Pole post pre-welding addressing method based on deep learning

Abstract

The invention discloses a post pre-welding addressing method based on deep learning, and the method comprises the following steps: obtaining a power battery image, and carrying out the labeling of a label, so as to form a data set; constructing a recognition network based on a yolox detection model framework and a lightweight pooling method, wherein the recognition network comprises a feature extraction module and a multi-kernel convolution attention module; training the identification network by using the data set to obtain an identification model; inputting the power battery image into the recognition model to obtain a pole interested image, and performing dimension reduction processing to obtain a grey-scale map; performing convolution operation on the grey-scale image based on a Gaussian filter to output a binary image, and filling a black area in the binary image; and carrying out circle center positioning and circumference identification on the filled binarized image to obtain coordinates of a pole welding area. According to the method provided by the invention, the purpose of real-time addressing before welding of the post can be achieved.

Description

Technical Field

[0001] The present invention belongs to the technical field of automated welding, and in particular relates to a pole welding pre-addressing method based on deep learning. Background Art

[0002] Due to the rapid development of clean and green energy, the new energy vehicle industry is expanding. In response, my country's demand and requirements for power batteries are also growing rapidly. The core technology of the new energy vehicle industry is power battery technology, and its quality directly affects the performance of new energy vehicles. In order to ensure the stability and reliability of automobile power supply, it is necessary to improve the welding quality between battery pack poles in the power battery pack production process.

[0003] To improve the welding quality between poles, it is necessary to accurately locate the welding area of ​​the poles and perform precise welding within a fixed range. Positioning errors may lead to defects such as welding deviation, pores, and undercuts, which will reduce the conductivity of the battery and cause material waste. Traditional visual inspection methods are difficult to adapt to fast and efficient production environments, and the requirements for the production environment are relatively high.

[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: obtain pole addressing data, and the servo belt camera takes the marked points to locate the coordinates; Step 3: the pole addressing data is converted into the pole coordinates in the current camera coordinate system; Step 4: the pole coordinates are subtracted from the robot welding coordinates to obtain the pole relative to the laser coordinates; Step 5: the relative coordinates are converted into laser welding coordinates; Step 6: the welding coordinates are sent to PLC to guide laser welding.

[0005] Patent document CN117086480A discloses a battery module laser welding method and system method based on vision guidance, including: step S1, PLC controls the transfer cart to move the module to the side panel welding station; step S2, lifts the module, removes dust and turns on nitrogen protection; step S3, after visually positioning and measuring the distance of the side panel in the module through a camera and a rangefinder, controls the six-axis robot to move the welding nozzle to weld the side panel; step S4, after visually positioning and measuring the distance of the pole, transplants the connecting piece to the pole, and moves the transfer cart to the connecting piece welding station; step S5, lifts the module, removes dust and turns on nitrogen protection; step S6, after visually positioning and measuring the distance of the battery module, controls the six-axis robot to move the galvanometer to weld the connecting piece.

[0006] In the above technical solution, the welding target must be completely exposed within the camera's field of view. Each time the entire image is processed, it takes a relatively longer time. Slight changes in the welding environment can easily affect the effect of image processing, thereby affecting the subsequent polar coordinate data extraction. Summary of the invention

[0007] The purpose of the present invention is to provide a pole welding pre-addressing method based on deep learning, which can achieve the purpose of real-time addressing before pole welding.

[0008] In order to achieve the purpose of the present invention, the following technical solution is provided: a method for addressing before pole welding based on deep learning, comprising the following steps: Obtain a power battery image, and label the poles and pole information in the power battery image, and form a data set with the power battery image and the label; Based on the yolox detection model framework and lightweight pooling method, a recognition network is constructed, which includes a feature extraction module and a multi-core convolution attention module: The feature extraction module includes a multi-scale feature extractor, which is used to extract feature images of pole positions in battery images at multiple scales; The multi-core convolution attention module downsamples feature images of multiple scales by a lightweight pooling method, and concatenates the sampled features obtained by sampling from the channel dimension to obtain an overall feature map as a polar image of interest for output; The recognition network is trained using a training set to obtain a recognition model for extracting an image of interest of a pole; Input the power battery image into the recognition model to obtain a pole image of interest, and perform dimensionality reduction processing on the pole image of interest to obtain a corresponding grayscale image; Performing a convolution operation on the grayscale image based on a pre-constructed Gaussian filter to output a binary image including only pixel values ​​0 and 255, and filling the black area in the binary image using a morphological method; The Hough gradient circle detection method is used to locate the center and identify the circumference of the filled binary image to obtain the coordinates of the pole welding area.

