A method for detecting tab welding defects based on deep learning
By analyzing the grayscale and binary features of electrode welding images using a deep learning-based method, and training a CNN network to identify welding defects, the problem of low accuracy in manual detection was solved, achieving efficient and accurate electrode welding defect detection.
Patent Information
- Application Number
- CN202410322562.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-03-20
- Publication Date
- 2024-11-01
- Estimated Expiration
- 2044-03-20
AI Technical Summary
Existing methods for detecting welding defects in electrode tabs mainly rely on manual experience and lack systematic inspection procedures, resulting in low accuracy and efficiency, and making it easy to misjudge welding defects.
A deep learning-based approach was adopted to collect electrode welding images, construct grayscale and binary images, analyze the shape features of the black areas, calculate the weld breakage index and grayscale distribution features, and use a CNN convolutional neural network for training to identify welding defects.
It improves the accuracy of electrode welding defect detection, reduces misjudgments caused by factors such as shadows, and achieves automated and accurate defect detection.
Smart Images

Figure CN118154560B_ABST
Abstract
Description
Technical Field
[0001] The present application relates to the field of image processing technology, and in particular to a method for detecting tab welding defects based on deep learning. Background Art
[0002] With the rapid development of the new energy industry, the market demand for lithium-ion power batteries has gradually increased. The tab welding process is a key part of the entire production process chain of lithium batteries. Its welding quality and welding accuracy will affect the performance and service life of the entire lithium battery, and in severe cases, it can directly lead to product scrapping. The tabs of conventional soft-pack batteries are flaky, soft, and multi-layer structures. Common welding methods include laser welding, ultrasonic welding, etc., which are prone to some welding defects during the welding process. At present, the domestic detection of defects generated after tab welding mainly relies on manual experience and naked eye inspection and judgment. There is a lack of corresponding and relatively systematic defect detection procedures for guidance, and it is greatly affected by human subjective factors, making it difficult to ensure the accuracy and efficiency of detection. With the rapid development of computer technology and image processing technology, the industry hopes to use machine vision technology to quickly and accurately realize the defect detection of lithium battery tabs after welding.
[0003] In the welding of the tabs, blowout is one of the common welding defects. When the quality of the lithium battery tabs after welding is automatically inspected, the detection of blowout defects is also prone to misjudgment. Blowout is mainly caused by the explosive reaction during the welding process, which causes metal flow and deformation. The stains and shadows around the solder joints on the PCM pole pieces are easily confused with the exploded solder joints, or are affected by nearby blowouts, resulting in the misjudgment of normal solder joints as blowouts. This interference factor that is easily misjudged as blowout is not completely eliminated, which is the main reason affecting the accuracy of defect detection. Summary of the invention
[0004] In order to solve the above technical problems, the present invention provides a tab welding defect detection method based on deep learning to solve the existing problems.
[0005] A method for detecting tab welding defects based on deep learning of the present invention adopts the following technical solution:
[0006] An embodiment of the present invention provides a method for detecting tab welding defects based on deep learning, the method comprising the following steps:
[0007] Collecting tab welding images;
[0008] The grayscale image and binary image of the tab welding image are used as the welding grayscale image and the welding binary image respectively; all black areas in the welding binary image are obtained; the shape characteristics of each black area in the welding binary image are analyzed to obtain the welding index of each black area; each suspected welding area is obtained according to the welding index of each black area; the omnidirectional welding propagation index of each suspected welding area is obtained according to the grayscale distribution characteristics of each suspected welding area; the welding grayscale curvature index of each suspected welding area is obtained according to the curvature of the contour of each suspected welding area and the radius of the welding spot; the welding curvature grayscale propagation characteristic index of each suspected welding area is obtained according to the omnidirectional welding propagation index and the welding grayscale curvature index of each suspected welding area;
[0009] The CNN convolutional neural network is trained according to the grayscale propagation characteristic index of the blown weld curvature in each suspected blown weld area; the defect detection of the tab welding is completed based on the trained CNN convolutional neural network.
