An AI-based system and method for monitoring and analyzing solder joint defects

Abstract

This invention discloses an artificial intelligence-based system and method for monitoring and analyzing weld defects, relating to the field of welding quality inspection technology. The method includes: acquiring surface images of composite current collector adapter tabs using a high-resolution industrial camera; identifying the precise location of each weld point; dynamically adjusting the angle, focal length, and light intensity parameters of the robotic arm and high-resolution industrial camera to capture images of each weld point at the optimal position based on the precise location of each weld point; extracting feature information of the weld points using image processing technology; automatically identifying the weld defect status of each weld point, classifying it into different weld defect levels, and marking the weld defect location and level on the weld point image; displaying the weld defect identification results on a display screen in real time, and printing the results into a report. This invention can automatically and accurately detect the weld point location and weld defect status of composite current collector adapter tabs, improving detection efficiency and accuracy.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of welding quality detection, and particularly relates to a virtual welding monitoring and analyzing system and method based on artificial intelligence. BACKGROUND

[0002] With the continuous development of battery technology, composite current collectors are increasingly widely used in battery manufacturing. In the battery manufacturing process, the tab is a key component for connecting the battery to the external circuit, and the welding quality directly affects the performance and safety of the battery. However, due to various factors in the welding process, virtual welding may occur between the tab and the composite current collector, which can seriously affect the normal operation of the battery. Therefore, it is particularly important to detect the virtual welding of the composite current collector adapter tab.

[0003] Existing virtual welding detection methods mostly use manual visual inspection or simple mechanical detection, which are not only inefficient, but also easily affected by human factors, resulting in low accuracy of the detection results. SUMMARY

[0004] The present application aims to provide a virtual welding monitoring and analyzing system and method based on artificial intelligence to solve the problems in the prior art.

[0005] To achieve the above-mentioned purpose, the present application provides the following technical scheme: a virtual welding monitoring and analyzing method based on artificial intelligence, the method comprising the steps of:

[0006] S1, acquiring the surface image of the composite current collector adapter tab through a high-resolution industrial camera; after image preprocessing of the surface image of the composite current collector adapter tab, the accurate position of each welding point is identified through a target recognition algorithm;

[0007] S2, controlling the movement of the mechanical arm, high-resolution industrial camera and light, according to the accurate position of each welding point, dynamically adjusting the angle, focal length and light intensity parameters of the mechanical arm, high-resolution industrial camera and light at the best position to take pictures of each welding point;

[0008] S3, after taking pictures of the welding points, using image processing technology to improve the clarity of the details of the welding points and extract the feature information of the welding points;

[0009] S4, using a virtual welding recognition model to automatically identify the virtual welding condition of each welding point, classifying different virtual welding levels according to the severity of the virtual welding, and marking the virtual welding position and virtual welding level on the welding point image;

[0010] S5, displaying the image of the welding point area and the virtual welding recognition result on the display screen in real time, and printing the virtual welding recognition result into a report through a printer.

[0011] According to the above scheme, the image preprocessing of the composite current collector adapter electrode surface image includes image grayscale processing, image denoising processing, and edge detection;

[0012] The image grayscale processing converts the acquired color composite current collector adapter tab surface image into a grayscale image, reducing data dimensionality;

[0013] The image denoising process uses Gaussian filtering to remove noise from the image, as shown in the following formula:

[0014]

[0015] Where I(x,y) represents the pixel values ​​of the original image; filtered (x,y) represents the pixel value of the image after Gaussian filtering; k represents the radius of the Gaussian kernel; G(u,v) represents the Gaussian kernel;

[0016] The edge detection utilizes the Canny edge detection algorithm to extract edges from the image.

[0017] According to the above scheme, the target recognition algorithm includes: training a solder joint recognition model using a large dataset of composite current collector adapter electrode surface images with precise solder joint locations marked; and identifying the precise location of the solder joint in the composite current collector adapter electrode surface image using the trained solder joint recognition model.

[0018] The composite current collector adapter electrode surface image dataset includes images under different lighting conditions, angles, and backgrounds;

[0019] The solder joint recognition model includes convolutional layers, pooling layers, and fully connected layers;

[0020] The convolutional layer divides the surface image of the composite current collector adapter tab into multiple small regions, and each region is convolved with a convolutional kernel. The convolutional kernel slides on the image to generate a two-dimensional feature map. Each pixel value of the feature map corresponds to the convolution result of the convolutional kernel at a certain position on the image. By using multiple different convolutional kernels, multiple feature maps are generated, and each feature map captures different local features.