[0009] The present invention introduces the convolutional network model into the pre-welding addressing method, which weakens the dependence of welding target finding on the camera field of view. At the same time, it uses multi-core convolutional attention to enhance the features of the target image, and uses a lightweight soft pooling module to accelerate the detection and positioning of the target. The Hough gradient method is combined with the neural network model to realize the integrated detection and positioning, which is used to quickly obtain higher-precision polar coordinates.

[0010] Specifically, the column information includes an image ID, an image path, an image name, and a pixel height and width of the image, wherein the pixel height and width of the image are represented by four coordinates of a rectangular frame.

[0011] Specifically, a CCD camera is used to capture a large number of power battery images from multiple angles, including images of power batteries in various complex environments. The number of power battery poles in the images varies, ensuring a multi-target data set. Use the labelimg tool to annotate the battery pole images and create a corresponding label file for each image. The corresponding label file contains the category label of the pole, the category label is Battery Pole, and the label file meets the xml label file standard in PascalVOC format; The battery final data set is divided into a training set, a validation set and a test set in a ratio of 8:1:1.

[0012] Specifically, the multi-scale feature extractor includes a 3×3 feature extractor, a 5×5 feature extractor and a 7×7 feature extractor.

[0013] Specifically, the output result expression of the multi-core convolution attention module is as follows: ; ;W and b represent the weight and bias parameters of the fully connected layer, represents the activation function, Represents the learnable weight representation of the final output result, Represents the overall feature map obtained by concatenation.

[0014] Specifically, the operation process of the lightweight pooling method is as follows: the kernel size of the pooling is set to k, the k×k area is defined as R, called the pooling window, the eigenvalues ​​in each pooling window are linearly transformed to calculate the weight, and the linear weight is used to replace the exponential weight in the original structure soft pooling. The weight is expressed as: ; Normalize the weights after linear transformation to obtain the distribution of weights, expressed as: ; Use weights to perform weighted summation on the eigenvalues ​​to obtain the result of lightweight soft pooling, expressed as: ;in, represents the learnable bias term, Represents the feature image to be downsampled.

[0015] Specifically, during the training process, the parameter batch_size 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 cosine learning rate was used to continuously adjust the number of training times and learning rate.

[0016] Specifically, 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, assign different weights to replace the value of the current pixel according to the distance of the multiple neighboring pixels from the current pixel, and select a fixed threshold for binarization according to the color characteristics of the target to obtain a binary image.

[0017] Specifically, the specific process of the Hough gradient circle detection method is as follows: Get all the 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, bidirectionally extending according to the positive and negative directions of the corresponding normal line, and voting for the pixel points added by the extended line to increase the count value in the accumulator; When all edge points have voted, the pixel coordinates with the highest count value in the accumulator are obtained to complete the center point positioning and circle identification of the pole.

[0018] 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 a gradient magnitude and a gradient direction as a result output.

[0019] Compared with the prior art, the present invention has the following beneficial effects: A multi-core convolutional attention module that focuses on multi-scale information for attention weight inference is used to enhance the detail features in the obtained data set. A lightweight soft pooling method is used to simplify the weight calculation while ensuring the pooling capability. At the same time, it ensures that the features are differentiable and can be used for reverse gradient transfer. The convolutional network model is then combined to achieve the pre-welding addressing and positioning of the pole. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 A flow chart of a pole welding pre-addressing method based on deep learning provided in this embodiment; Figure 2 A schematic diagram of a real machine with data annotation provided for this embodiment; Figure 3 A schematic diagram of the structure of the identification network provided in this embodiment; Figure 4 A schematic diagram of a lightweight soft pooling method provided in this embodiment; Figure 5 A schematic diagram of the change of the positioning loss curve of the model training provided in this embodiment; Figure 6 A schematic diagram of the change of the confidence loss curve of the model training provided in this embodiment; Figure 7 A schematic diagram of an image of a region of interest of a pole provided in this embodiment; Figure 8A schematic diagram of a binarized image provided in this embodiment; Fig. 9 A schematic diagram of circle center positioning and circle circumference identification provided in this embodiment. DETAILED DESCRIPTION

[0021] In order to make the purpose, technical scheme and advantages of the embodiments of the present invention clearer, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. The components of the embodiments of the present invention generally described and shown in the drawings here can be arranged and designed in various different configurations. Therefore, the following detailed description of the embodiments of the present invention provided in the drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the present invention. Based on the embodiments in the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0022] like Figure 1 As shown, a pole welding pre-addressing method based on deep learning provided in this embodiment includes the following steps: like Figure 2 As shown, power battery images are collected and manually annotated, and the annotated power battery images are proportionally divided to construct a data set for training and testing.

