Weld joint identification method based on visual inspection

Abstract

The invention discloses a weld joint identification method based on visual inspection. The method comprises an image acquisition module, a data preprocessing module, a deep learning detection module, an image segmentation and contour extraction module, a weld joint parameter calculation module, a result visualization module, a real-time optimization module and the like. According to the method, the detection precision is improved, the generalization ability of the model in a complex scene is remarkably improved, the adaptability is enhanced, the real-time performance of the system is remarkably improved by optimizing an algorithm structure and a hardware acceleration scheme, a comprehensive evaluation function is added, and multi-dimensional evaluation indexes are provided for the welding seam quality; in addition, the method can also be applied to welding quality control in industrial production, helps enterprises to reduce the rework rate and material waste, and is expected to save 10%-15% of production cost for related welding industries. The objective and reliable welding seam quality evaluation method is provided, the fairness and accuracy of special operation examinations are improved, and the skill level and work safety of special operators can be improved.

Description

Technical Field

[0001] The present invention relates to the field of welding technology, and in particular to a weld identification method based on visual detection. Background Art

[0002] The special operation operation qualification examination usually includes two parts: theoretical knowledge assessment and practical operation evaluation. In the evaluation of welding subjects, the conventional manual scoring method is easily affected by subjective judgment, external interference (such as welding smoke, strong light), inconsistent scoring standards and other factors, resulting in the lack of objectivity and accuracy in the evaluation results.

[0003] At present, weld inspection methods based on computer vision are gradually being applied to special operation examinations. Most of these methods use target detection and classification algorithms (such as YOLO, Faster R-CNN, etc.) to analyze weld quality. However, the existing technologies still have the following problems in practical applications: 1. Insufficient detection accuracy: Existing algorithms often make false detections or missed detections when dealing with small target detection and complex background interference in specific welding scenarios. Especially when there are subtle defects such as pores and slag inclusions on the weld surface, it is difficult for the model to accurately capture these details.

[0004] 2. Limited adaptability: Since existing models are usually trained based on general datasets (such as COCO or VOC), their characteristics are quite different from the actual scenes of welding operations. The insufficiency and mismatch of training data lead to poor performance of the model in the actual test environment.

[0005] 3. Insufficient real-time performance: Quality assessment in the welding process requires a quick response, but some existing algorithms are limited by computational complexity and it is difficult to achieve real-time reasoning on resource-constrained edge devices.

[0006] 4. Lack of targeted optimization: Existing technologies lack specific optimization strategies for weld detection and quality assessment, such as how to accurately identify weld length, width differences and other subtle features. Summary of the invention

[0007] In order to solve the problems in the above-mentioned background technology, defects in welds can be detected more accurately, and rapid response capabilities can be achieved to adapt to the complex scene requirements of special operations. A weld recognition method that can combine high-precision target detection, image segmentation and intelligent analysis is needed for comprehensive detection and evaluation of weld quality. Therefore, the present invention provides a weld recognition method based on visual detection.

[0008] The specific technical solution of the present invention is: a weld identification method based on visual detection, comprising the following steps: S1. Capturing high-resolution images of the welding process by an industrial camera, wherein the industrial camera uses automatic exposure control, denoising filtering and image stabilization algorithms; S2. preprocessing the collected high-resolution images, wherein the preprocessing includes unifying the image resolution and adjusting it to 1920×1080 pixels, and then using denoising filtering and histogram equalization technology to enhance the contrast of the weld area in the image, and removing duplicate or low-quality image data. The preprocessed data is used as input for subsequent model training and inference; S3. Build a deep learning target detection model to identify key defects in weld images based on the YOLO series of algorithms. The defects include pores, slag inclusions, and unformed areas. The detection model is trained with labeled data, and a multi-scale feature extraction mechanism is introduced during the training process to improve the detection capability of small target defects. S4. Analyze the detected defect positions and use image segmentation technology to extract the weld contour to provide basic data for subsequent parameter calculations to ensure accurate positioning of the weld area; S5. Weld parameter calculation: Through the weld profile in S4, the key parameters of the weld are automatically calculated, including weld length L, weld width W, and weld difference ΔW. Through statistics and analysis of the cross-sectional data of the weld, the comprehensive score S of the weld quality is quantitatively evaluated; the weld length L is calculated by the arc length of the weld profile curve, and the discrete point method or curve fitting method is used. The weld profile is represented by n discrete points, and the point coordinates are (x i， y i ), the total length of the weld is calculated as follows,