[0021] The pooling layer reduces the dimensionality of the feature map through downsampling operations, thereby reducing computational complexity while retaining the main features;

[0022] The fully connected layer flattens the feature map of the pooling layer into a one-dimensional vector, performs a linear transformation through the weight matrix and bias term to generate a new feature vector, and uses an activation function to perform a non-linear transformation to enhance the expressive power of the solder joint recognition model.

[0023] According to the above scheme, the three-dimensional spatial coordinates of each weld point are obtained to locate the precise position of the weld point in three-dimensional space; based on the precise position data of the weld point, the angle, focal length and light intensity of the robotic arm and high-resolution industrial camera are adjusted.

[0024] When adjusting the angle, focal length, and light intensity of the robotic arm and high-resolution industrial camera, deep reinforcement learning is used to dynamically adjust these parameters through interaction with the environment, thereby finding the optimal parameter configuration.

[0025] Integrate the parameters of the robotic arm, high-resolution industrial camera angle, focal length, and light intensity into a single s t , is used to represent states, each state representing the current configuration, as shown in the following formula:

[0026] s t =[θ a1 ,θ a2 ,θ c ,f lens ,I light ];

[0027] Where, θ a1 and θ a2 Represented as the angle of the robotic arm; θ c Expressed as the angle of a high-resolution industrial camera; f lens Expressed as focal length; I light Indicated as light intensity;

[0028] The adjustment parameters of the robotic arm, high-resolution industrial camera angle, focal length, and light intensity are integrated into a single parameter. t , is used to represent the action space, which represents the actions that can be executed, as shown in the following formula;

[0029] a t =[Δθ a1 ,Δθ a2 ,Δθ c ,Δf lens ,ΔI light ];

[0030] Where, Δθ a1 and Δθ a2 This is expressed as the adjustment angle of the robotic arm; Δθ c This refers to the adjustment angle of a high-resolution industrial camera; Δf lens This is expressed as the adjustment amount for the focal length; ΔI light This represents the adjustment amount for light intensity;

[0031] A reward function is used to measure the effect of taking a certain action in a certain state. The reward function is defined as follows:

[0032] r t =R(s) t ,a t ) = Q(image) - P;

[0033] Where, r t R(s) represents the reward effect obtained at time t. t ,a t ) represents the reward function; Q(image) represents the image quality evaluation index, used to measure the sharpness and detail of the image; P represents the penalty term, which prevents the adjustment angle, focus adjustment amount, and light intensity of robotic arms or high-resolution industrial cameras from exceeding the reasonable range;

[0034] In deep reinforcement learning, actions are taken based on the current state to obtain feedback, which includes rewards and punishments, and behavioral strategies are adjusted based on this feedback.

[0035] The cumulative reward is maximized by optimizing the policy function, and the optimal policy is found using gradient ascent, as shown in the following formula:

[0036]

[0037] Where J(β) represents the expected cumulative reward under the policy parameter β; E πβ Represented as in policy π β The expected value is γ; T represents the total number of time steps; t Represented as a discount factor, it controls the importance of future rewards;

[0038] Policy gradient update, the formula is as follows:

[0039]

[0040] Where β represents the parameters of the policy network, controlling the angle, focal length, and light intensity of the robotic arm and camera; α represents the learning rate, controlling the step size for each update; Represented as the gradient operation with respect to β; E[] represents the expected value; logπ β (a t |s t ) represents the state s t Choose action a t The logarithmic probability of A; t Let this be represented by the dominance function, indicating the likelihood of taking action a at time t. t The degree of good or bad compared to taking other actions.

[0041] According to the above scheme, when the quality of the captured solder joint image is blurry or lacks detail, feedback is automatically provided, and the angle, focal length, and light intensity parameters of the robotic arm and high-resolution industrial camera are adjusted to recapture the solder joint image until a high-definition and detailed solder joint image is obtained; the adjustment process is recorded, and feedback and optimization are performed.

[0042] According to the above scheme, the image processing technology includes Gaussian filtering, contrast enhancement, and edge enhancement processing on the solder joint image;

[0043] The contrast enhancement process utilizes histogram equalization to adjust the pixel values ​​of the solder joint image, making the histogram distribution of the solder joint image more uniform and improving the contrast.