[0010] Preferably, the black area in the welding binary image is specifically an area composed of pixels with a value of 0 in the welding binary image.
[0011] Preferably, the shape characteristics of each black area in the welding binary image are analyzed to obtain the welding index of each black area, specifically:
[0012] For each black area, the minimum bounding rectangle of the black area is obtained, and the ratio of the length to the width of the minimum bounding rectangle is used as the first ratio; the mode of the areas of all the minimum bounding rectangles in the welding binary image is obtained, and the ratio of the area of the minimum bounding rectangle to the mode is used as the second ratio; the product of the first ratio and the second ratio is used as the explosion welding index of each black area.
[0013] Preferably, the method of obtaining each suspected weld burst area according to the weld burst index of each black area is specifically as follows:
[0014] The upper quartile of the deep welding index of all black areas is obtained, and the black areas with deep welding index greater than the upper quartile are regarded as suspected deep welding areas.
[0015] Preferably, the omnidirectional weld propagation index of each suspected weld area is obtained according to the grayscale distribution characteristics of each suspected weld area, specifically:
[0016] Obtain the entropy value of the gray level co-occurrence matrix of each suspected weld explosion area in the four directions of 0°, 45°, 90° and 135°; obtain the lower quartile Ent of all the entropy values c ;
[0017] The expression of the omnidirectional welding propagation index of each suspected welding area is:
[0018]
[0019] In the formula, R d is the omnidirectional welding propagation index of suspected welding area d, is the entropy value of the gray-level co-occurrence matrix of the suspected fried weld area d in the c direction, Var{} is the variance function, and exp() represents an exponential function with a natural constant as the base.
[0020] Preferably, the grayscale curvature index of each suspected weld burst region is obtained according to the curvature degree of the contour of each suspected weld burst region, specifically:
[0021] For each suspected blown weld area, the centroid method is used to obtain the center of the suspected blown weld area, and the curvature of each point on the edge curve of the suspected blown weld area is obtained; the blown weld grayscale curvature index expression of each suspected blown weld area is:
[0022]
[0023] In the formula, E p is the grayscale curvature index of the suspected weld area p, Q p is the mode of the curvature of all pixels on the edge curve of the suspected blown solder joint area p, r represents the radius of the normal solder joint, g l,j , g l,j-1 is the gray value of the jth and j-1th pixels on the line connecting the lth edge pixel of the suspected weld area p and the center of the area, sgn() is the sign function, l n represents the total number of pixels on the line connecting the lth edge pixel point of the suspected deep-weld region p and the region center, including the lth edge pixel point and the region center, and n represents the number of edge pixels of the suspected deep-weld region p.
[0024] Preferably, the radius of the normal welding spot includes:
[0025] An edge detection algorithm is used to obtain a welding edge image of the tab welding image, and a Hough circle detection algorithm is used to obtain the radius of all circular targets of the welding edge image, and the mode of all the radii is used as the radius of a normal welding spot.
[0026] Preferably, the blown weld curvature grayscale propagation characteristic index of each suspected blown weld area is specifically the product of the omnidirectional blown weld propagation index of each suspected blown weld area and the blown weld grayscale curvature index.
[0027] Preferably, the CNN convolutional neural network is trained according to the grayscale propagation characteristic index of the weld curvature of each suspected weld area, specifically:
[0028] The tab welding image samples whose grayscale propagation feature index of the blown-up welding curvature is greater than 0 are marked as 1, and the tab welding image samples whose grayscale propagation feature index of the blown-up welding curvature is less than or equal to 0 are marked as 0, and the CNN convolutional neural network is trained.
[0029] Preferably, the defect detection of the tab welding is completed according to the trained CNN convolutional neural network, including:
[0030] The tab welding image is used as the input of the trained CNN convolutional neural network, and the output is the tab welding image with defects.
