A method for detecting surface particles of a battery pole piece

Abstract

The application discloses a kind of battery pole piece surface particulate matter detection methods, comprising the following steps: utilize teacher network and student network to build teacher-student network;Utilize student network to build twin neural network;Sample set is constructed and data is preprocessed;Teacher-student network is trained using sample set, twin neural network is trained using sample set;Detection sample is input into trained teacher-student network, and the difference graph output by teacher-student network is obtained, when the difference graph is defect-free, the detection result of this detection sample is no exception;When the difference graph exists defect, the detection result of this detection sample is abnormal, detection sample is input into trained twin neural network again, and the intersection graph of twin neural network is obtained, the intersection of intersection graph and difference graph is obtained Complete defect position;The application has the advantages that: detection precision is high, with the data set that can be used for training, easy to detect and meet real-time requirements.

Description

Technical Field

[0001] This invention relates to the field of electrode testing, and more specifically to a method for detecting particulate matter on the surface of battery electrodes. Background Technology

[0002] In lithium battery production, particulate matter can adhere to the electrode surface due to improper dust management or debris generated during the electrode fabrication process. This can cause self-discharge of the core, breaking down the ceramic film and leading to a short circuit. Because the particles are small and randomly distributed, current lithium battery manufacturers primarily use brushes to remove them during the electrode rolling process. However, since the electrodes and particles are usually tightly adhered, brushes cannot remove most of the particles and can easily damage defective electrodes. With the development of digital imaging, digital image-based particle detection of lithium battery electrode surfaces, especially deep learning-based particle detection, has shown extremely superior performance. Existing technologies mainly include:

[0003] 1. Anomaly detection on polarimetric surfaces based on traditional digital images: Traditional methods mainly include anomaly detection based on template matching, anomaly detection based on feature training, anomaly detection based on frequency domain analysis, and anomaly detection techniques based on sparse coding reconstruction.

[0004] 2. Anomaly Detection Based on Deep Learning: This category employs deep learning methods, primarily including anomaly detection techniques based on distance metrics, methods based on classification surface reconstruction, methods based on image reconstruction, and anomaly detection techniques based on knowledge distillation. For example, the method, apparatus, detection equipment, and storage medium for detecting defects in battery electrodes disclosed in Chinese Patent Publication No. CN109598721A is an example of an anomaly detection method based on deep learning.

[0005] During the research process, the following problems were found with the existing technologies: Traditional digital image-based methods for detecting particulate matter on electrode surfaces suffer from difficulties in designing artificial features and poor generalization. Furthermore, traditional methods have low detection accuracy, making it difficult to meet the requirements of modern production. Deep learning-based methods require a large amount of data. Due to the characteristics of electrode surface particulate matter and electrode production, there is a lack of suitable training datasets. The small size of the particles makes them difficult to detect, and real-time detection is required. A single deep learning method cannot meet these requirements.

[0006] Based on the above analysis, current technologies for detecting particulate matter on the surface of lithium-ion battery electrodes have significant limitations. Summary of the Invention

[0007] The technical problem to be solved by the present invention is that the existing technology for detecting particulate matter on the surface of lithium-ion battery electrode has low detection accuracy, lacks a dataset that can be used for training, and the particulate matter is small in size, making it difficult to detect and meet real-time requirements.

[0008] This invention solves the above-mentioned technical problems through the following technical means: a method for detecting particulate matter on the surface of battery electrodes, comprising the following steps:

[0009] Step 1: Construct a teacher-student network using teacher and student networks;

[0010] Step 2: Construct a twin neural network using the student network;

[0011] Step 3: Construct a sample set and perform data preprocessing;

[0012] Step 4: Train the teacher-student network using the sample set, and train the Siamese neural network using the sample set;

[0013] Step 5: Input the test sample into the trained teacher-student network to obtain the difference map output by the teacher-student network. If there is no defect in the difference map, the test result of the test sample is no abnormality; if there is a defect in the difference map, the test result of the test sample is abnormal. Input the test sample into the trained Siamese neural network to obtain the intersection map of the Siamese neural network. Take the intersection of the intersection map and the difference map to obtain the complete defect location.

[0014] Furthermore, the teacher network is a ResNet101, and the student network is a ResNeXt50.

[0015] Further, step one includes:

[0016] Select multiple teacher networks and the same number of student networks. Connect the teacher networks sequentially to form the first network module, and connect the student networks sequentially to form the second network module.