[0009] This formula calculates the total length of the weld based on the cumulative Euclidean distance between two points. When L ≥ the set standard value, the weld meets the weld standard and is complete. When L < the set standard value, the weld is not fully welded or has poor fusion. The weld width W is the width value measured in the cross-sectional direction of the weld. It is calculated using the equidistant scanning method and defines the weld width points (x l，i ,y l，i ) and (x r，i ,y r，i ), the average width of the weld is calculated as follows,

[0010] Among them, x l，i , x r，i are the x-coordinates of the left and right edges of the weld, n is the number of sampling points in the weld cross section, W std is the standard weld width, W std The value is 5mm~8mm, W avg<W min , weld shrinkage or insufficient penetration occurs, W avg >W max , the weld is too wide, affecting the structural strength; The weld difference ΔW is measured by calculating the difference between the maximum width and the minimum width. The calculation formula is as follows:

[0011] Among them, W max =max(W1, W2, W3, W4, W5..., W n ), W min =min(W1, W2, W3, W4, W5..., W n ), ΔW<allowable range, the weld width is evenly distributed and the welding quality is good; ΔW>allowable range, the weld is uneven, resulting in stress concentration and affecting the structural strength; The comprehensive weld quality score S is weightedly calculated based on the weld length L, the weld width W, and the weld difference ΔW.

[0012] α1, α2, α3 are weighted coefficients, satisfying α1+α2+α3=1, L std , W std , ΔW max Standard welding parameters: S≥0.9, excellent quality, good weld quality, and parameters meet the standards; 0.8≤S<0.9, qualified quality, and overall quality requirements are met; 0.6≤S<0.8, there may be weld defects, and inspection is recommended; S<0.6, unqualified quality, requiring repair or re-welding; S6. Generate a quality report to provide comprehensive support for welding quality evaluation and assessment. The quality report includes test results, evaluation data and visual images. The test results are presented in a graphical manner, and the defect location, category and related parameters such as weld length and width are marked on the original image.

[0013] Furthermore, the present invention combines CUDA parallel computing technology and lightweight network optimization strategy in model deployment, and adopts multi-threaded design to achieve efficient collaboration of image acquisition, processing and result display, and optimizes hardware resource utilization to ensure that the system can respond quickly in resource-constrained scenarios.

[0014] Furthermore, in step S4, a lightweight target detection model or a geometric fitting algorithm is used to replace some segmentation steps to adapt to different hardware configurations and usage requirements. Furthermore, lightweight object detection models include YOLOv4-tiny.

[0015] The present invention can accurately identify subtle defects in welds by designing a deep learning model optimized for welding scenarios and combining it with a multi-scale feature extraction mechanism. In complex welding scenarios, the defect detection accuracy of the present invention reaches more than 95%, which is 10%-15% higher than that of existing technologies (such as YOLO, Faster R-CNN, etc.), especially in small target defect detection, with the false detection rate and missed detection rate reduced by 8% and 12% respectively; the present invention combines domain knowledge and large-scale welding data sets for model training, which significantly improves the generalization ability of the model in complex scenarios. In practical applications, the model can adapt to different welding processes and a variety of welding materials, and maintain stable detection performance under different lighting conditions and background interference. The average detection accuracy of the model in diversified welding scenarios exceeds 92%, which is more than 20% higher than that of general target detection models (such as models trained based on COCO data sets); the present invention significantly improves the real-time performance of the system by optimizing the algorithm structure and hardware acceleration solutions (such as CUDA parallel computing technology and lightweight network design), and can be used on edge devices (such as NVIDIA Jetson series), the average inference time of the system is reduced to less than 50 milliseconds, which can meet the needs of real-time quality assessment in the welding process. Compared with the existing technology, the inference speed is improved by 30%-40%, and the hardware resource occupancy rate is reduced by 25%. In addition, the present invention can not only detect the length and width of the weld, but also evaluate the uniformity of the weld width, count the maximum width, minimum width and width difference, and provide multi-dimensional evaluation indicators for the weld quality. The weld contour extraction accuracy of the present invention reaches more than 98%, and the width measurement error is controlled within ±1 mm, which is significantly better than the traditional manual measurement method (the error is usually ±5 mm).