[0031] The present invention has at least the following beneficial effects:
[0032] The present invention analyzes the welding image of the tab, constructs a blown weld index through the shape characteristics of the minimum circumscribed rectangle of the black area in the binary image, and determines the suspected blown weld area according to the shape characteristics of the normal solder joint and the blown weld area; for the suspected blown weld area, by calculating the omnidirectional blown weld propagation index and the blown weld grayscale curvature index, the normal solder joints misjudged as blown welds due to factors such as shadows are effectively distinguished according to the grayscale characteristics and contour characteristics; the blown weld curvature grayscale propagation characteristic index is constructed, and a convolutional neural network is trained to judge the tab welding image, so as to improve the recognition accuracy of blown weld defects and reduce the misjudgment caused by shadows. BRIEF DESCRIPTION OF THE DRAWINGS
[0033] In order to more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings required for use in the embodiments or the prior art descriptions are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0034] Figure 1 A flowchart of a method for detecting tab welding defects based on deep learning provided by the present invention;
[0035] Figure 2 The image of the welding of the tab is blown up;
[0036] Figure 3 This is a misjudged normal tab welding image;
[0037] Figure 4 The flowchart for obtaining the grayscale propagation characteristic index of the deep-weld curvature is shown in Figure 2. DETAILED DESCRIPTION
[0038] In order to further explain the technical means and effects adopted by the present invention to achieve the predetermined invention purpose, the following is a detailed description of the specific implementation method, structure, features and effects of a deep learning-based tab welding defect detection method proposed by the present invention in combination with the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" does not necessarily refer to the same embodiment. In addition, specific features, structures or characteristics in one or more embodiments may be combined in any suitable form.
[0039] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.
[0040] The following is a detailed description of a method for detecting tab welding defects based on deep learning provided by the present invention in conjunction with the accompanying drawings.
[0041] An embodiment of the present invention provides a method for detecting tab welding defects based on deep learning.
[0042] Specifically, the following method for detecting tab welding defects based on deep learning is provided. Figure 1 , the method comprises the following steps:
[0043] Step S001: Acquire a tab welding image, and then pre-process the tab welding image.
[0044] This embodiment mainly considers analyzing the images in the tab welding and detecting the blown weld defects. When collecting the tab welding images, there may be shadows near the welding points, or the normal welding points may be affected by the nearby blown welds, which may lead to the false judgment of blown welds.
[0045] The image of the welded tab is collected by using a combined light source consisting of a spherical integral light source and a coaxial light source, and a high-resolution camera to obtain the original tab welding image. The original tab welding image is used as input, and the pre-processed image is obtained by using the median filtering technology. The median filtering technology is a well-known technology, and the specific process is not repeated here.
[0046] At this point, the tab welding image is obtained.
[0047] Step S002: Obtain a binary welding image; take the minimum circumscribed rectangle for each black area in the binary image to construct a weld burst index; calculate the omnidirectional weld burst propagation index and weld burst grayscale curvature index for each suspected weld burst area, and construct a weld burst curvature grayscale propagation characteristic index.
[0048] The color space conversion technology is used for the tab welding image to obtain the welding grayscale image. Since the color of the burnout and solder joints is darker than the color of the protection circuit board, the welding grayscale image is binarized to obtain the welding binary image. Figure 2 As shown in the figure, the left side is a grayscale image and the right side is a binary image; the misjudged normal tab welding image is as follows Figure 3 As shown, the left side is a grayscale image and the right side is a binary image.
[0049] The area composed of pixels with a value of 0 in the welding binary image is regarded as a black area. The black area represents the area of the blown weld or the normal weld spot. The minimum bounding rectangle is taken for each black area. In this embodiment, the minimum bounding rectangle of the black area i is taken as an example. The blown weld index is calculated based on the minimum bounding rectangle. The expression is:
[0050]
[0051] In the formula, A i is the welding index of black area i, are the length and width of the minimum bounding rectangle of black area i, S i is the area of the minimum bounding rectangle of black region i, and S is the mode of the minimum bounding rectangle areas of all black regions in the welding binary image. As a first ratio, as the second ratio.