[0023] More specifically, in order to achieve a model effect that can be used in multiple places after one training, it is necessary to ensure that a sufficient number of images are collected. During the collection process, full consideration should be given to image collection in all situations, such as when the camera is slightly tilted, some poles are beyond the camera's field of view, and pole images under actual welding conditions. Image collection should be achieved from multiple angles to ensure that the target pole can be accurately detected in a variety of situations after one training.

[0024] When labeling data, the pole area should be accurately selected on the labeling software, and the image data and the labeling data xml file should be matched one by one. The xml tag file includes the image ID, image path, image name, pixel height and width of the image, where the pixel height and width of the image are represented by four coordinates of a rectangular box, including xmin, ymin, xmax, ymax, where (xmin, ymin) is the coordinate of the upper left vertex of the rectangular box, and (xmax, ymax) is the coordinate of the lower right vertex of the rectangular box. Accurate correspondence can help accurately locate the target position during model training and perform feature learning of the target position.

[0025] Based on the yolox detection model framework and lightweight pooling method, a recognition network is constructed, which includes a feature extraction module and a multi-core convolution attention module: The feature extraction module includes a multi-scale feature extractor, which is used to extract feature images of pole positions in battery images at multiple scales; The multi-core convolutional attention module downsamples feature images of multiple scales through a lightweight pooling method, and splices the sampled features obtained from the channel dimension to obtain an overall feature map as a polar image of interest for output.

[0026] like Figure 3 As shown, based on the yolox network model adopted in this embodiment, a multi-core convolution attention module is introduced, and a multi-scale feature extractor is used to extract features of different scales on 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 convolution kernels, 3×3, 5×5, and 7×7, are used for X.

[0027] For each convolution kernel, feature extraction is performed to obtain a feature map, which is expressed as , downsampling is performed using global average pooling, and the result is expressed as , and concatenate the sampled features from the channel dimension to obtain the overall feature map represented as , the characteristic results obtained by the above three steps are specifically expressed as: ; ; ; in represents the convolution process, represents kernels of different sizes, The feature map is averaged and the above results are further extracted using the fully connected layer to obtain , and then the weighted normalization is performed after mapping it using the activation function to obtain the final result of the multi-core convolution attention module, which is specifically expressed as: ; ; where W and b are the weight and bias parameters of the fully connected layer, is the selected activation function, It is a learnable weight representation of the final output result.

[0028] like Figure 4As shown, it is a schematic diagram of the lightweight pooling method used in this embodiment. By setting the kernel size of the pooling to k, the k×k area is defined as R, which is called the pooling window. The eigenvalues ​​in each pooling window are linearly transformed to calculate the weight. Here, we use linear weights to replace the exponential weights in the original structure soft pooling. The weights are expressed as: ; Then the weights after linear transformation are normalized to obtain the distribution of weights, which is expressed as: ; Finally, the weights are used to perform weighted summation on the eigenvalues ​​to obtain the result of lightweight soft pooling, which is expressed as: ;in, represents the learnable bias term, Represents the feature image to be downsampled.

[0029] The ReLU activation function ensures that the obtained weights are non-negative, b is a learnable bias term, and the lightweight soft pooling calculation method avoids complex exponential operations and complex normalization processes by using simple linear changes, while ensuring the differentiability of the weights. The ReLU expression is: ; The recognition network is trained using the data set to obtain a recognition model for extracting the image of interest of the pole.

[0030] More specifically, this embodiment inputs the divided training set images into the yolox network model that combines the multi-core convolution attention module and the lightweight soft pooling structure for training, sets the parameter batch_size to 16, uses the random gradient descent with momentum of 0.9, sets the initial learning rate to 0.01 and uses the cosine learning rate to adjust continuously, sets a reasonable number of training times, and judges whether the model training is underfitting or overfitting according to the convergence of the model's loss data. Each training completes the change of the output loss curve, and reasonably adjusts parameters such as the number of training times and the learning rate according to the curve changes.

