A Deep Learning-Based Pre-Bond Addressing Method for Pole Brakes

Abstract

This invention discloses a deep learning-based pre-welding addressing method for electrode posts, comprising the following steps: acquiring and labeling power battery images to form a dataset; constructing a recognition network based on the YOLOx detection model framework and a lightweight pooling method, wherein the recognition network includes a feature extraction module and a multi-kernel convolutional attention module; training the recognition network using the dataset to obtain a recognition model; inputting the power battery images into the recognition model to obtain the electrode post image of interest, and performing dimensionality reduction processing to obtain a grayscale image; performing a convolution operation on the grayscale image based on a Gaussian filter to output a binarized image, and filling in the black areas in the binarized image; performing center positioning and circumference recognition on the filled binarized image to obtain the coordinates of the electrode post welding area. The method provided by this invention can achieve real-time pre-welding addressing of electrode posts.

Description

Technical Field

[0001] This invention belongs to the field of automated welding technology, and in particular relates to a deep learning-based method for pre-welding addressing of poles. Background Technology

[0002] The rapid development of clean and green energy has led to the continuous expansion of the new energy vehicle industry. Consequently, my country's demand for and requirements of power batteries have also increased rapidly. Power battery technology is the core technology of the new energy vehicle industry, and its quality directly affects the performance of new energy vehicles. To ensure a stable and reliable power supply for vehicles, it is necessary to improve the welding quality between the battery terminals during the power battery pack production process.

[0003] Improving the welding quality between terminals requires accurately locating the welding area and performing precise welding within a fixed range. Incorrect positioning can lead to defects such as misaligned welds, porosity, and undercut, reducing the battery's conductivity and wasting materials. Traditional visual inspection methods are ill-suited to fast-paced, high-efficiency production environments and have relatively high requirements for the production environment.

[0004] Patent document CN114571073A discloses a laser welding visual positioning guidance method, system, device, and medium, including the following steps: Step 1: Product module is in place; Step 2: Acquire pole addressing data, and use a servo-driven camera to capture marked points to locate coordinates; Step 3: Convert the pole addressing data into pole coordinates in the current camera coordinate system; Step 4: Subtract the robot welding coordinates from the pole coordinates to obtain the pole coordinates relative to the laser; Step 5: Convert the relative coordinates into laser welding coordinates; Step 6: Send the welding coordinates to the PLC to guide laser welding.

[0005] Patent document CN117086480A discloses a vision-guided laser welding method and system for battery modules, including: Step S1, PLC-controlled transfer trolley moves the module to the side plate welding station; Step S2, lifting the module, removing dust, and activating nitrogen protection; Step S3, after visually locating and measuring the side plate inside the module using a camera and rangefinder, controlling a six-axis robot to move the welding nozzle to weld the side plate; Step S4, after visually locating and measuring the electrode post, transferring the connecting piece onto the electrode post, and moving the transfer trolley to the connecting piece welding station; Step S5, lifting the module, removing dust, and activating nitrogen protection; Step S6, after visually locating and measuring the battery module, controlling a six-axis robot to move the galvanometer to weld the connecting piece.

[0006] The above technical solution requires that the welding target must be fully exposed within the camera's field of view, which takes a relatively longer time to process the entire image each time. Small changes in the welding environment can easily affect the image processing effect, thereby affecting the subsequent extraction of polar coordinate data. Summary of the Invention

[0007] The purpose of this invention is to provide a deep learning-based pre-soldering addressing method for poles, which can achieve real-time pre-soldering addressing of poles.

[0008] To achieve the objectives of this invention, the following technical solution is provided: a deep learning-based pre-welding addressing method for pole pieces, comprising the following steps:

[0009] Acquire images of the power battery and label the terminals and terminal information in the power battery images. Combine the power battery images and labels into a dataset.

[0010] A recognition network is constructed based on the YOLOx detection model framework and a lightweight pooling method. The recognition network includes a feature extraction module and a multi-kernel convolutional attention module.

