Blade battery pole piece defect detection model construction method based on GAN

Abstract

The invention discloses a GAN-based blade battery pole piece defect detection model construction method. The method comprises the following steps: acquiring a plurality of target pole piece images with defects; preprocessing the image to obtain a preprocessed image, and extracting an effective area with defects in the image; contour information is obtained for the effective areas, and the contour information is accurately classified according to feature parameters; performing data enhancement on the classified data to obtain an enhanced data set; in the data set, defect types and position information are used as labels, a neural network is trained, and a neural network model is used as a defect detection model of the blade battery pole piece. According to the invention, the data set after data enhancement is used as the training set to be input into the network, so that the problem that the training samples are seriously insufficient due to multiple defect types of the battery pole piece and low probability of occurrence of individual samples is solved, and the detection accuracy of the battery pole piece defect detection model is greatly improved; and various defects and position information of the blade battery pole piece can be quickly and accurately detected.

Description

Technical Field

[0001] The present invention belongs to the field of battery defect detection, and more specifically, relates to a method for constructing a blade battery pole piece defect detection model based on GAN. Background Art

[0002] The reliability and performance of blade batteries are crucial to the growing market of new energy vehicles. In this context, ensuring the quality of battery pole pieces has become a key challenge in the manufacturing process. However, during the manufacturing process of lithium batteries, various factors can cause scratches, dark spots and other defects on the battery pole pieces. These defects affect the performance and life of the battery, and in severe cases, may even cause safety accidents such as battery short circuits and fires.

[0003] There are deep learning methods to detect battery pole defects. However, traditional deep learning technology relies on a large amount of manually labeled sample data to train the detection model. In addition, battery poles are noisy, with many types of defective samples and a low probability of occurrence for individual samples, which leads to a serious shortage of training samples. At the same time, defective samples are usually small in size and have scarce appearance features, which increases the difficulty of manual labeling and easily introduces labeling errors. Summary of the invention

[0004] Purpose of the invention: In response to the above defects or improvement needs of the prior art, the present invention provides a method for constructing a blade battery pole piece defect detection model based on GAN, the purpose of which is to achieve automatic, accurate and rapid detection of blade battery defects under small sample conditions.

[0005] Technical solution: The present invention provides a method for constructing a blade battery pole piece defect detection model based on GAN, comprising the following steps: S1, collect several images of target pole pieces with defects; S2, preprocessing the acquired image to obtain a preprocessed image to extract a valid area with defects in the image; S3, obtaining contour information for the extracted effective area and accurately classifying the defects in the effective area according to the characteristic parameters; S4, performing data enhancement on the classified image data to obtain an enhanced image data set; S5. In the data set, the defect type and location information are used as labels to form a training set, the neural network is trained through the training set, and the trained neural network model is used as the defect detection model of the blade battery electrode.

[0006] Furthermore, step S2, preprocessing the collected image includes: The collected image is converted into a grayscale image and gamma-enhanced to improve the extractability of defects; the gamma-enhanced grayscale image is binarized, and the binarized image is opened and closed to remove the influence of small objects and the disconnection between adjacent defects; the binarized image processed by the opening and closing operations is pixel-traversed to extract the effective area containing the defects.

[0007] Further, step S3, obtaining contour information for the extracted effective area and accurately classifying it according to the characteristic parameters includes: Filling the extracted effective area containing defects, acquiring contour information according to the filled image, and classifying the defects in the effective area according to feature parameters based on the contour information; The characteristic parameters include the aspect ratio of the defect contour, the circularity, the area of ​​the defect, the gray value and the gray mean of the defect.

[0008] Further, in step S4, the data enhancement includes: Crop and randomly flip the classified data; The cropped and randomly flipped data are used to obtain a new training data set using a pole defect data augmentation method based on a generative adversarial network (GAN).

[0009] Furthermore, the new training data set is obtained by using a pole defect data enhancement method based on a generative adversarial network (GAN) to obtain the data after clipping and random flipping, including: The improved GAN-based generative adversarial network generates target defect images for expanding the training data set. The improved GAN-based generative adversarial network is based on the underlying architecture of DCGAN. A multi-scale residual block and a CA module are embedded in the generator G of the network to enhance feature channels and spatial attention. A spectral normalization module is embedded in the discriminator D of the network to improve the stability of discriminator training and the quality of generated images.

[0010] Furthermore, the network generator G is composed of a multi-layer convolutional structure, including an upsampling layer, a fully connected layer, a multi-layer convolution, a multi-scale residual block and a CA module.

[0011] Furthermore, the network discriminator D includes a series of spectral normalization layers, convolutional layers and random dropout layers as well as Mish functions and Dropout layers.

[0012] Furthermore, the network loss function is a HingeLoss loss function.

[0013] Furthermore, in step S5, the neural network is a YOLOV5s network.