[0044] The histogram equalization first obtains the cumulative distribution function of the solder joint image, as shown in the following formula:

[0045]

[0046] Where CDF represents the cumulative distribution function; CDF(r k P(r) represents the cumulative probability distribution of gray levels from 0 to k; i ) is the grayscale value r in the solder joint image. i The probability distribution;

[0047] Then, by mapping each pixel value using the cumulative distribution function of the image, the enhanced solder joint image is obtained, as shown in the following formula:

[0048] I enhanced (x,y)=round(255×CDF(I(x,y)));

[0049] Among them, I enhanced (x,y) represents the enhanced solder joint image; I(x,y) represents the pixel value of the original image; CDF(I(x,y)) represents the cumulative probability;

[0050] The edge enhancement process utilizes the Sobel operator to calculate the gradient of the solder joint image, enhancing the edges and more clearly defining the shape of the solder joint.

[0051] According to the above scheme, the feature information of the solder joint includes the geometric shape features and texture features of the solder joint;

[0052] The geometric features of the solder joint include area, perimeter, and circularity; the area is the total number of pixels within the solder joint area; the perimeter is the total number of pixels at the edge of the solder joint; and the circularity is the degree to which the solder joint is nearly circular.

[0053] The texture features of the solder joints are extracted using a gray-level co-occurrence matrix, with texture features extracted from the co-occurrence matrix under different directions and distances.

[0054] According to the above scheme, a solder joint recognition model is trained using a large dataset of solder joint images labeled with solder joint conditions. The solder joint image dataset contains solder joint images with various solder joint conditions and labels them with solder joint severity levels, including mild, moderate, and severe solder joints.

[0055] Mild solder joint defects: The solder joint has an incomplete appearance and poor contact, but does not affect the function;

[0056] The term "moderate cold solder joint" refers to solder joints with obviously poor appearance and poor contact, which may affect functionality.

[0057] The severe cold solder joint: the solder joints have almost no welded connection, which seriously affects the function;

[0058] The extracted solder joint feature information is input into the faulty solder joint identification model;

[0059] The cold solder joint identification model analyzes whether each solder joint is a normal solder joint based on its characteristic information. For normal solder joints, the cold solder joint identification model will not mark them as cold solder joints. For solder joints suspected of being cold solder joints, the model proceeds to the next step of cold solder joint level classification.

[0060] For solder joints suspected of having poor solder joints, the poor solder joint identification model classifies them into different levels of poor solder joints based on the solder joint feature information;

[0061] For solder joints identified as having poor solder joints, the poor solder joint identification model marks the solder joint image when outputting the solder joint image, selects the solder joint with poor solder joint, and marks the poor solder joint level next to the solder joint.

[0062] According to the above scheme, the operator can view the detailed information and cold solder joint level of each solder joint through the display screen; the cold solder joint identification results are printed into a report through the display screen and printer; the report includes the location of the solder joint, whether it is a cold solder joint, the cold solder joint level, and a characteristic description of the cold solder joint.

[0063] An artificial intelligence-based cold solder joint monitoring and analysis system includes: an image acquisition module, an image processing module, a cold solder joint identification module, and an output module;

[0064] The image acquisition module includes a high-resolution industrial camera and a robotic arm. The high-resolution industrial camera includes a light source and lens suitable for the welding environment and is fixed on the robotic arm. It is used to acquire images of the surface of the composite current collector adapter welding tab and identify the precise location of the weld point. Based on the identified precise location of the weld point, the robotic arm dynamically adjusts the shooting angle and distance to capture images of the weld point from multiple angles and positions, and transmits the captured images of the weld point to the image processing module.

[0065] The image processing module is used to receive the solder joint image transmitted by the image acquisition module, perform image preprocessing and feature extraction on the image, identify the feature information of the solder joint, and transmit the feature information of the solder joint to the cold solder joint identification module.

[0066] The solder joint identification module constructs a solder joint identification model through training on a large number of solder joint samples; it receives solder joint features transmitted by the image processing module, uses the solder joint identification model to automatically identify the solder joint condition of each solder joint, classifies it into different solder joint levels according to the severity of the solder joint, and outputs solder joint images, marking the location and level of the solder joint on the solder joint images.

[0067] The output module includes a display screen and a printer, which are used to display the image of the solder joint area and the results of the cold solder joint identification in real time on the display screen, and print the results of the cold solder joint identification into a report through the printer for easy recording and archiving.