[0011] The feature extraction module includes a multi-scale feature extractor, which is used to extract feature images of the pole positions in battery images at multiple scales.

[0012] The multi-kernel convolutional attention module downsamples feature images at multiple scales using a lightweight pooling method, and concatenates the sampled features along the channel dimension to obtain an overall feature map as an image of interest.

[0013] The recognition network is trained using the training set to obtain a recognition model for extracting polar images of interest.

[0014] The image of the power battery is input into the recognition model to obtain the pole image of interest, and the pole image of interest is subjected to dimensionality reduction processing to obtain the corresponding grayscale image.

[0015] The grayscale image is convolved based on a pre-constructed Gaussian filter to output a binarized image that includes only 0 and 255 pixel values, and morphological methods are used to fill in the black areas in the binarized image.

[0016] The Hough gradient circle detection method is used to locate the center of the circle and identify the circumference of the filled binarized image in order to obtain the coordinates of the pole welding area.

[0017] This invention introduces a convolutional network model into the pre-welding addressing method, reducing the dependence of welding target finding on the camera's field of view. Simultaneously, it employs multi-kernel convolutional attention to enhance the features of the target image and uses a lightweight soft pooling module to accelerate target detection and localization. The Hough gradient method is combined with a neural network model to achieve integrated detection and localization, enabling rapid acquisition of higher-precision polar coordinates.

[0018] Specifically, the polar information includes image ID, image path, image name, image pixel height and width, wherein the pixel height and width of the image are represented by the four coordinates of a rectangle.

[0019] Specifically, a large number of power battery images are captured from multiple different angles using a CCD camera, including power battery images under various complex environments. The number of power battery terminals in the images varies to ensure a multi-target dataset.

[0020] The labelimg tool is used to annotate the images of the power battery poles. A corresponding label file is created for each image. The corresponding label file contains the category label of the pole, which is "Battery Pole". The label file meets the PascalVOC format XML label file standard.

[0021] The final battery dataset is divided into a training set, a validation set, and a test set in a ratio of 8:1:1.

[0022] Specifically, the multi-scale feature extractor includes a 3×3 feature extractor, a 5×5 feature extractor, and a 7×7 feature extractor.

[0023] Specifically, the output expression of the multi-kernel convolutional attention module is as follows: ; Where W and b represent the weights and bias parameters of the fully connected layer, This represents the activation function. The learnable weight representation of the final output. This represents the overall feature map obtained by splicing.

[0024] Specifically, the lightweight pooling method operates as follows: The pooling kernel size is set to k, and this k×k region is defined as R, called the pooling window. A linear transformation is performed on the feature values ​​within each pooling window to calculate the weights. Linear weights replace the exponential weights in the original soft pooling structure, and their weights are expressed as: The weights after linear transformation are normalized to obtain their distribution, which is expressed as follows: We use weights to perform a weighted summation of the eigenvalues ​​to obtain the lightweight soft pooling result, which is expressed as: ;in, This represents a learnable bias term. This represents the feature image to be downsampled.

[0025] Specifically, during training, the batch_size parameter is set to 16, stochastic gradient descent with momentum of 0.9 is used, the initial learning rate is set to 0.01, and the number of training iterations and the learning rate are continuously adjusted using a cosine learning rate.

[0026] Specifically, the convolution operation of the Gaussian filter is as follows:

[0027] For each pixel in the image, select the current pixel and its multiple neighboring pixels. Based on the distance of the multiple neighboring pixels from the current pixel, assign different weights to perform weighted replacement of the current pixel's value. Then, select a fixed threshold based on the color characteristics of the target to perform binarization, thereby obtaining a binarized image.

[0028] Specifically, the process of the Hough gradient circle detection method is as follows:

[0029] Obtain all edges in the image and determine the normal direction of each edge;

[0030] An accumulator is set for each pixel in the image. Starting from the edge point on the edge, the accumulator is extended in both directions according to the positive and negative directions of the corresponding normal. The pixels where the extension lines are added are voted on to increase the count value in the accumulator.