[0014] The present application also requests protection for a GAN-based blade battery pole piece defect detection model, wherein the detection model is constructed using the above-mentioned GAN-based blade battery pole piece defect detection model construction method.

[0015] Beneficial effects: The GAN-based blade battery pole piece defect detection model construction method provided by the present invention includes collecting several target pole piece images with defects; preprocessing the collected images to extract the effective area with defects in the image; obtaining contour information for the extracted effective area and accurately classifying it according to feature parameters; performing data enhancement on the classified data to obtain an enhanced data set; in the data set, using the defect type and location information as labels to form a training set, and training the neural network through the training set to obtain a blade battery pole piece defect detection model, which can quickly and accurately detect a variety of defects and location information of the blade battery pole piece.

[0016] The electrode defect data enhancement method based on the generative adversarial network (GAN) in the present invention achieves enhanced feature channels and spatial attention by embedding multi-scale residual blocks and CA modules in the generator G of the DCGAN underlying architecture; and embedding a spectral normalization module in the discriminator D to improve the discriminator training stability and the quality of generated images. This method solves the problem of serious shortage of training samples caused by high noise in the cell electrode, many types of defective samples, and low probability of occurrence of individual samples. A small sample training blade battery electrode detection model is realized. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] Figure 1 A flowchart of a method for constructing a blade battery pole piece defect detection model based on GAN in an embodiment of the present invention; Figure 2 Schematic diagram of a blade battery pole piece image acquisition system according to an embodiment of the present invention; Figure 3 This is a structural diagram of a generator G based on an improved GAN generative adversarial network according to an embodiment of the present invention; Figure 4 4 is a structural diagram of a generative adversarial network discriminator D based on an improved GAN according to an embodiment of the present invention. DETAILED DESCRIPTION

[0018] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not intended to limit the present invention. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0019] refer to Figure 1The method for constructing a blade battery electrode defect detection model based on GAN provided in an embodiment of the present invention comprises the following steps: S1, collect several images of target pole pieces with defects; Furthermore, the embodiment of the present invention uses a blade battery pole piece image acquisition system to acquire images, such as Figure 2 As shown, the system includes a camera control unit, a CCD camera, a light source control unit, a light source, a pole piece conveying mechanism, and a pole piece flipping mechanism. The pole piece to be inspected is transported to the desired position by the pole piece conveying mechanism, and the camera control unit controls the camera to cooperate with the light source controlled by the light source control unit to collect the single-side image of the pole piece. After completion, the pole piece flipping mechanism completes the flipping of the pole piece, thereby realizing the collection of the image of the other side of the pole piece.

[0020] S2, preprocessing the acquired image to obtain a preprocessed image to extract a valid area with defects in the image; S21, converting the collected image into a grayscale image; S22, gamma enhancement is performed on the grayscale image to improve the extractability of defects; S23, performing binarization processing on the grayscale image after gamma enhancement; S24, performing opening and closing operations on the binary image to remove the influence of small objects and the disconnection between adjacent defects; S25, performing pixel traversal on the binary image processed by the opening and closing operations to extract the effective area containing the defects.

[0021] In some embodiments, step S24, an opening and closing operation is performed on the binary image, the opening operation is a combination of erosion followed by expansion, and the closing operation is a combination of expansion followed by erosion. Different structural elements B and different opening and closing operation times are selected according to the actual size of the image. The principle is to first use a large structural element to perform several rounds of closing operations to connect multiple areas that may be slightly far apart but still belong to the same defect impact range; then, use a small structural element to perform several rounds of opening operations to remove extremely small islands that do not affect the regional judgment; finally, use a small structural element to perform several rounds of closing operations to reconnect the subtle connections of the area that may have been disconnected in the previous opening operation.

[0022] S3, obtaining contour information for the extracted effective area and accurately classifying the defects in the effective area according to the characteristic parameters; Filling the extracted effective area containing defects, acquiring contour information according to the filled image, and classifying the defects in the effective area according to feature parameters based on the contour information; The characteristic parameters include the aspect ratio of the defect contour, the circularity, the area of ​​the defect, the gray value and the gray mean of the defect.

[0023] S4, performing data enhancement on the classified data to obtain an enhanced data set; S41, cropping and randomly flipping the classified data; S42. A new training data set is obtained by using a pole defect data enhancement method based on a generative adversarial network (GAN) on the cropped and randomly flipped data.

[0024] Furthermore, the electrode defect data enhancement method based on a generative adversarial network (GAN) in step S42 includes: The improved GAN-based generative adversarial network generates target defect images for expanding the training data set. The improved GAN-based generative adversarial network is based on the underlying architecture of DCGAN. Multi-scale residual blocks and CA modules are embedded in the generator G of the network to enhance feature channels and spatial attention; a spectral normalization module is embedded in the network discriminator D to achieve the improvement of discriminator training stability and generated image quality.