[0016] Therefore, the present invention has achieved remarkable technical effects in terms of detection accuracy, adaptability, real-time performance, and evaluation function, and provides an efficient and reliable solution for special operation practical assessment and welding quality control, and has broad application prospects and important social and economic benefits. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 A schematic flow chart of a weld identification method provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0018] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention.

[0019] This embodiment uses a welding production workshop as the application scenario. There are multiple fixed industrial cameras in the workshop to collect real-time images and video data during the welding process. In order to ensure the comprehensiveness and real-time nature of data collection, all cameras use industrial cameras of the same model and resolution, and uniformly use edge computing terminals (such as those based on NVIDIA Jetson) for data collection and preliminary processing.

[0020] In the present invention, the system as a whole includes modules such as image acquisition, data preprocessing, deep learning detection, image segmentation and contour extraction, weld parameter calculation, result visualization and real-time optimization.

[0021] like Figure 1 FIG. 1 is a flow chart of a weld seam recognition method provided by an embodiment of the present invention. The present invention provides a weld seam recognition method based on visual detection, and the method comprises the following steps: S1. Continuous video data of the welding process is collected through industrial cameras. The video data is extracted into static images in PNG format at preset time intervals (for example, every 1000 milliseconds) using frame extraction technology to ensure image quality. Automatic exposure control, image stabilization and illumination compensation technologies are used during the collection process to reduce the impact of ambient light fluctuations and camera vibrations. The collected image data is stored in corresponding folders according to the camera number for subsequent unified management and processing.

[0022] S2. The collected high-resolution images are preprocessed, and the preprocessing includes unifying the image resolution, adjusting it to 1920×1080 pixels, and then using Gaussian filtering to denoise the image, using histogram equalization technology to enhance the contrast of the weld area in the image, and removing blurry or repeated frames through an image clarity evaluation algorithm (such as using the Laplace operator to calculate the variance of the image). The preprocessed image data is imported into the image annotation platform, and the welds and their defects (such as pores, slag inclusions, and unformed areas) in the image are annotated in combination with manual experience. The annotation results generate corresponding annotation files in XML or JSON format, and are organized into training data sets in VOC or COCO format.

[0023] S3. Build a deep learning target detection model, select YOLOv5x as the basic target detection model, train the model through the data set, introduce data enhancement (such as rotation, scaling, flipping, etc.) during the training process to increase data diversity, and adopt a multi-scale training strategy to ensure that the model can maintain high detection accuracy at different resolutions. In order to further improve the recognition ability of small defects, the defects include pores, slag inclusions and unformed areas. The detection model is trained by annotated data, and a multi-scale feature extraction mechanism is introduced into the model architecture. Feature branches with different resolutions (for example, feature maps of 120×120, 60×60, and 30×30, respectively) are used for feature fusion. Cross-validation and early stopping strategies are used during the training process. Finally, a model with an accuracy of more than 95% in the weld defect detection task is obtained, and the trained model is saved as a .pt file, which is subsequently converted to a format supported by ONNX and the target platform.

[0024] S4. Deploy the trained deep learning model to the edge computing terminal, receive the real-time collected image data for reasoning, perform target detection on each frame of the image, and the model automatically outputs the bounding box and category information of the defect area. For the detected weld area, use the image segmentation algorithm (such as the segmentation network based on yolov8-seg or directly use the mask generated in the model) to extract the weld contour to provide accurate regional information for subsequent parameter calculation.

[0025] S5. Using the weld contour obtained by segmentation, the minimum circumscribed rectangle algorithm is used to calculate the weld length L and weld width W. At the same time, the maximum and minimum values ​​of the cross-sectional width in the local area of ​​the weld are counted, and then the weld difference ΔW is calculated. The calculated parameters such as weld length L, weld width W and weld difference ΔW and the detection results (defect category, confidence, etc.) are organized into structured data, and the weld quality comprehensive score S is quantitatively evaluated. A test report is generated, and the detection frame, contour and parameter data are superimposed on the original image through image processing libraries such as OpenCV to achieve intuitive visualization of the results; The weld length L is calculated by the arc length of the weld contour curve. The weld contour is represented by n discrete points using the discrete point method or curve fitting method. The point coordinates are (x i， y i ), the total length of the weld is calculated as follows,