[0031] Use the validation set to verify the improved yolox network model obtained through training, obtain the validation set results, analyze the differences between the training set results and the validation set results, and focus on observing whether the target is missed and whether the detection rate of the target is not much different from the detection rate on the validation set. If the validation result is not ideal, continue with the adjustment strategy in the previous step until the performance of the model on the training set and the validation set is balanced.

[0032] According to the existing training parameter settings, the model is trained, the number of iterations is given, and each iteration is completed by preliminarily judging whether the current model has converged through the change of the model's loss function curve. Specifically, for a specific given example, the number of iterations is 200, such as Figure 5 and Figure 6As shown in the figure, the loss of the model is relatively large at the beginning of training. In the 50-160th round of iteration, the loss value fluctuates within a certain range, but is still above 0.5, indicating that the loss value is not very small at this time, and the error of the model is relatively large. In the last few dozen rounds, the loss value further decreases to below 0.5 and remains stable in a small range.

[0033] The power battery image is input into the recognition model to obtain the pole image of interest, such as Figure 7 As shown, the polar column image of interest is subjected to dimensionality reduction processing to obtain a corresponding grayscale image; The grayscale image is convolved based on a pre-constructed Gaussian filter to output Figure 8 The binary image shown includes only 0 and 255 pixel values, and the black area in the binary image is filled using a morphological method.

[0034] More specifically, in this embodiment, by obtaining the pole image of interest obtained by model detection as a preprocessing object, the image is subjected to dimensionality reduction processing to obtain a grayscale image; Construct a Gaussian filter with a convolution kernel size of 3×3, and use the filter to perform a convolution operation on the grayscale image. For each pixel in the image, select the current pixel and its eight neighboring pixels, and assign different weights to replace the value of the current pixel according to the distance of the eight neighboring pixels from the current pixel. According to the color characteristics of the target, a suitable threshold is selected for binarization to obtain an image with only 0 and 255 pixel values ​​for subsequent gradient calculation. The morphological method is used to fill the small black areas in the binary image.

[0035] like Fig. 9 As shown, the Hough gradient circle detection method is used to locate the center and identify the circumference of the filled binary image to obtain the coordinates of the pole welding area.

[0036] More specifically, edge detection is performed on the image, and the normal direction of each edge in the image is calculated; An accumulator is set for each pixel of the image. For each edge point, according to the positive and negative directions of the normal, the pixel points where the normal extension line intersects within a certain range are voted, and the count value in the accumulator is increased.

[0037] When all edge points have voted, the pixel coordinates with the highest count value in the accumulator are obtained to complete the pole center point positioning and circle identification.

[0038] Furthermore, the normal direction of the edge pixel is obtained by performing convolution calculation on the x and y directions of each pixel in the image in both horizontal and vertical directions to obtain the gradient amplitude and gradient direction. For the horizontal direction, the gradient kernel is used , for the vertical direction, the gradient kernel is used , specifically expressed as follows: ; Convolve the above two gradient kernels with the image respectively, and calculate the gradient information of each pixel in the horizontal and vertical directions, as follows: ; ;in is the pixel value in the original image, and are weights in two different directions. After the above gradient information calculation, the gradient G and amplitude of the edge pixel can be obtained using the formula , the specific calculation is as follows: ; Further, the voting range and circle identification are specifically manifested in that a maximum radius value and a minimum radius value are obtained by pre-estimating the approximate radius range of the battery pole area. During the voting search process, the voting range is between the two radius values. Each vote not only records the number of normal intersections of edge pixels, but also records the search radius corresponding to the current number of intersections. The search radius recorded in the accumulator corresponding to the final center coordinate point is the radius of the pole welding area, where Fig. 9 Shown is what this implementation provides.

[0039] In summary, the present invention aims at the problem of pole addressing and positioning before welding of power battery modules, and establishes a real-time pole addressing method before welding, which has weak dependence on the camera and low requirements on the processing environment. Compared with the traditional visual detection method, the present invention introduces the convolutional network model into the pre-welding addressing method, which weakens the dependence of welding target finding on the camera field of view. At the same time, multi-core convolutional attention is used to enhance the features of the target image, and a lightweight soft pooling module is used to accelerate the detection and positioning of the target. The Hough gradient method is combined with the neural network model to realize the integrated detection and positioning, which is used to quickly obtain higher-precision pole coordinates.

[0040] In addition, the terms "upper", "lower", "inner", "outer", "front", "rear" are used for descriptive purposes only and should not be understood as indicating or implying relative importance. Unless otherwise specifically stated, the relative steps, numerical expressions and values ​​of the components and steps described in these embodiments do not limit the scope of the present invention.