[0025] In some embodiments, the structure of the generator G in the generative adversarial network based on the improved GAN is as follows: Figure 3 As shown in the figure, it is composed of a multi-layer convolution structure, and combines upsampling layers, fully connected layers, multi-layer convolutions, multi-scale residual blocks and CA modules. The generation process starts with a fully connected layer, followed by batch normalization, upsampling and convolution layers to gradually increase the spatial dimension and reduce the depth of the feature map, and then apply multi-scale residual blocks and CA modules for feature optimization to finally generate the target image.

[0026] The structure of the discriminator D is as follows Figure 4 As shown in the figure, it consists of a series of spectral normalization layers, convolutional layers and random dropout layers. It inputs an image of fixed size and performs deep processing through multiple convolutional layers and fully connected layers. In addition, batch normalization, Mish nonlinear activation function and Dropout layer are added to prevent overfitting. Finally, the Flatten layer and a fully connected layer output vectors, which ensure the model's ability to identify the authenticity of the generated images.

[0027] In some embodiments, the loss function of the generative adversarial network is a HingeLoss loss function.

[0028] S5. In the data set, the defect type and location information are used as labels to form a training set, the neural network is trained through the training set, and the trained neural network model is used as the defect detection model of the blade battery electrode.

[0029] Furthermore, in step S5, the neural network is a YOLOV5s network.

[0030] The present application also requests protection for a GAN-based blade battery pole piece defect detection model, wherein the detection model is constructed using the above-mentioned GAN-based blade battery pole piece defect detection model construction method.

[0031] It will be easily understood by those skilled in the art that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the present invention.

Claims

1. A method for constructing a blade battery pole piece defect detection model based on GAN, characterized in that: The following steps are involved: S1, collect several images of target pole pieces with defects; S2, preprocessing the acquired image to obtain a preprocessed image to extract a valid area with defects in the image; S3, obtaining contour information for the extracted effective area and accurately classifying the defects in the effective area according to the characteristic parameters; S4, performing data enhancement on the classified image data to obtain an enhanced image data set; S5. In the data set, the defect type and location information are used as labels to form a training set, the neural network is trained through the training set, and the trained neural network model is used as the defect detection model of the blade battery electrode.

2. The GAN-based blade battery pole piece defect detection model construction method according to claim 1, characterized in that: Step S2, preprocessing the collected image includes: The collected image is converted into a grayscale image and gamma-enhanced to improve the extractability of defects; the gamma-enhanced grayscale image is binarized, and the binarized image is opened and closed to remove the influence of small objects and the disconnection between adjacent defects; the binarized image processed by the opening and closing operations is pixel-traversed to extract the effective area containing the defects.

3. The GAN-based blade battery electrode defect detection model construction method according to claim 1, characterized in that: Step S3, obtaining contour information for the extracted effective area and accurately classifying it according to the characteristic parameters, includes: Filling the extracted effective area containing defects, acquiring contour information according to the filled image, and classifying the defects in the effective area according to feature parameters based on the contour information; The characteristic parameters include the aspect ratio of the defect contour, the circularity, the area of ​​the defect, the gray value and the gray mean of the defect.

4. The method for constructing a blade battery electrode defect detection model based on GAN according to claim 1, characterized in that: Step S4, the data enhancement includes: Crop and randomly flip the classified data; The cropped and randomly flipped data are used to obtain a new training data set using a pole defect data augmentation method based on a generative adversarial network (GAN).

5. The method for constructing a blade battery pole piece defect detection model based on GAN according to claim 4, characterized in that: The new training data set is obtained by using the pole defect data enhancement method based on the generative adversarial network GAN to obtain the data after the clipping and random flipping processing, including: The improved GAN-based generative adversarial network generates target defect images for expanding the training data set. The improved GAN-based generative adversarial network is based on the underlying architecture of DCGAN. A multi-scale residual block and a CA module are embedded in the generator G of the network to enhance feature channels and spatial attention. A spectral normalization module is embedded in the discriminator D of the network to improve the stability of discriminator training and the quality of generated images.

6. The GAN-based blade battery pole piece defect detection model construction method according to claim 5, characterized in that: The network generator G is composed of a multi-layer convolutional structure, including an upsampling layer, a fully connected layer, a multi-layer convolution, a multi-scale residual block and a CA module.

7. The method for constructing a blade battery pole piece defect detection model based on GAN according to claim 5, characterized in that: The network discriminator D includes a series of spectral normalization layers, convolutional layers and random dropout layers as well as Mish functions and Dropout layers.

8. The method for constructing a blade battery pole piece defect detection model based on GAN according to claim 5, characterized in that: The network loss function is the HingeLoss loss function.

9. The method for constructing a blade battery pole piece defect detection model based on GAN according to claim 5, characterized in that: Step S5, the neural network is a YOLOV5s network.

10. A blade battery pole piece defect detection model based on GAN, characterized in that: The detection model is constructed using the GAN-based blade battery pole piece defect detection model construction method as described in any one of claims 1 to 9.

Images