[0017] Furthermore, step two includes:

[0018] Two identical structures to the second network module are selected to form the third and fourth network modules, and the weights are shared between the third and fourth network modules.

[0019] Furthermore, step three includes:

[0020] A sample set is constructed, which includes positive samples, test samples, and small samples of particulate matter on the electrode surface. The positive samples are defect-free samples with no particulate matter on the electrode surface, the small samples of particulate matter on the electrode surface are defective samples with particulate matter on the electrode surface collected manually, and the test samples are electrode surface image samples collected in real time.

[0021] The sample set is preprocessed, including image size processing, image data format processing, and data augmentation.

[0022] Furthermore, the training of the teacher-student network using a sample set includes:

[0023] Freeze the parameters of the teacher network pre-trained on ImageNet, input positive samples into the first and second network modules of the teacher-student network, adjust the parameters of the second network module, train the teacher-student network, and stop training when the loss function loss_t is minimized, thus obtaining the trained teacher-student network.

[0024] Furthermore, the formula for the loss function loss_t is as follows:

[0025] loss t =I1loss1+I2loss2+I3loss3+I4loss4…+I k loss_k

[0026] Where loss_k is the loss of the feature map output by the k-th layer of the teacher-student network, k is the total number of layers in the teacher-student network, and I k I1, I2, I3, and I4 are the weights corresponding to the loss of the feature map output by the k-th layer of the teacher-student network, respectively.

[0027] In the first network module, the i-th teacher network *tech_i* and in the second network module, the i-th student network *stu_i* are the i-th layers of the teacher-student network, where i = 1, 2, 3, 4...k. Then, the loss *loss_i* of the feature map output by the i-th layer of the teacher-student network is...

[0028] loss_i = l ival +l idir

[0029]

[0030]

[0031] Among them, ι ival Let N be the loss value, j be the j-th neuron, and N be the loss value. j Let a be the number of neurons. j Let l be the activation value of the j-th neuron. idir For directional loss, vec(tech_i) represents the feature vector of the i-th teacher network tech_i in the first network module. T Let |vec(tech_i)|| represent the transpose of the eigenvector of the i-th teacher network tech_i in the first network module, and let ||vec(tech_i)|| represent the modulus of the eigenvector of the i-th teacher network tech_i in the first network module.

[0032] Furthermore, the teacher-student network has four layers, and the formula for the loss function loss_t is as follows:

[0033] loss_t=I1loss_1+I2loss_2+I3loss_3+I4loss_4.

[0034] Furthermore, the method of training the Siamese neural network using a sample set includes:

[0035] Small samples of polar particles are input into the third and fourth network modules of the Siamese neural network, respectively. The parameters of the Siamese neural network are continuously adjusted, and the Siamese neural network is iteratively trained. The loss function loss_l is calculated in each iteration until the loss function loss_l is minimized, at which point the iteration stops, and the trained Siamese neural network is obtained.

[0036] Furthermore, the formula for the loss function loss_l is as follows:

[0037]

[0038] in, This represents the Euclidean distance between the corresponding pixels in the output feature maps of the third and fourth network modules.

[0039] The advantages of this invention are:

[0040] (1) This invention integrates the detection results of the teacher-student network and the detection results of the Siamese neural network to improve the detection accuracy. It also constructs a sample set and uses the sample set to train the network, thus having a dataset that can be used for training. In addition, the detection of particulate matter on the electrode surface is performed through model training, which is easy to detect, has a short detection time, and improves the real-time performance of the detection.

[0041] (2) The teacher network parameters of this invention are derived from the large dataset ImageNet. The student network has a similar structure to the teacher network and only uses normal samples during training. Therefore, for normal samples, the teacher network and student network should have similar outputs. For defective samples, the teacher network has a high probability of correctly outputting defect information, while the student network, having never seen defective samples, will only output normal samples. Therefore, by subtracting the pixel differences of the different output images of the teacher network and student network for abnormal samples, a difference map can be obtained, which can identify most abnormal defects. The Siamese neural network has the same structure and shares weights, so its output for the same image should be extremely similar. Only small particulate samples are used during training. Therefore, when the test sample is a small particulate sample, the intersection of the two output images of the Siamese neural network is taken to obtain the intersection map, which can reveal the defect location. Finally, the intersection of the intersection map and the difference map is taken to obtain the complete defect location of the entire pole piece.