[0052] If the black area is a normal solder joint area, since the normal solder joint is circular and the sizes of all normal solder joints are basically the same, the area of the minimum circumscribed rectangle of this black area is equal, and the length and width of its minimum circumscribed rectangle are also approximately equal, that is, the aspect ratio is close to 1; and because in the welding of the tabs, most of the solder joints are normal solder joints, the mode of the areas of all minimum circumscribed rectangles is the area of the minimum circumscribed rectangle of the normal solder joints, so the explosion index of the black area is closer to 1; if the black area is an explosion welding area or a suspected explosion welding area, since this area is not circular and is caused by the flow and deformation of the metal caused by the explosion, the minimum circumscribed rectangle area of the explosion welding area or the suspected explosion welding area is larger and the aspect ratio is not necessarily 1, so And S i >S. In summary, the welding index A of the black area i i Approximately equal to 1, the black area i is more likely to be a normal solder joint area; the hot solder index A of the black area i i >1, and A i The larger the value is, the more likely the black area i is a suspected blown weld area.
[0053] Calculate the weld index of all black areas, and mark the black areas with weld index greater than the upper quartile as suspected weld areas. Analyze the suspected weld area. If there is a weld defect in the suspected weld area, the cause of the weld is the pressure and heat generated by the explosion that spreads throughout the welding area, causing the metal to flow and deform. Therefore, the pixel values in the suspected weld area should be almost consistent, and deformation occurs in all directions around the weld. If the suspected weld area is not a weld, the cause of the area is the shadow around the weld. Due to the uneven lighting conditions, the shadow is not necessarily uniform, so the pixel values in the suspected weld area are inconsistent, and it cannot be guaranteed that deformation occurs in all directions around the weld.
[0054] This embodiment takes the suspected blown weld area d as an example, takes the suspected blown weld area d, the offset value 1 and the offset vector (Δx, Δy) as input, calculates the gray level co-occurrence matrix, and outputs the gray level co-occurrence matrix in four directions of 0°, 45°, 90° and 135° respectively. Then, the entropy Ent is calculated for each gray level co-occurrence matrix. d The calculation of the gray level co-occurrence matrix and the gray level co-occurrence matrix entropy is a well-known technique, and the specific process is not repeated here. The omnidirectional welding propagation index of the suspected welding area d is calculated, and the expression is:
[0055]
[0056] In the formula, R d is the omnidirectional welding propagation index of suspected welding area d, is the entropy of the gray-level co-occurrence matrix of the suspected weld area d in the c direction, Ent c is the lower quartile of the entropy of the gray-level co-occurrence matrix of all suspected weld areas in the c direction, Var{} is the variance function, and exp() represents an exponential function with a natural constant as the base.
[0057] If there is a weld defect in the suspected weld area, the pixel values in the suspected weld area are almost the same, and deformation occurs in all directions around the weld point, so the entropy of the gray level co-occurrence matrix in each direction is small and the difference between the entropies of the gray level co-occurrence matrix in each direction is small, thus The smaller the variance is, the larger the omnidirectional welding propagation index of the suspected welding area is. If the suspected welding area is not welding, the pixel values in the suspected welding area are inconsistent, and it is impossible to guarantee that deformation occurs in all directions around the welding point. Therefore, the entropy of the gray-level co-occurrence matrix in each direction is large, and the difference between the entropies of the gray-level co-occurrence matrix in each direction is also large, so The larger the value is, the larger the variance is, so the smaller the omnidirectional welding propagation index of the suspected welding area is.
[0058] The omnidirectional hot solder joint propagation index eliminates the possibility that shadow factors may cause misjudgment of hot solder joints. However, there is still a type of error, that is, hot solder joints occur at a certain solder joint, and the resulting hot solder joint area affects the normal solder joints adjacent to it. When using the omnidirectional hot solder joint propagation index for judgment, the normal solder joints affected by the hot solder joint area may also be judged as hot solder joints, so further screening is required. In response to this situation, by analyzing the gray value change and curvature change characteristics of the normal solder joints judged as hot solder joints, further screening is performed to avoid judging normal solder joints as hot solder joints.