[0068] Compared with the prior art, the beneficial effects of the present invention are:

[0069] 1. This invention can automatically and accurately detect the poor soldering of composite current collector adapter lugs, improving detection efficiency and accuracy;

[0070] 2. This invention employs deep reinforcement learning to capture clear and accurate images of solder joints based on their locations, thereby improving the reliability of the detection results;

[0071] 3. This invention is applicable to the testing of composite current collector adapter welding tabs of different specifications and types, and has wide applicability. Attached Figure Description

[0072] Fig. 1 This is a flowchart illustrating the steps of an artificial intelligence-based method for monitoring and analyzing cold solder joints according to the present invention.

[0073] Fig. 2 This is a schematic diagram of the structure of an artificial intelligence-based cold solder joint monitoring and analysis system according to the present invention. Detailed Implementation

[0074] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0075] Example: Figs. 1-2 As shown, the present invention provides a technical solution: a method for monitoring and analyzing cold solder joints based on artificial intelligence;

[0076] Images of the composite current collector adapter tab surface are acquired using a high-resolution industrial camera; image preprocessing is performed on the composite current collector adapter tab surface images, including image grayscale conversion, image denoising, and edge detection;

[0077] The image grayscale processing converts the acquired 2048x2048x3 (RGB) color composite current collector adapter tab surface image into a 2048x2048 grayscale image, reducing data dimensionality.

[0078] The image denoising process uses Gaussian filtering to remove noise from the image, as shown in the following formula:

[0079]

[0080] Where I(x,y) represents the pixel values ​​of the original image; filtered (x,y) represents the pixel value of the image after Gaussian filtering; k represents the radius of the Gaussian kernel; G(u,v) represents the Gaussian kernel;

[0081] The edge detection utilizes the Canny edge detection algorithm to extract edges from the image.

[0082] A solder joint recognition model was trained using a large dataset of composite current collector adapter electrode tab surface images with precise solder joint locations annotated; and the trained solder joint recognition model was used to identify the precise locations of solder joints in the composite current collector adapter electrode tab surface images.

[0083] The composite current collector adapter electrode surface image dataset includes images under different lighting conditions, angles, and backgrounds;

[0084] The solder joint recognition model includes convolutional layers, pooling layers, and fully connected layers;

[0085] The convolutional layer divides the surface image of the composite current collector adapter tab into multiple small regions, and each region is convolved with a convolutional kernel. The convolutional kernel slides on the image to generate a two-dimensional feature map. Each pixel value of the feature map corresponds to the convolution result of the convolutional kernel at a certain position on the image. By using multiple different convolutional kernels, multiple feature maps are generated, and each feature map captures different local features.

[0086] The pooling layer reduces the dimensionality of the feature map through downsampling operations, thereby reducing computational complexity while retaining the main features;

[0087] The fully connected layer flattens the feature map of the pooling layer into a one-dimensional vector, performs a linear transformation through the weight matrix and bias term to generate a new feature vector, and uses an activation function to perform a non-linear transformation to enhance the expressive power of the solder joint recognition model.

[0088] The solder joint recognition model identifies the precise location of each solder joint in the surface image of the composite current collector adapter tab.

[0089] Obtain the three-dimensional spatial coordinates of each weld point to pinpoint its precise location in three-dimensional space; adjust the angle, focal length, and light intensity of the robotic arm and high-resolution industrial camera based on the precise location data of the weld points.

[0090] When adjusting the angle, focal length, and light intensity of the robotic arm and high-resolution industrial camera, deep reinforcement learning is used to dynamically adjust these parameters through interaction with the environment, thereby finding the optimal parameter configuration.

[0091] Integrate the parameters of the robotic arm, high-resolution industrial camera angle, focal length, and light intensity into a single s t , used to represent state, the formula is as follows:

[0092] s t =[θ a1 ,θ a2 ,θ c ,f lens ,I light ];

[0093] Where, θ a1 and θ a2 Represented as the angle of the robotic arm; θ c Expressed as the angle of a high-resolution industrial camera; f lens Expressed as focal length; I light Indicated as light intensity;

[0094] The initial state s0 is: s0 = [0,0,0,50,100];

[0095] The adjustment parameters of the robotic arm, high-resolution industrial camera angle, focal length, and light intensity are integrated into a single parameter. t , is used to represent the action space, and the formula is as follows;