[0042] (3) This invention can make full use of the hierarchical features extracted by different network depths by weighting the loss between different depths of the teacher-student neural network; by grouping convolution to lighten the student network, the detection time is reduced, which meets the real-time requirements of lithium battery electrode surface particulate matter detection; since the twin neural network has the same weight structure, the training is easy to converge, which meets the small sample training requirements of lithium battery electrode surface particulate matter; and this invention adopts a two-stage detection method, which can combine the detection results of the teacher-student neural network and the twin neural network to improve the detection accuracy; at the same time, through knowledge distillation, it also has a good detection capability for small targets. Attached Figure Description

[0043] Figure 1 This is a flowchart of a method for detecting particulate matter on the surface of a battery electrode, as disclosed in an embodiment of the present invention.

[0044] Figure 2 This is a schematic diagram of the model structure in a method for detecting particulate matter on the surface of a battery electrode as disclosed in an embodiment of the present invention. Detailed Implementation

[0045] 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 in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, 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.

[0046] like Figure 1 As shown, the present invention provides a method for detecting particulate matter on the surface of battery electrodes, comprising the following steps:

[0047] S1. Construct a teacher-student network using teacher and student networks; the specific process is as follows:

[0048] Multiple teacher networks and an equal number of student networks are selected. The teacher networks are sequentially connected in series to form the first network module, and the student networks are sequentially connected in series to form the second network module. In this embodiment, there are four teacher networks, namely... Figure 2 There are four student networks, namely tech_1 to tech_4. Figure 2In the table, stu_1 to stu_4, the teacher network uses ResNet101, which is an existing technology (see He K, Zhang X, Ren S, et al. Deep Residual Learning for Image Recognition[J].IEEE, 2016.DOI:10.1109 / CVPR.2016.90.). The student network uses ResNeXt50, which is also an existing technology (see Xie S, Girshick R, Dollár, Piotr, et al. Aggregated Residual Transformations for DeepNeural Networks[J].IEEE, 2016.DOI:10.1109 / CVPR.2017.634.). The structure of the teacher-student network is shown in Table 1.

[0049] Table 1. Structure of the Teacher-Student Network

[0050]

[0051]

[0052] In the first network module, the i-th teacher network tech_i and the i-th student network stu_i in the second network module are the i-th layers of the teacher-student network. The output feature maps of each layer of the teacher-student network have the same dimension, as follows:

[0053] feature_maps=[(256,56,56),(512,28,28),(1024,14,14),,(2048,7,7)].

[0054] S2. Construct a Siamese neural network using the student network; the specific process is as follows:

[0055] The Siamese neural network is constructed using the student networks from step S1 above. Specifically, two networks with the same structure as the second network module are selected to form the third and fourth network modules. The third and fourth network modules share weights, thus requiring fewer samples to optimize the Siamese neural network. During subsequent training, a small sample of preprocessed polar particles is used to train the Siamese neural network, optimizing the loss function to approach zero. In this embodiment, both the third and fourth network modules include four student networks. The third network module is as follows: Figure 2 As shown in stu_5 to stu_8, the fourth network module is as follows: Figure 2 The numbers shown are stu_9 to stu_12.

[0056] S3. Construct a sample set and perform data preprocessing; the specific process is as follows:

[0057] A sample set is constructed, which includes positive samples, test samples, and small samples of particulate matter on the electrode surface. The positive samples are defect-free samples with no particulate matter on the electrode surface. The small samples of particulate matter on the electrode surface are defective samples with particulate matter on the electrode surface that are collected manually. Only a small number of samples are needed, which are easy to obtain. The test samples are real-time image samples of the electrode surface. The test samples are samples for which it is not known in advance whether there are particulate matter on the electrode surface. The final result can only be obtained by model detection.

[0058] The sample set is preprocessed, including image resizing, image data formatting, and data augmentation. In this embodiment, the image size is scaled to (256, 256, 3), and the scaled images are then standardized.