[0059] First, the Canny edge detection algorithm is used on the tab welding image to obtain the welding edge image; the Hough circle detection is used on the welding edge image to output the center and radius of the circular target in the image. Hough circle detection is a well-known technology, and the specific process is not repeated here. The mode of the radius of all circular targets is used as the radius of the normal welding point.
[0060] Since blown welds can cause metal flow and deformation, this deformation occurs in all directions around the weld, and this deformation causes the edge curve of the blown weld area to show irregular bending, so that the curvature at each point on the edge curve is different and irregular; if a normal weld is judged to be blown weld, the irregular bending of the edge curve caused by the blown weld will only appear in one direction around the normal weld, and not in all directions. For the unaffected part, the curvature of each pixel point on the edge curve is basically the same and equal to the inverse of the radius of the normal weld.
[0061] Taking the suspected blown weld area p as an example, starting from the center of the suspected blown weld area p, select any point on the edge of the suspected blown weld area p, connect the point with the center of the suspected blown weld area p to obtain a line segment l, and take the gray value g of each pixel on the line segment l. Then calculate the curvature of each point on the edge curve of the suspected blown weld area p, and then calculate the blown weld grayscale curvature index:
[0062]
[0063] In the formula, E p is the grayscale curvature index of the suspected weld area p, Q p is the mode of the curvature of all pixels on the edge curve of the suspected blown solder joint area p, r is the radius of the normal solder joint, g l,j , g l,j-1 is the gray value of the jth and j-1th pixels on the line connecting the lth edge pixel of the suspected weld area p and the center of the area, sgn() is the sign function, l nrepresents the total number of pixels on the line connecting the lth edge pixel point of the suspected weld area p and the center of the area, including the lth edge pixel point and the center of the area, and n represents the number of edge pixels of the suspected weld area p. It should be noted that the center of the suspected weld area is calculated using the centroid method, which will not be repeated in this embodiment. It should be further noted that, considering the existence of the mode of the curvature of all pixels on the edge curve of the suspected weld area p, it means that it has been deformed in all directions and there is no arc part, so it can be determined as a weld. At this time, Q p =0.
[0064] If there is a weld defect in the suspected weld area p, the gray value in the center is the smallest, and the gray value gradually increases as it approaches the edge of the area, so g l,j -g l,j-1 >0, and the curvature of each pixel on the edge curve is irregular, and the mode of the curvature Q p There is no relationship between the radius r of the normal solder joint, so So at this time, the gray curvature index E of the explosion welding p >0. If the suspected blown solder area p is caused by a normal solder joint being misjudged as blown solder, the gray value gradually decreases from the center of the normal solder joint to the edge of the area, so that g l,j -g ,j-1 <0, and the mode of the curvature of all pixels on the edge curve should be equal to the curvature of the pixels on the edge curve of the normal solder joint, that is, the reciprocal of the radius r of the normal solder joint, so is approximately equal to 0, so at this time E p ≤0.
[0065] Each suspected hot weld area that is initially screened out has a two-dimensional feature vector (R, E), where R is the omnidirectional hot weld propagation index of the suspected hot weld area, and E is the hot weld grayscale curvature index of the suspected hot weld area. Taking the suspected hot weld area m as an example, the hot weld curvature grayscale propagation feature index of the suspected hot weld area is calculated based on this two-dimensional feature vector, and the expression is:
[0066] H m =E m ×R m
[0067] In the formula, H m is the grayscale propagation characteristic index of the weld curvature of the suspected weld region m, E m is the grayscale curvature index of the suspected weld zone m, R m is the omnidirectional hot-welding propagation index of the suspected hot-welding area m. The acquisition process of the hot-welding curvature grayscale propagation characteristic index is as follows: Figure 4 shown.