[0096] a t =[Δθ a1 ,Δθ a2 ,Δθ c ,Δf lens ,ΔI light ];

[0097] Where, Δθ a1 and Δθ a2 This is expressed as the adjustment angle of the robotic arm; Δθ c This refers to the adjustment angle of a high-resolution industrial camera; Δf lens This is expressed as the adjustment amount for the focal length; ΔI light This represents the adjustment amount for light intensity;

[0098] The robotic arm's adjustment angle is adjusted in ranges of [-5°, 5°] each time; the high-resolution industrial camera's adjustment angle is adjusted in ranges of [-3°, 3°] each time; the focal length is adjusted in ranges of [-10, 10] each time; and the light intensity is adjusted in ranges of [-20, 20] each time.

[0099] A reward function is used to measure the effect of taking a certain action in a certain state. The reward function is defined as follows:

[0100] r t =R(s) t ,a t ) = Q(image) - P;

[0101] Where, r t R(s) represents the reward effect obtained at time t. t ,a t ) represents the reward function; Q(image) represents the image quality evaluation index, used to measure the sharpness and detail of the image; P represents the penalty term, which prevents the adjustment angle, focus adjustment amount, and light intensity of robotic arms or high-resolution industrial cameras from exceeding the reasonable range;

[0102] The cumulative reward is maximized by optimizing the policy function, and the optimal policy is found using gradient ascent, as shown in the following formula:

[0103]

[0104] Where J(β) represents the expected cumulative reward under the policy parameter β; E πβ Represented as in policy π β The expected value is γ; T represents the total number of time steps; t Represented as a discount factor;

[0105] Policy gradient update, the formula is as follows:

[0106]

[0107] Where β represents the parameters of the policy network, controlling the angle, focal length, and light intensity of the robotic arm and camera; α represents the learning rate, controlling the step size for each update; Represented as the gradient operation with respect to β; E[] represents the expected value; logπ β (a t |s t ) represents the state s t Choose action a t The logarithmic probability of A; t Let this be represented by the dominance function, indicating the likelihood of taking action a at time t.t Compared to the merits or demerits of taking other actions;

[0108] Adjust the robotic arm, high-resolution industrial camera angle, focal length, and light intensity parameters to capture images of each weld point from the optimal position;

[0109] After capturing images of the solder joints, image processing techniques are used to improve the clarity of the solder joint details;

[0110] The image processing techniques include Gaussian filtering, contrast enhancement, and edge enhancement of the solder joint image.

[0111] The contrast enhancement process utilizes histogram equalization to adjust the pixel values ​​of the solder joint image, making the histogram distribution of the solder joint image more uniform and improving the contrast.

[0112] The histogram equalization first obtains the cumulative distribution function of the solder joint image, as shown in the following formula:

[0113]

[0114] Where CDF represents the cumulative distribution function; CDF(r k P(r) represents the cumulative probability distribution of gray levels from 0 to k; i ) is the grayscale value r in the solder joint image. i The probability distribution;

[0115] The enhanced solder joint image is obtained by mapping each pixel value, as shown in the following formula:

[0116] I enhanced (x,y)=round(255×CDF(I(x,y)));

[0117] Among them, I enhanced (x,y) represents the enhanced solder joint image; I(x,y) represents the pixel value of the original image; CDF(I(x,y)) represents the cumulative probability;

[0118] The edge enhancement process utilizes the Sobel operator to calculate the gradient of the solder joint image and enhance the edges.

[0119] Extract the feature information of the solder joint; the feature information of the solder joint includes the geometric shape features and texture features of the solder joint;

[0120] The geometric features of the solder joint include area, perimeter, and circularity; the area is the total number of pixels within the solder joint area; the perimeter is the total number of pixels at the edge of the solder joint; and the circularity is the degree to which the solder joint is nearly circular.

[0121] The texture features of the solder joints are extracted using a gray-level co-occurrence matrix, with texture features extracted from the co-occurrence matrix under different directions and distances.

[0122] A solder joint recognition model was trained using a large dataset of solder joint images labeled with solder joint conditions. The solder joint image dataset contains solder joint images with various solder joint conditions and is labeled with solder joint severity levels, including mild, moderate, and severe solder joints.