[0059] S4. Train the teacher-student network using the sample set, and train the Siamese neural network using the sample set; the process of training the teacher-student network using the sample set is as follows:

[0060] The teacher network uses a pre-trained model with parameters frozen during training. The student network uses a simpler structure similar to the teacher network, and is lightweighted using grouped convolutions to accelerate model training and improve the efficiency of the next step, student-student twin neural network detection. A loss function is calculated for each module of both the teacher and student networks, and weights are assigned to the loss function of each module to obtain the total loss of the teacher-student network. This loss function is then optimized to approach zero. Specifically... Figure 2 As shown, the parameters of the teacher network pre-trained on ImageNet are frozen. Positive samples are input into the first and second network modules of the teacher-student network. The parameters of the second network module are adjusted, and the teacher-student network is trained until the loss function loss_t is minimized, at which point training stops, resulting in the trained teacher-student network. The formula for the loss function loss_t is as follows:

[0061] loss t =I1loss1+I2loss2+I3loss3+I4loss4…+I k loss_k

[0062] Where loss_k is the loss of the feature map output by the k-th layer of the teacher-student network, k is the total number of layers in the teacher-student network, and I kI1, I2, I3, and I4 are the weights corresponding to the loss of the feature map output by the k-th layer of the teacher-student network, respectively. I1, I2, I3, and I4 are the weights corresponding to the loss of the feature maps output by the first to fourth layers of the teacher-student network, respectively. The i-th teacher network tech_i in the first network module and the i-th student network stu_i in the second network module are the i-th layers of the teacher-student network, i = 1, 2, 3, 4...k. In this embodiment, the teacher-student network has four layers. Therefore, the loss is calculated for the feature map output by each layer of the teacher-student network, and the loss of four feature maps is calculated in total. The formula of the loss function loss_t is loss_t = I1loss_1 + I2loss_2 + I3loss_3 + I4loss_4.

[0063] loss_i = l ival +l idir

[0064]

[0065]

[0066] Among them, l ival Value loss refers to the numerical loss between the teacher network output and the student network output on the feature map, i.e., the average difference in values ​​at the corresponding neurons, where j is the j-th neuron, and N... j Let a be the number of neurons. j Let l be the activation value of the j-th neuron. idir The directional loss is the similarity loss between the teacher and student neural networks, using cosine loss. vec(tech_i) represents the feature vector of the i-th teacher network *tech_i* in the first network module. T Let |vec(tech_i)|| represent the transpose of the eigenvector of the i-th teacher network tech_i in the first network module, and let ||vec(tech_i)|| represent the modulus of the eigenvector of the i-th teacher network tech_i in the first network module.

[0067] The process of training a Siamese neural network using a sample set is as follows:

[0068] When a small sample of polar particles from the same batch is simultaneously input into both the upper and lower branches of a Siamese neural network, each branch produces an output. The loss function is the Euclidean distance between the corresponding pixels in the feature maps of the two outputs. Specifically... Figure 2As shown, small samples of polar particulate matter are input into the third and fourth network modules of the Siamese neural network, respectively. The parameters of the Siamese neural network are continuously adjusted, and iterative training is performed. The loss function `loss_l` is calculated in each iteration until it is minimized, at which point the iteration stops, resulting in a trained Siamese neural network. The formula for the loss function `loss_l` is as follows:

[0069]

[0070] in, This represents the Euclidean distance between the corresponding pixels in the output feature maps of the third and fourth network modules.

[0071] S5. Input the test sample into the trained teacher-student network to obtain the difference map output by the teacher-student network. When there is no defect in the difference map, the test result of the test sample is no abnormality; when there is a defect in the difference map, the test result of the test sample is abnormal. Input the test sample into the trained Siamese neural network again to obtain the intersection map of the Siamese neural network. Take the intersection of the intersection map and the difference map to obtain the complete defect location.

[0072] The detection principle of this invention is as follows: The teacher network parameters are derived from the large dataset ImageNet. The student network has a similar structure to the teacher network and is trained using only normal samples. Therefore, for normal samples, the teacher network and the student network should have similar outputs. For defective samples, the teacher network has a high probability of correctly outputting defect information, while the student network, having never seen defective samples, will only output normal samples. Therefore, by subtracting the pixel differences between the different output images of the teacher network and the student network for abnormal samples, a difference map can be obtained, which can identify most abnormal defects. The Siamese neural network has the same structure and shares weights, so its output for the same image should be extremely similar. During training, only small particulate samples are used. Therefore, when the test sample is a small particulate sample, the intersection of the two output images of the Siamese neural network is taken to obtain the intersection map, which can reveal the complete defect location.

