge data extracted by the edge operator. The reconstructed groove edge line contains h2 groove edge data.

[0043] For example, Figure 5 the red lines in Figures (m) and (n) are the groove edge lines reconstructed by the established YOLO-II model. Compared with the original groove edge lines extracted by the existing image processing, the influence of welding spatter on the extraction accuracy of the groove edge line is excluded, and the detection accuracy of the groove edge position is improved, laying a technical foundation for intelligent welding.

Claims

1. A dual YOLO model anti-interference visual detection method for welding groove edge position, characterized by The following steps are involved: 1) Collect n frames of global images of the welding area when the arc is located at different positions on the groove, annotate each frame of the global image of the welding area and the corresponding data set of multiple welding feature position frame coordinate texts to form n data sets, and establish a YOLO-I model based on the n data sets. The output of the YOLO-I model is a global image of the welding area containing multiple welding feature position frames; Collect m feature images containing the groove edge of welding interferences of different sizes, shapes, and quantities, and annotate them to obtain a dataset containing the feature images of the groove edge and the coordinate text of the corresponding interference area position information, forming m datasets. A YOLO-II model is established based on the m datasets. The output of the YOLO-II model is the feature image of the groove edge containing the position information of the interference area. 2) A global image of the welding area of ​​the groove to be inspected is collected and input into the YOLO-I model to generate a ROI window containing the groove edge. The ROI images of the right and left edges of the groove are intercepted and generated through this window. 3) Extract the feature image containing the groove edge through the groove edge ROI image processing method; 4) The extracted feature image of the groove edge is input into the YOLO-II model to obtain the position information of the interference area. The confidence of the interference area position information is compared with the confidence threshold. When the confidence threshold is greater than the confidence threshold, the interference area position information is the interference area of ​​the welding groove edge. The original groove edge line is extracted through the edge operator, and the interference area position is mapped in situ onto the original groove edge line. The interference area data is eliminated and the groove edge line is reconstructed.

2. The anti-interference visual detection method for welding groove edge position using a dual YOLO model according to claim 1 is characterized in that: In step 4), when reconstructing the groove edge line, the groove edge data groove edge line is reconstructed using a linear principal component analysis method based on Hough transform according to the retained groove edge data.

3. The anti-interference visual detection method for welding groove edge position using a dual YOLO model according to claim 1 is characterized in that: In step 4), if there is no interference data on the groove edge line, in step 4), the edge operator is used to extract the original groove edge data, and the extracted original groove edge data is directly reconstructed by the principal component analysis method based on Hough transform.

4. The anti-interference visual detection method for welding groove edge position using a dual YOLO model according to claim 1 is characterized in that: Step 2) is replaced by: collecting a global image of the welding area of ​​the groove to be inspected, obtaining the arc contour and feature point coordinates through the arc contour extraction combination method, comparing the lateral coordinates of the arc feature points with the image width coefficient, and obtaining the ROI images of the left and right edges of the groove.

5. The anti-interference visual detection method for welding groove edge position using a dual YOLO model according to claim 4 is characterized in that: The horizontal coordinate of the arc feature point is compared with the image width coefficient. When the horizontal coordinate of the arc feature point is less than the image width coefficient, it means that the arc is located on the left side of the welding groove. Then the groove right edge ROI window is set. Otherwise, the groove left edge ROI window is set. According to the obtained groove edge ROI window, the groove right edge and groove left edge ROI images are intercepted and generated.

6. The anti-interference visual detection method for welding groove edge position using a dual YOLO model according to claim 1, characterized in that: In step 2), the confidence of the welding feature position box in the ROI window output by the YOLO-Ⅰ model is compared with the confidence threshold. When both are greater than the confidence threshold, the identified welding feature position box is a valid area.

7. The anti-interference visual detection method for welding groove edge position using a dual YOLO model according to claim 4 is characterized in that: The arc contour extraction combined method includes image filtering, image thresholding, and morphological operation image processing algorithms to separate the arc from the background and obtain a global binary image containing the arc contour. By traversing the pixel points of the arc contour in the global binary image, the coordinates of the highest point in the arc contour height direction are found as the arc feature point.

8. The anti-interference visual detection method for welding groove edge position using a dual YOLO model according to claim 1, characterized in that: The groove edge ROI image processing method includes image filtering for denoising, morphological operation, and Otsu method adaptive threshold processing method to extract a feature image containing the groove edge.

9. The method for anti-interference visual detection of welding groove edge position using a dual YOLO model according to claim 1, characterized in that: The datasets accounting for 70%, 20% and 10% of the n datasets and m datasets are divided into three parts: training sample set, verification sample set and test sample set, and input them into the YOLO program respectively.

10. A system for implementing the anti-interference visual detection method according to any one of claims 1 to 9, characterized in that: The invention comprises a welding torch, a welding torch control cabinet, an infrared camera, a filtering system, an arc position generator, an image real-time transmission module, a computer image processing module and a camera fixture. The infrared camera is aligned with the center line of the welding groove and is arranged at an angle to the horizontal plane. The infrared camera is respectively connected to the output end of the arc position generator and the input end of the image real-time transmission module via a video cable. The input end of the arc position generator is connected to the welding torch control cabinet via a wire. The welding torch control cabinet is connected to the arc motion driver of the narrow gap welding torch. The image real-time transmission module comprises an image acquisition card and image acquisition software and is connected to the computer image processing module via a video cable. The filtering system is coaxially arranged at the front end of the infrared camera. The welding wire passes through the welding torch and extends into the welding groove to generate an arc. The arc position generator collects the position signal of the arc. The filtering system is coaxially installed at the front end of the infrared camera. After receiving the arc position signal, the infrared camera collects a frame of a global image of the welding area in which the arc is located on the left wall or the right wall of the groove, or the center of the groove, as well as a feature image including the groove edge and containing welding interference of different sizes, shapes and quantities.

Images