[0123] The extracted solder joint feature information is input into the faulty solder joint identification model;

[0124] Minor cold solder joint: The solder joint is not fully visible and has poor contact, but it does not affect the function;

[0125] Moderate cold solder joint: The solder joint has an obviously poor appearance and poor contact, which may affect the function;

[0126] Severe cold solder joints: The solder joints barely form a welded connection, severely affecting functionality;

[0127] The cold solder joint identification model analyzes whether each solder joint is a normal solder joint based on its characteristic information. For normal solder joints, the cold solder joint identification model will not mark them as cold solder joints. For solder joints suspected of being cold solder joints, the model proceeds to the next step of cold solder joint level classification.

[0128] For solder joints suspected of having poor solder joints, the poor solder joint identification model classifies them into different levels of poor solder joints based on the solder joint feature information;

[0129] For solder joints identified as having poor solder joints, the poor solder joint identification model marks the solder joint image when outputting the solder joint image, selects the solder joint with poor solder joint, and marks the poor solder joint level next to the solder joint.

[0130] The image of the solder joint area and the results of the cold solder joint identification are displayed on the screen in real time. The results of the cold solder joint identification are printed into a report through the screen and printer. The report includes the location of the solder joint, whether it is a cold solder joint, the level of cold solder joint, and a characteristic description of the cold solder joint.

[0131] This invention provides another technical solution: an artificial intelligence-based cold solder joint monitoring and analysis system, which includes: an image acquisition module, an image processing module, a cold solder joint identification module, and an output module;

[0132] The image acquisition module includes a high-resolution industrial camera and a robotic arm. The high-resolution industrial camera includes a light source and lens suitable for the welding environment and is fixed on the robotic arm. It is used to acquire images of the surface of the composite current collector adapter welding tab and identify the precise location of the weld point. Based on the identified precise location of the weld point, the robotic arm dynamically adjusts the shooting angle and distance to capture images of the weld point from multiple angles and positions, and transmits the captured images of the weld point to the image processing module.

[0133] The image processing module is used to receive the solder joint image transmitted by the image acquisition module, perform image preprocessing and feature extraction on the image, identify the feature information of the solder joint, and transmit the feature information of the solder joint to the cold solder joint identification module.

[0134] The solder joint identification module constructs a solder joint identification model through training on a large number of solder joint samples; it receives solder joint features transmitted by the image processing module, uses the solder joint identification model to automatically identify the solder joint condition of each solder joint, classifies it into different solder joint levels according to the severity of the solder joint, and outputs solder joint images, marking the location and level of the solder joint on the solder joint images.

[0135] The output module includes a display screen and a printer, which are used to display the image of the solder joint area and the results of the cold solder joint identification in real time on the display screen, and print the results of the cold solder joint identification into a report through the printer for easy recording and archiving.

[0136] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A method for monitoring and analyzing cold solder joints based on artificial intelligence, characterized in that: The method includes the following steps: S1. Acquire images of the surface of the composite current collector adapter tab using a high-resolution industrial camera; After image preprocessing of the composite current collector transfer welding electrode tab surface image, the precise location of each welding point is identified by the target recognition algorithm; S2. Control the movement of the robotic arm, high-resolution industrial camera, and lights. Based on the precise position of each weld point, use deep reinforcement learning to dynamically adjust the angle, focal length, and light intensity parameters of the robotic arm and high-resolution industrial camera to capture each weld point at the optimal position. S3. After capturing the solder joint image, image processing technology is used to improve the clarity of the solder joint details and extract the feature information of the solder joint. S4. Using a cold solder joint identification model, automatically identify the cold solder joint status of each solder joint, classify it into different cold solder joint levels according to the severity of the cold solder joint, and mark the location and level of the cold solder joint on the solder joint image; S5. Display the image of the solder joint area and the results of the cold solder joint identification on the screen in real time, and print the cold solder joint identification results into a report through a printer.

2. The method for monitoring and analyzing cold solder joints based on artificial intelligence according to claim 1, characterized in that: The image preprocessing of the composite current collector adapter electrode surface image includes image grayscale processing, image denoising processing, and edge detection; The image grayscale processing converts the acquired color composite current collector adapter tab surface image into a grayscale image, reducing data dimensionality; The image denoising process uses Gaussian filtering to remove noise from the image, as shown in the following formula: Where I(x,y) represents the pixel values ​​of the original image; filtered (x,y) represents the pixel value of the image after Gaussian filtering; k represents the radius of the Gaussian kernel; G(u,v) represents the Gaussian kernel; The edge detection utilizes the Canny edge detection algorithm to extract edges from the image.