[0068] Omnidirectional welding propagation index R of suspected welding area m m The larger it is, the more likely it is a blown weld. The blown weld grayscale curvature index E of the suspected blown weld area m m >0, the more likely it is a blown weld; the blown weld grayscale curvature index E of the suspected blown weld area m m ≤0, the more likely it is that the solder joint is misjudged as a normal solder joint.
[0069] At this point, the grayscale propagation characteristic coefficient of the blown weld curvature of each suspected blown weld area is obtained.
[0070] Step S003: Detect the blown weld defects in the tab welding by using deep learning according to the blown weld curvature grayscale propagation characteristic coefficient of each suspected blown weld area.
[0071] A CNN convolutional neural network is used. According to the grayscale propagation characteristic index H of the blown weld curvature of each blown weld sample, samples with H>0 are regarded as real blown welds, and samples with H≤0 are regarded as normal welds. The Labelme image labeling tool is used to assign corresponding labels to each blown weld sample. The labels are set to 0 and 1 respectively, where 0 represents a normal weld and 1 represents a blown weld defect. The CNN convolutional neural network is trained. The training of the CNN convolutional neural network is a well-known technology, and the specific process will not be repeated. The lug welding image is used as input, and a CNN convolutional neural network is used. The binary cross-entropy loss (Binary Cross-EntropyLoss) is used as the loss function. The Adam optimization algorithm is used to output the defective lug welding image.
[0072] At this point, the detection of welding defects in the tab welding is completed.
[0073] In summary, the embodiment of the present invention analyzes the welding image of the tab, constructs a blown weld index through the shape characteristics of the minimum circumscribed rectangle of the black area in the binary image, and determines the suspected blown weld area according to the shape characteristics of the normal weld spot and the blown weld area; for the suspected blown weld area, by calculating the omnidirectional blown weld propagation index and the blown weld grayscale curvature index, the normal weld spots that are misjudged as blown welds due to factors such as shadows are effectively distinguished according to the grayscale characteristics and contour characteristics; the blown weld curvature grayscale propagation feature index is constructed, and a convolutional neural network is trained to judge the tab welding image, so as to improve the recognition accuracy of blown weld defects and reduce misjudgments caused by shadows.
[0074] It should be noted that the sequence of the above embodiments of the present invention is only for description and does not represent the advantages and disadvantages of the embodiments. The above is a description of a specific embodiment of this specification. In addition, the processes depicted in the accompanying drawings do not necessarily require the specific order or continuous order shown to achieve the desired results. In some embodiments, multitasking and parallel processing are also possible or may be advantageous.
[0075] The various embodiments in this specification are described in a progressive manner, and the same or similar parts between the various embodiments can be referenced to each other, and each embodiment focuses on the differences from other embodiments.
[0076] The embodiments described above are only used to illustrate the technical solutions of the present application, rather than to limit them. Modifications to the technical solutions recorded in the aforementioned embodiments, or equivalent replacement of some of the technical features therein, do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present application, and should all be included in the protection scope of the present application.