[0031] Once all edge points have completed voting, obtain the coordinates of the pixel with the highest count value in the accumulator to complete the center point location and circumference recognition of the pole.

[0032] Specifically, the normal direction is determined by performing convolution calculations on the x and y directions of each pixel in the image in both horizontal and vertical directions to obtain the gradient magnitude and gradient direction as the output results.

[0033] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0034] A multi-kernel convolutional attention module that focuses on multi-scale information for attention weight inference is used to enhance the detailed features in the acquired dataset. A lightweight soft pooling method simplifies weight calculation while ensuring pooling capability and feature differentiability, which can be used for backpropagation. Thus, combined with a convolutional network model, pre-welding addressing and localization of pole pieces is achieved. Attached Figure Description

[0035] Figure 1 A flowchart of the deep learning-based pre-welding addressing method for poles provided in this embodiment;

[0036] Figure 2 This is a schematic diagram of the actual machine used for data annotation provided in this embodiment;

[0037] Figure 3 This is a schematic diagram of the structure of the recognition network provided in this embodiment;

[0038] Figure 4 This is a schematic diagram of the lightweight soft pooling method provided in this embodiment;

[0039] Figure 5 This is a schematic diagram illustrating the changes in the localization loss curve during model training provided in this embodiment.

[0040] Figure 6 This is a schematic diagram illustrating the change in the confidence loss curve during model training provided in this embodiment.

[0041] Figure 7 This is a schematic diagram of the region of interest image of the pole column provided in this embodiment;

[0042] Figure 8 This is a schematic diagram of the binarized image provided in this embodiment;

[0043] Figure 9 This is a schematic diagram of the center positioning and circumference recognition provided in this embodiment. Detailed Implementation

[0044] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0045] like Figure 1 As shown in this embodiment, a deep learning-based pre-welding addressing method for poles includes the following steps:

[0046] like Figure 2 As shown, images of power batteries are collected and manually labeled. The labeled power battery images are then proportionally divided to construct a dataset for training and testing.

[0047] More specifically, in order to achieve a model effect that can be used in multiple places after a single training, it is necessary to ensure that a sufficient number of images are collected. During the collection process, all situations should be fully considered, such as when the camera is slightly tilted, some poles are outside the camera's field of view, and pole images under actual welding conditions. Images should be collected from multiple angles to ensure that the target pole can be accurately detected in various situations after a single training.

[0048] When annotating data, the polar region should be accurately selected in the annotation software, and the image data and the annotation data XML file should be matched one-to-one. The XML tag file includes the image ID, image path, image name, and the pixel height and width of the image. The pixel height and width of the image are represented by four coordinates of a rectangle, including xmin, ymin, xmax, and ymax. (xmin, ymin) are the coordinates of the top left vertex of the rectangle, and (xmax, ymax) are the coordinates of the bottom right vertex of the rectangle. Precise matching can help the model accurately locate the target position during training and perform feature learning at the target position.

[0049] A recognition network is constructed based on the YOLOx detection model framework and a lightweight pooling method. The recognition network includes a feature extraction module and a multi-kernel convolutional attention module.

[0050] The feature extraction module includes a multi-scale feature extractor, which is used to extract feature images of the pole positions in battery images at multiple scales.

[0051] The multi-kernel convolutional attention module downsamples feature images at multiple scales using a lightweight pooling method, and concatenates the sampled features along the channel dimension to obtain an overall feature map as a polarimetric image of interest for output.

[0052] like Figure 3 As shown, this embodiment uses the YOLOx network model and introduces a multi-kernel convolutional attention module. A multi-scale feature extractor is used to extract features from the input feature map X, X∈R^(W×H×C), where W and H represent the width and height of the input feature map, respectively, and C represents the number of channels. Three different convolutional kernels, 3×3, 5×5, and 7×7, are used for X to extract features at different scales.