3. The artificial intelligence-based method for monitoring and analyzing cold solder joints according to claim 1, characterized in that: The target recognition algorithm includes training a solder joint recognition model using a large dataset of composite current collector adapter electrode tab surface images with precise solder joint locations marked; and using the trained solder joint recognition model to identify the precise location of the solder joints in the composite current collector adapter electrode tab surface images. The composite current collector adapter electrode surface image dataset includes images under different lighting conditions, angles, and backgrounds; The solder joint recognition model includes convolutional layers, pooling layers, and fully connected layers; The convolutional layer divides the surface image of the composite current collector adapter tab into multiple small regions, and each region is convolved with a convolutional kernel. The convolutional kernel slides on the image to generate a two-dimensional feature map. Each pixel value of the feature map corresponds to the convolution result of the convolutional kernel at a certain position on the image. By using multiple different convolutional kernels, multiple feature maps are generated, and each feature map captures different local features. The pooling layer reduces the dimensionality of the feature map through downsampling operations, thereby reducing computational complexity while retaining the main features; The fully connected layer flattens the feature map of the pooling layer into a one-dimensional vector, performs a linear transformation through the weight matrix and bias term to generate a new feature vector, and uses an activation function to perform a non-linear transformation to enhance the expressive power of the solder joint recognition model.

4. The artificial intelligence-based method for monitoring and analyzing cold solder joints according to claim 1, characterized in that: Obtain the three-dimensional spatial coordinates of each weld point to pinpoint its precise location in three-dimensional space; adjust the angle, focal length, and light intensity of the robotic arm and high-resolution industrial camera based on the precise location data of the weld points. When adjusting the angle, focal length, and light intensity of the robotic arm and high-resolution industrial camera, deep reinforcement learning is used to dynamically adjust these parameters through interaction with the environment, thereby finding the optimal parameter configuration. Integrate the parameters of the robotic arm, high-resolution industrial camera angle, focal length, and light intensity into a single s t , used to represent state, the formula is as follows: s t =[θ a1 ,i a2 ,i c ,f lens ,I light ]; Where, θ a1 and θ a2 Represented as the angle of the robotic arm; θ c Indicated as the angle of a high-resolution industrial camera; f lens Expressed as focal length; I light Indicated as light intensity; The adjustment parameters of the robotic arm, high-resolution industrial camera angle, focal length, and light intensity are integrated into a single parameter. t , is used to represent the action space, and the formula is as follows: a t =[Δθ a1 ,Dth a2 ,Dth c ,Δf lens ,D I light ]; Where, Δθ a1 and Δθ a2 This is expressed as the adjustment angle of the robotic arm; Δθ c This refers to the adjustment angle of a high-resolution industrial camera; Δf lens This is expressed as the adjustment amount for the focal length; ΔI light This represents the adjustment amount for light intensity; A reward function is used to measure the effect of taking a certain action in a certain state. The reward function is defined as follows: r t =R(s t ,a t )=Q(image)-P; Where, r t R(s) represents the reward effect obtained at time t. t ,a t ) represents the reward function; Q(image) represents the image quality evaluation index, used to measure the sharpness and detail of the image; P represents the penalty term, which prevents the adjustment angle, focus adjustment amount, and light intensity of robotic arms or high-resolution industrial cameras from exceeding the reasonable range; The cumulative reward is maximized by optimizing the policy function, and the optimal policy is found using gradient ascent, as shown in the following formula: Where J(β) represents the expected cumulative reward under the policy parameter β; E πβ Represented as in policy π β The expected value is γ; T represents the total number of time steps; t Represented as a discount factor; Policy gradient update, the formula is as follows: Where β represents the parameters of the policy network; α represents the learning rate, which controls the step size of each update; Represented as the gradient operation with respect to β; E[] represents the expected value; logπ β (a t |s t ) represents the state s t Choose action a t Logarithmic probability; A t Let this be represented by the dominance function, indicating the likelihood of taking action a at time t. t The degree of good or bad compared to taking other actions.

5. The artificial intelligence-based method for monitoring and analyzing cold solder joints according to claim 4, characterized in that: When the captured solder joint image is blurry or lacks detail, the system automatically provides feedback, adjusts the parameters of the robotic arm, high-resolution industrial camera angle, focal length, and light intensity, and recaptures the solder joint image until a high-definition and detailed solder joint image is obtained; the adjustment process is recorded, and feedback and optimization are performed.