[0041] Of course, the above description is only a specific embodiment of the present invention and is not intended to limit the scope of implementation of the present invention. All equivalent changes or modifications made according to the structure, characteristics and principles described in the patent application scope of the present invention should be included in the patent application scope of the present invention.

[0042] Finally, it should be noted that the above-described embodiments are only specific implementations of the present invention, which are used to illustrate the technical solutions of the present invention, rather than to limit them. The protection scope of the present invention is not limited thereto. Although the present invention is described in detail with reference to the above-described embodiments, ordinary technicians in the field should understand that any technician familiar with the technical field can still modify the technical solutions recorded in the above-described embodiments within the technical scope disclosed by the present invention, or can easily think of changes, or make equivalent replacements for some of the technical features therein; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the protection scope of the present invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.

Claims

1. A pole welding pre-addressing method based on deep learning, characterized in that: The following steps are involved: Obtain a power battery image, and label the poles and pole information in the power battery image, and form a data set with the power battery image and the label; Based on the yolox detection model framework and lightweight pooling method, a recognition network is constructed, which includes a feature extraction module and a multi-core convolution attention module: The feature extraction module includes a multi-scale feature extractor, which is used to extract feature images of pole positions in battery images at multiple scales; The multi-core convolution attention module downsamples feature images of multiple scales by a lightweight pooling method, and splices the sampled features obtained by sampling from the channel dimension to obtain an overall feature map as a polar image of interest for output; The recognition network is trained using the data set to obtain a recognition model for extracting the image of interest of the pole; Input the power battery image into the recognition model to obtain a pole image of interest, and perform dimensionality reduction processing on the pole image of interest to obtain a corresponding grayscale image; Performing a convolution operation on the grayscale image based on a pre-constructed Gaussian filter to output a binary image including only pixel values ​​0 and 255, and filling the black area in the binary image using a morphological method; The Hough gradient circle detection method is used to locate the center and identify the circumference of the filled binary image to obtain the coordinates of the pole welding area.

2. The method for addressing before pole welding based on deep learning according to claim 1, characterized in that: The column information includes image ID, image path, image name, pixel height and width of the image, wherein the pixel height and width image are represented by four coordinates of a rectangular frame.

3. The method for addressing before pole welding based on deep learning 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 method for addressing before pole welding based on deep learning according to claim 1, characterized in that: The output expression of the multi-core convolution attention module is as follows: ; ;W and b represent the weight and bias parameters of the fully connected layer, represents the activation function, Represents the learnable weight representation of the final output result, Represents the overall feature map obtained by concatenation.

5. The method for addressing before pole welding based on deep learning according to claim 1, characterized in that: The operation process of the lightweight pooling method is as follows: The kernel size of the pooling is set to k, and the k×k area is defined as R, which is called the pooling window. The eigenvalues ​​in each pooling window are linearly transformed to calculate the weights. The linear weights are used to replace the exponential weights in the original structure soft pooling. The weights are expressed as: ; Normalize the weights after linear transformation to obtain the distribution of weights, expressed as: ; Use weights to perform weighted summation on the eigenvalues ​​to obtain the result of lightweight soft pooling, expressed as: in, represents the learnable bias term, Represents the feature image to be downsampled.

6. The method for addressing before pole welding based on deep learning according to claim 1, characterized in that: During the training process, the batch_size parameter was set to 16, the stochastic gradient descent with momentum of 0.9 was used, the initial learning rate was set to 0.01, and the cosine learning rate was used to continuously adjust the number of training times and the learning rate.

7. The method for addressing before pole welding based on deep learning 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, assign different weights to replace the value of the current pixel according to the distance of the multiple neighboring pixels from the current pixel, and select a fixed threshold for binarization according to the color characteristics of the target to obtain a binary image.

8. The method for addressing before pole welding based on deep learning according to claim 1, characterized in that: The specific process of the Hough gradient circle detection method is as follows: Get all the 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, bidirectionally extending according to the positive and negative directions of the corresponding normal line, and voting for the pixel points added by the extended line to increase the count value in the accumulator; When all edge points have voted, the pixel coordinates with the highest count value in the accumulator are obtained to complete the center point positioning and circle identification of the pole.

9. The method for addressing before pole welding based on deep learning 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 a gradient magnitude and a gradient direction as a result output.

Images