[0053] For each type of convolutional kernel, feature extraction is performed to obtain a feature map, which is represented as follows: Downsampling is performed using global average pooling, and the result is expressed as follows: The sampled features are then concatenated along the channel dimension to obtain the overall feature map, which is represented as follows: The feature results obtained from the above three steps are specifically expressed as follows: ; ;

[0054] ;

[0055] in Represents the convolution process. Representing kernels of different sizes, The feature map is subjected to average pooling, and the result obtained is further extracted using a fully connected layer. After mapping using an activation function and then weighted normalization, the final result of the multi-kernel convolutional attention module is obtained, specifically as follows: ; Where W and b are the weights and bias parameters of the fully connected layer. It is the selected activation function. It is the learnable weight representation of the final output.

[0056] like Figure 4 The diagram shows the lightweight pooling method used in this embodiment. The pooling kernel size is set to k, and the k×k region is defined as R, called the pooling window. The feature values ​​within each pooling window are linearly transformed to calculate the weights. Here, we replace the exponential weights in the original soft pooling structure with linear weights, which are expressed as follows: Then, the weights after the linear transformation are normalized to obtain the weight distribution, which is expressed as: Finally, the eigenvalues ​​are weighted and summed to obtain the lightweight soft pooling result, which is represented as: ;in, This represents a learnable bias term. This represents the feature image to be downsampled.

[0057] The ReLU activation function guarantees that the resulting weights are non-negative. `b` is a learnable bias term. This lightweight soft pooling method avoids complex exponential operations and normalization processes by using a simple linear transformation, while ensuring the differentiability of the weights. The ReLU expression is: The recognition network is trained using the dataset to obtain a recognition model for extracting polar images of interest.

[0058] More specifically, in this embodiment, the divided training set images are input into a YOLOv network model that combines a multi-kernel convolutional attention module and a lightweight soft pooling structure for training. The batch size is set to 16, stochastic gradient descent with a momentum of 0.9 is used, the initial learning rate is set to 0.01 and continuously adjusted using a cosine learning rate, and a reasonable number of training iterations are set. The convergence of the model's loss data is used to determine whether the model training is underfitting or overfitting. After each training iteration, the change in the output loss curve is recorded, and the number of training iterations and learning rate are adjusted accordingly based on the curve changes.

[0059] The improved YOLOv network model trained is validated using a validation set. The validation set results are then analyzed to determine the differences between the training set and validation set results. The focus is on whether any targets are missed and whether the target detection rate is similar to that on the validation set. If the validation results are not ideal, the adjustment strategy from the previous step should be continued until the model's performance on the training and validation sets is balanced.

[0060] Based on the existing training parameters, the model is trained. Given a specified number of iterations, the convergence of the model is initially determined by observing the change in the model's loss function curve after each iteration. Specifically, for a given instance, the number of iterations is 200. Figure 5 and Figure 6 As shown, the model's loss is relatively large at the beginning of training. During the 50-160 iterations, the loss value fluctuates within a certain range but remains above 0.5, indicating that the loss value is not very small at this point and the model's error is relatively large. In the last few dozen iterations, the loss value further decreases to below 0.5 and remains stable within a small range.

[0061] The image of the power battery is input into the recognition model to obtain the image of interest for the electrode post, such as... Figure 7 As shown, the polar image of interest is reduced in dimension to obtain the corresponding grayscale image;

[0062] The grayscale image is convolved based on a pre-constructed Gaussian filter to output as follows: Figure 8 The image shown is a binarized image containing only 0 and 255 pixel values, and the black areas in the binarized image are filled using morphological methods.

[0063] More specifically, in this embodiment, the polar image of interest obtained by model detection is used as the preprocessing object, and the image is dimensionality reduced to obtain a grayscale image;

[0064] Construct a Gaussian filter with a kernel size of 3×3, and use this filter to perform convolution operation on the grayscale image. For each pixel in the image, select the current pixel and its 8 neighboring pixels, and assign different weights to the 8 neighboring pixels based on their distance from the current pixel to perform weighted replacement of the current pixel's value.