6. The artificial intelligence-based method for monitoring and analyzing cold solder joints according to claim 1, characterized in that: The image processing techniques include Gaussian filtering, contrast enhancement, and edge enhancement of the solder joint image. The contrast enhancement process utilizes histogram equalization to adjust the pixel values ​​of the solder joint image, making the histogram distribution of the solder joint image more uniform and improving the contrast. The histogram equalization first obtains the cumulative distribution function of the solder joint image, as shown in the following formula: Where CDF represents the cumulative distribution function; CDF(r k P(r) represents the cumulative probability distribution of gray levels from 0 to k; i ) is the grayscale value r in the solder joint image. i The probability distribution; The enhanced solder joint image is obtained by mapping each pixel value, as shown in the following formula: I enhanced (x,y)=round(255×CDF(I(x,y))); Among them, I enhanced (x,y) represents the enhanced solder joint image; I(x,y) represents the pixel value of the original image; CDF(I(x,y)) represents the cumulative probability; The edge enhancement process uses the Sobel operator to calculate the gradient of the solder joint image and enhance the edges.

7. The artificial intelligence-based method for monitoring and analyzing cold solder joints according to claim 1, characterized in that: The feature information of the solder joint includes the geometric shape features and texture features of the solder joint; The geometric features of the weld joint include area, perimeter, and roundness; The texture features of the solder joints are extracted using a gray-level co-occurrence matrix, with texture features extracted from the co-occurrence matrix under different directions and distances.

8. The artificial intelligence-based method for monitoring and analyzing cold solder joints according to claim 1, characterized in that: A solder joint recognition model was trained using a large dataset of solder joint images labeled with solder joint conditions. The solder joint image dataset contains solder joint images with various solder joint conditions and is labeled with solder joint severity levels, including mild, moderate, and severe solder joints. The extracted solder joint feature information is input into the faulty solder joint identification model; The cold solder joint identification model analyzes whether each solder joint is a normal solder joint based on its characteristic information. For normal solder joints, the cold solder joint identification model will not mark them as cold solder joints. For solder joints suspected of being cold solder joints, the model proceeds to the next step of cold solder joint level classification. For solder joints suspected of having poor solder joints, the poor solder joint identification model classifies them into different levels of poor solder joints based on the solder joint feature information; For solder joints identified as having poor solder joints, the poor solder joint identification model marks the solder joint image when outputting the solder joint image, selects the solder joint with poor solder joint, and marks the poor solder joint level next to the solder joint.

9. The artificial intelligence-based method for monitoring and analyzing cold solder joints according to claim 1, characterized in that: Operators view detailed information and cold solder joint level for each solder joint on the display screen; the cold solder joint identification results are printed into a report via the display screen and printer; the report includes the location of the solder joint, whether it is a cold solder joint, the cold solder joint level, and a characteristic description of the cold solder joint.

10. A cold solder joint monitoring and analysis system based on artificial intelligence, characterized in that: The system includes: an image acquisition module, an image processing module, a cold solder joint identification module, and an output module; The image acquisition module includes a high-resolution industrial camera and a robotic arm. The high-resolution industrial camera includes a light source and lens suitable for the welding environment and is fixed on the robotic arm. It is used to acquire images of the surface of the composite current collector adapter welding tab and identify the precise location of the weld point. Based on the identified precise location of the weld point, the robotic arm dynamically adjusts the shooting angle and distance to capture images of the weld point from multiple angles and positions, and transmits the captured images of the weld point to the image processing module. The image processing module is used to receive the solder joint image transmitted by the image acquisition module, perform image preprocessing and feature extraction on the image, identify the feature information of the solder joint, and transmit the feature information of the solder joint to the cold solder joint identification module. The solder joint identification module constructs a solder joint identification model through training on a large number of solder joint samples; it receives solder joint features transmitted by the image processing module, uses the solder joint identification model to automatically identify the solder joint condition of each solder joint, classifies it into different solder joint levels according to the severity of the solder joint, and outputs solder joint images, marking the location and level of the solder joint on the solder joint images. The output module includes a display screen and a printer, which are used to display the image of the solder joint area and the results of the cold solder joint identification in real time on the display screen, and print the results of the cold solder joint identification into a report through the printer for easy recording and archiving.

Images