Claims
1. A method for detecting tab welding defects based on deep learning, characterized in that: The method comprises the following steps: Collecting tab welding images; The grayscale image and binary image of the tab welding image are used as the welding grayscale image and the welding binary image respectively; all black areas in the welding binary image are obtained; the shape characteristics of each black area in the welding binary image are analyzed to obtain the welding index of each black area; each suspected welding area is obtained according to the welding index of each black area; the omnidirectional welding propagation index of each suspected welding area is obtained according to the grayscale distribution characteristics of each suspected welding area; the welding grayscale curvature index of each suspected welding area is obtained according to the curvature of the contour of each suspected welding area and the radius of the welding spot; the welding curvature grayscale propagation characteristic index of each suspected welding area is obtained according to the omnidirectional welding propagation index and the welding grayscale curvature index of each suspected welding area; The CNN convolutional neural network is trained according to the grayscale propagation characteristic index of the blown weld curvature of each suspected blown weld area; the defect detection of the tab welding is completed according to the trained CNN convolutional neural network; The shape features of each black area in the welding binary image are analyzed to obtain the welding index of each black area, specifically: For each black area, the minimum bounding rectangle of the black area is obtained, and the ratio of the length to the width of the minimum bounding rectangle is used as the first ratio; the mode of the areas of all the minimum bounding rectangles in the welding binary image is obtained, and the ratio of the area of the minimum bounding rectangle to the mode is used as the second ratio; the product of the first ratio and the second ratio is used as the explosion welding index of each black area; The omnidirectional welding propagation index of each suspected welding area is obtained according to the grayscale distribution characteristics of each suspected welding area, specifically: Obtain the entropy value of the gray level co-occurrence matrix of each suspected weld explosion area in the four directions of 0°, 45°, 90° and 135°; obtain the lower quartile Ent of all the entropy values c ; The expression of the omnidirectional welding propagation index of each suspected welding area is: In the formula, R d is the omnidirectional welding propagation index of suspected welding area d, is the entropy value of the gray-level co-occurrence matrix of the suspected fried weld area d in the c direction, Var{} is the variance function, and exp() represents an exponential function with a natural constant as the base; The method of obtaining the grayscale curvature index of each suspected weld burst region according to the curvature degree of the contour of each suspected weld burst region is specifically: For each suspected blown weld area, the centroid method is used to obtain the center of the suspected blown weld area, and the curvature of each point on the edge curve of the suspected blown weld area is obtained; the blown weld grayscale curvature index expression of each suspected blown weld area is: In the formula, E p is the grayscale curvature index of the suspected weld area p, Q p is the mode of the curvature of all pixels on the edge curve of the suspected blown solder joint area p, r represents the radius of the normal solder joint, g l,j , g l,j-1 is the gray value of the jth and j-1th pixels on the line connecting the lth edge pixel of the suspected weld area p and the center of the area, sgn() is the sign function, l n represents the total number of pixels on the line connecting the lth edge pixel point of the suspected deep-weld region p and the region center, including the lth edge pixel point and the region center, and n represents the number of edge pixels of the suspected deep-weld region p; The weld burst curvature grayscale propagation characteristic index of each suspected weld burst area is specifically the product of the omnidirectional weld burst propagation index of each suspected weld burst area and the weld burst grayscale curvature index; The CNN convolutional neural network is trained according to the grayscale propagation characteristic index of the weld curvature of each suspected weld area, specifically: The tab welding image samples whose grayscale propagation feature index of the blown-up welding curvature is greater than 0 are marked as 1, and the tab welding image samples whose grayscale propagation feature index of the blown-up welding curvature is less than or equal to 0 are marked as 0, and the CNN convolutional neural network is trained.
2. A method for detecting tab welding defects based on deep learning as claimed in claim 1, characterized in that: The black area in the welding binary image is specifically an area composed of pixels with a value of 0 in the welding binary image.
3. A method for detecting tab welding defects based on deep learning as claimed in claim 1, characterized in that: The method of obtaining each suspected weld explosion area according to the weld explosion index of each black area is as follows: The upper quartile of the deep welding index of all black areas is obtained, and the black areas with deep welding index greater than the upper quartile are regarded as suspected deep welding areas.
4. A method for detecting tab welding defects based on deep learning as claimed in claim 1, characterized in that: The radius of the normal welding point includes: An edge detection algorithm is used to obtain a welding edge image of the tab welding image, and a Hough circle detection algorithm is used to obtain the radius of all circular targets of the welding edge image, and the mode of all the radii is used as the radius of a normal welding spot.
5. The method for detecting tab welding defects based on deep learning according to claim 1, characterized in that: The defect detection of the tab welding is completed according to the trained CNN convolutional neural network, including: The tab welding image is used as the input of the trained CNN convolutional neural network, and the output is the tab welding image with defects.