[0065] Based on the color characteristics of the target, an appropriate threshold is selected for binarization, resulting in an image with only 0 and 255 pixel values ​​for subsequent gradient calculation. Morphological methods are then used to fill in the small black areas in the binary image.

[0066] like Figure 9 As shown, the Hough gradient circle detection method is used to locate the center of the circle and identify the circumference of the filled binarized image in order to obtain the coordinates of the pole welding area.

[0067] More specifically, edge detection is performed on the image, and the normal direction of each edge in the image is calculated;

[0068] An accumulator is set for each pixel of the image. For each edge point, the pixels where the extensions of the normals intersect are voted on within a certain range according to the positive and negative directions of the normals, and the count value in the accumulator is increased.

[0069] Once all edge points have completed voting, obtain the coordinates of the pixel with the highest count value in the accumulator to complete the polar center point localization and circumference recognition.

[0070] Furthermore, the edge pixel normal direction is obtained by performing convolution calculations on the x and y directions of each pixel in the image in both horizontal and vertical directions to obtain the gradient magnitude and gradient direction. For the horizontal direction, a gradient kernel is used. For the vertical direction, a gradient kernel is used. The specific expression is as follows: The two gradient kernels mentioned above are convolved with the image respectively to calculate the gradient information of each pixel in the horizontal and vertical directions, as follows: ; ;in These are the pixel values ​​from the original image. and These are weights in two different directions. After calculating the gradient information described above, the gradient G and magnitude of the edge pixels can be obtained using the formula. The specific calculations are as follows: Furthermore, the voting range and circumference recognition are specifically manifested in that, by pre-estimating the approximate radius range of the battery terminal region, a maximum radius value and a minimum radius value are obtained. During the voting search process, the voting range lies between these two radius values. Each vote records not only the number of intersections of the normals of the edge pixels but also the search radius corresponding to the current intersection number. Finally, the search radius recorded in the accumulator corresponding to the center coordinate point is the radius of the terminal welding area, where, for example... Figure 9 The image shown is provided in this embodiment.

[0071] In summary, this invention addresses the problem of electrode post addressing and localization before welding power battery modules, establishing a real-time electrode post addressing method with low camera dependence and minimal requirements for the processing environment. Compared to traditional visual detection methods, this invention introduces a convolutional network model into the pre-welding addressing method, reducing the dependence of welding target finding on the camera's field of view. Simultaneously, it employs multi-kernel convolutional attention to enhance the feature of the target image and uses a lightweight soft pooling module to accelerate target detection and localization. The Hough gradient method is combined with a neural network model to achieve integrated detection and localization, enabling rapid acquisition of higher-precision electrode post coordinates.

[0072] Furthermore, the terms "upper," "lower," "inner," "outer," "front," and "rear" are used for descriptive purposes only and should not be construed as indicating or implying relative importance. Unless otherwise specifically stated, the relative steps, numerical expressions, and values ​​of the components and steps set forth in these embodiments do not limit the scope of the invention.

[0073] Of course, the above description is only a specific embodiment of the present invention and is not intended to limit the scope of the present invention. All equivalent changes or modifications made to the structure, features and principles described in the claims of the present invention should be included in the scope of the claims of the present invention.

[0074] Finally, it should be noted that the above-described embodiments are merely specific implementations of the present invention, used to illustrate the technical solutions of the present invention, and not to limit it. The scope of protection of the present invention is not limited thereto. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that any person skilled in the art can still modify or easily conceive of changes to the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention, or make equivalent substitutions for some of the technical features; and these modifications, changes, or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should all be covered within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims.