[0026] The formula calculates the total length of the weld based on the cumulative Euclidean distance between two points. When L ≥ the set standard value, the weld meets the standard and the weld is complete. When L < the set standard value, there may be problems with incomplete weld or poor fusion. The standard value is a fixed value. For example, if the length of the welding plate is 15 cm, after the welding of the welding plate, the weld length is greater than or equal to 10 cm, which is considered to be in compliance, and 10 is the set standard value. The weld width W is the width value measured in the cross-sectional direction of the weld. It is calculated using the equidistant scanning method and defines the weld width points (x l，i ,y l，i ) and (x r，i ,y r，i ), the average width of the weld is calculated as follows,

[0027] Among them, x l，i , x r，i are the x-coordinates of the left and right edges of the weld, n is the number of sampling points in the weld cross section, W std is the standard weld width, W std The value is 5mm~8mm. If W avg <W min , weld shrinkage or insufficient penetration may occur. avg >W max , the weld may be too wide, affecting the structural strength; The weld difference ΔW is measured by calculating the difference between the maximum width and the minimum width. The calculation formula is as follows:

[0028] Among them, the maximum weld width W max =max(W1, W2, W3, W4, W5..., W n ), minimum weld width W min =min(W1, W2, W3, W4, W5..., W n ), if ΔW is less than the allowable range, the weld width is evenly distributed and the welding quality is good; if ΔW is greater than the allowable range, the weld is uneven, which leads to stress concentration and affects the structural strength; The comprehensive weld quality score S is weightedly calculated based on the weld length L, the weld width W, and the weld difference ΔW.

[0029] α1, α2, α3 are weighted coefficients, satisfying α1+α2+α3=1 (e.g. α1=0.4, α2=0.3, α3=0.3), L std , W std , ΔW maxis the standard welding parameter, ΔW max 3mm; when S≥0.9, the welding quality is excellent, the weld quality is good, and the parameters meet the standards; when 0.8≤S<0.9, the welding quality is qualified, and the weld as a whole meets the quality requirements; when 0.6≤S<0.8, the welding quality needs to be adjusted, there may be weld defects, and inspection is recommended; when S<0.6, the welding quality is unqualified and needs to be repaired or re-welded.

[0030] The weld length L reflects whether the welding is complete, the weld width W reflects the degree of fusion and weld stability, and the weld difference ΔW reflects the welding uniformity. If it is too large, it will affect the welding strength. The comprehensive score S combines multiple indicators to provide a comprehensive weld quality assessment.

[0031] S6. Generate a quality report to provide comprehensive support for welding quality evaluation and assessment. The quality report includes test results, evaluation data and visual images. The test results are presented in a graphical manner, and the defect location, category and related parameters such as weld length and width are marked on the original image.

[0032] In addition, the present invention combines CUDA parallel computing technology and lightweight network optimization strategy in model deployment, and adopts multi-threaded design to achieve efficient collaboration of image acquisition, processing and result display, and optimizes hardware resource utilization to ensure that the system can respond quickly in resource-constrained scenarios.

[0033] As another embodiment of the present invention, in step S4, a lightweight target detection model (such as YOLOv4-tiny) or a geometric fitting algorithm may be used to replace part of the segmentation steps to adapt to different hardware configurations and usage requirements.

[0034] In addition, the weld identification method of the present invention can also be applied to welding quality control in industrial production, helping enterprises reduce rework rates and material waste, and is expected to save 10%-15% of production costs for welding-related industries.

[0035] The above is a schematic description of the present invention and its implementation methods, which is not restrictive. The accompanying drawings show only one implementation method of the present invention, and the actual structure is not limited thereto. Therefore, if a person skilled in the art is inspired by it and, without departing from the purpose of the invention, designs a structural method and an embodiment similar to the technical solution without creativity, they all belong to the protection scope of the present invention.