[0073] Through the above technical solutions, this invention provides a method for detecting particulate matter on the surface of lithium battery electrodes based on knowledge distillation and Siamese neural networks. By designing a teacher-student and Siamese neural network framework, designing a weighted loss function, and using grouped convolution to lightweight the student network, it achieves innovation in network model structure. The teacher-student network is trained with positive samples, and the resulting student network forms the Siamese neural network. Positive samples are used to train the teacher-student neural network, while small samples of electrode surface particulate matter are used to train the Siamese neural network. During detection, the detection samples are input into both the teacher-student network and the Siamese neural network, obtaining the detection output difference map of the teacher-student network and the detection output difference map of the Siamese neural network. The intersection of these difference maps yields the detection result for particulate matter on the electrode surface, thus improving the accuracy of the detection results.

[0074] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications 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.

Claims

1. A method for detecting particulate matter on the surface of battery electrodes, characterized in that, Includes the following steps: Step 1: Construct a teacher-student network using teacher and student networks; Step 2: Construct a twin neural network using the student network; Step 3: Construct a sample set and perform data preprocessing; Step 4: Train the teacher-student network using the sample set, and train the Siamese neural network using the sample set; The step of training the teacher-student network using a sample set includes: freezing the parameters of the teacher network pre-trained on ImageNet, inputting positive samples into the first and second network modules of the teacher-student network, adjusting the parameters of the second network module, and training the teacher-student network until the loss function is reached. Training stops when the loss function is minimized, resulting in a well-trained teacher-student network; The formula is as follows: in, For Teacher-Student Network The loss of the feature map output by the layer, The total number of layers in the teacher-student network. For Teacher-Student Network The weights corresponding to the loss of the feature map output by the layer. , , , These are the weights corresponding to the loss of the feature maps output from the first to the fourth layers of the teacher-student network; The i-th teacher network in the first network module With the i-th student network in the second network module Let i be the i-th layer of the teacher-student network, where i = 1, 2, 3, 4... Then the teacher-student network Loss of the feature map output by the layer for in, For value loss, For the first One neuron, The number of neurons. For the first The activation value of each neuron. For directional loss, This represents the i-th teacher network in the first network module. eigenvectors, This represents the i-th teacher network in the first network module. The transpose of the eigenvectors. This represents the i-th teacher network in the first network module. The magnitude of the eigenvectors; Step 5: Input the test sample into the trained teacher-student network to obtain the difference map output by the teacher-student network. If there is no defect in the difference map, the test result of the test sample is no abnormality; if there is a defect in the difference map, the test result of the test sample is abnormal. Input the test sample into the trained Siamese neural network to obtain the intersection map of the Siamese neural network. Take the intersection of the intersection map and the difference map to obtain the complete defect location.

2. The method for detecting particulate matter on the surface of a battery electrode according to claim 1, characterized in that, The teacher network is ResNet101, and the student network is ResNeXt50.

3. The method for detecting particulate matter on the surface of a battery electrode according to claim 1, characterized in that, Step one includes: Select multiple teacher networks and the same number of student networks. Connect the teacher networks sequentially to form the first network module, and connect the student networks sequentially to form the second network module.

4. The method for detecting particulate matter on the surface of a battery electrode according to claim 3, characterized in that, Step two includes: Two identical structures to the second network module are selected to form the third and fourth network modules, and the weights are shared between the third and fourth network modules.

5. The method for detecting particulate matter on the surface of a battery electrode according to claim 4, characterized in that, Step three includes: A sample set is constructed, which includes positive samples, test samples, and small samples of particulate matter on the electrode surface. The positive samples are defect-free samples with no particulate matter on the electrode surface, the small samples of particulate matter on the electrode surface are defective samples with particulate matter on the electrode surface collected manually, and the test samples are electrode surface image samples collected in real time. The sample set is preprocessed, including image size processing, image data format processing, and data augmentation.

6. The method for detecting particulate matter on the surface of a battery electrode according to claim 1, characterized in that, The teacher-student network has four layers, and the loss function... The formula is as follows: 。 7. The method for detecting particulate matter on the surface of a battery electrode according to claim 5, characterized in that, The method of training a Siamese neural network using a sample set includes: Small samples of polar particulate matter were input into the third and fourth network modules of the Siamese neural network, respectively. The parameters of the Siamese neural network were continuously adjusted, and iterative training was performed. The loss function was calculated in each iteration. The value, until the loss function When the value is minimized, stop iterating to obtain the trained Siamese neural network.

8. The method for detecting particulate matter on the surface of a battery electrode according to claim 7, characterized in that, The loss function The formula is as follows: in, This represents the Euclidean distance between the corresponding pixels in the output feature maps of the third and fourth network modules.

Images