Claims

1. A deep learning-based pre-soldering addressing method for pole pieces, characterized in that, Includes the following steps: Acquire images of the power battery and label the terminals and terminal information in the power battery images. Combine the power battery images and labels into a dataset. A recognition network is constructed based on the YOLOx detection model framework and a lightweight pooling method. The recognition network includes a feature extraction module and a multi-kernel convolutional attention module. The feature extraction module includes a multi-scale feature extractor, which is used to extract feature images of the pole positions in battery images at multiple scales. The multi-kernel convolutional attention module downsamples feature images at multiple scales using a lightweight pooling method, and concatenates the sampled features along the channel dimension to obtain an overall feature map as an image of interest. The recognition network is trained using the dataset to obtain a recognition model for extracting polar images of interest; The image of the power battery is input into the recognition model to obtain the pole image of interest, and the pole image of interest is subjected to dimensionality reduction processing to obtain the corresponding grayscale image. The grayscale image is convolved based on a pre-constructed Gaussian filter to output a binarized image that includes only 0 and 255 pixel values, and morphological methods are used to fill in the black areas in the binarized image. The Hough gradient circle detection method is used to locate the center of the circle and identify the circumference of the filled binarized image in order to obtain the coordinates of the pole welding area.

2. The deep learning-based pre-welding addressing method for pole pieces according to claim 1, characterized in that, The polar information includes image ID, image path, image name, and the image's pixel height and width, wherein the pixel height and width are represented by the four coordinates of a rectangle.

3. The deep learning-based pre-welding addressing method for pole pieces according to claim 1, characterized in that, The multi-scale feature extractor includes a 3×3 feature extractor, a 5×5 feature extractor, and a 7×7 feature extractor.

4. The deep learning-based pre-welding addressing method for pole pieces according to claim 1, characterized in that, The output expression of the multi-kernel convolutional attention module is as follows: ; Where W and b represent the weights and bias parameters of the fully connected layer, This represents the activation function. The learnable weight representation of the final output. This represents the overall feature map obtained by splicing.

5. The deep learning-based pre-welding addressing method for pole pieces according to claim 1, characterized in that, The operation process of the lightweight pooling method is as follows: The pooling kernel size is set to k, and the k×k region is defined as R, called the pooling window. A linear transformation is performed on the feature values ​​within each pooling window to calculate the weights. Linear weights replace the exponential weights in the original soft pooling structure, and their weights are expressed as: The weights after linear transformation are normalized to obtain their distribution, which is expressed as follows: We use weights to perform a weighted summation of the eigenvalues ​​to obtain the lightweight soft pooling result, which is expressed as: in, This represents a learnable bias term. This represents the feature image to be downsampled.

6. The deep learning-based pre-welding addressing method for pole pieces according to claim 1, characterized in that, During training, the batch_size parameter was set to 16, stochastic gradient descent with momentum of 0.9 was used, the initial learning rate was set to 0.01, and the number of training iterations and the learning rate were continuously adjusted using a cosine learning rate.

7. The deep learning-based pre-welding addressing method for pole pieces according to claim 1, characterized in that, The convolution operation of the Gaussian filter is as follows: For each pixel in the image, select the current pixel and its multiple neighboring pixels. Based on the distance of the multiple neighboring pixels from the current pixel, assign different weights to perform weighted replacement of the current pixel's value. Then, select a fixed threshold based on the color characteristics of the target to perform binarization, thereby obtaining a binarized image.

8. The deep learning-based pre-welding addressing method for pole pieces according to claim 1, characterized in that, The specific process of the Hough gradient circle detection method is as follows: Obtain all edges in the image and determine the normal direction of each edge; An accumulator is set for each pixel in the image. Starting from the edge point on the edge, the accumulator is extended in both directions according to the positive and negative directions of the corresponding normal. The pixels where the extension lines are added are voted on to increase the count value in the accumulator. Once all edge points have completed voting, obtain the coordinates of the pixel with the highest count value in the accumulator to complete the center point location and circumference recognition of the pole.

9. The deep learning-based pre-welding addressing method for pole pieces according to claim 8, characterized in that, The normal direction is determined by performing convolution calculations on the x and y directions of each pixel in the image in both horizontal and vertical directions to obtain the gradient magnitude and gradient direction as the output results.

Images