Claims

1. A weld identification method based on visual inspection, characterized in that: The identification method comprises the following steps: S1. Capturing high-resolution images of the welding process by an industrial camera, wherein the industrial camera uses automatic exposure control, denoising filtering and image stabilization algorithms; S2. Preprocessing the collected high-resolution image, wherein the preprocessing includes unifying the image resolution, using denoising filtering and histogram equalization technology to enhance the contrast of the weld area in the image, and removing duplicate or low-quality image data; S3. Build a deep learning target detection model to identify key defects in weld images based on the YOLO series algorithm. The detection model is trained with labeled data, and a multi-scale feature extraction mechanism is introduced during the training process to improve the detection capability of small target defects. S4. Analyze the detected defect positions and extract the weld contour using image segmentation technology; S5. Weld parameter calculation: through the weld profile in S4, the key parameters of the weld are automatically calculated, including the weld length L, weld width W, and weld difference ΔW. Through statistics and analysis of the cross-sectional data of the weld, the comprehensive score S of the weld quality is quantitatively evaluated; S6. Generate quality reports to provide comprehensive support for welding quality evaluation and assessment.

2. The weld seam recognition method based on visual inspection according to claim 1, characterized in that: The unified resolution of the S2 is 1920×1080 pixels.

3. The weld seam recognition method based on visual inspection according to claim 1, characterized in that: The key defects in S3 include pores, slag inclusions and unformed areas.

4. The weld seam recognition method based on visual inspection according to claim 1, characterized in that: The weld length L in S5 is calculated by the arc length of the weld contour curve, using the discrete point method or curve fitting method. The weld contour is represented by n discrete points, and the point coordinates are (x i， y i ), the total length of the weld is calculated as follows: ； This formula calculates the total length of the weld based on the cumulative Euclidean distance between two points. When L ≥ the set standard value, the weld meets the weld standard and is complete. When L < the set standard value, the weld is not fully welded or has poor fusion.

5. The weld seam recognition method based on visual inspection according to claim 1, characterized in that: The weld width W in S5 is the width value measured in the cross-sectional direction of the weld, which is calculated by the equidistant scanning method and defines the width points (x l，i ,y l，i ) and (x r，i ,y r，i ), the average width of the weld is calculated as follows: ； Among them, x l，i , x r，i are the x-coordinates of the left and right edges of the weld, n is the number of sampling points in the weld cross section, W std is the standard weld width, W std The value is 5mm~8mm, W avg <W min , weld shrinkage or insufficient penetration occurs, W avg >W max , the weld is too wide, affecting the structural strength.

6. The weld seam recognition method based on visual inspection according to claim 1, characterized in that: The weld difference ΔW in S5 is measured by calculating the difference between the maximum width and the minimum width, and the calculation formula is as follows: ； Among them, W max =max(W1, W2, W3, W4, W5..., W n ), W min =min(W1, W2, W3, W4, W5..., W n ), ΔW<allowable range, the weld width is evenly distributed and the welding quality is good; ΔW>allowable range, the weld is uneven, resulting in stress concentration and affecting the structural strength.

7. A weld seam recognition method based on visual inspection according to claim 4, 5 or 6, characterized in that: The comprehensive weld quality score S is weightedly calculated based on the weld length L, the weld width W, and the weld difference ΔW. ； α1, α2, α3 are weighted coefficients, satisfying α1+α2+α3=1, L std , W std , ΔW max For standard welding parameters, S≥0.9, excellent quality, good weld quality, and parameters meet the standards; 0.8≤S<0.9, qualified quality, and overall compliance with quality requirements; 0.6≤S<0.8, there may be weld defects, and inspection is recommended; S<0.6, unqualified quality, requiring repair or re-welding.

8. The weld seam identification method based on visual inspection according to claim 1, characterized in that: The quality report in S6 includes the test results, evaluation data and visual images. The test results are presented in a graphical manner, and the defect location, category and related parameters such as weld length and width are marked on the original image.

9. The weld seam recognition method based on visual inspection according to claim 1, characterized in that: The recognition method combines CUDA parallel computing technology and lightweight network optimization strategy in deployment, and adopts multi-threaded design to achieve efficient collaboration of image acquisition, processing and result display. By optimizing hardware resource utilization, it ensures that the system can respond quickly in resource-constrained scenarios.

10. The weld seam recognition method based on visual inspection according to claim 1, characterized in that: In step S4, a lightweight target detection model or a geometric fitting algorithm is used to replace part of the image segmentation steps to adapt to different hardware configurations and usage requirements.

Images