Retired lithium ion battery rapid sorting method based on electrochemical impedance spectroscopy and semi-parameter clustering algorithm

Abstract

The invention discloses a rapid sorting method for retired lithium ion batteries based on an electrochemical impedance spectroscopy and a semi-parameter clustering algorithm, which comprises the following steps of: firstly, performing electrochemical impedance spectroscopy test and relaxation time distribution analysis on the retired lithium ion batteries to obtain electrochemical impedance spectroscopy data and relaxation time distribution peak value data; and then, extracting features from the electrochemical impedance spectroscopy data by using a convolutional auto-encoder, extracting features from relaxation time distribution peak data by using a full-connection neural network, carrying out feature splicing on the two features, and inputting the spliced features into a deep neural network to quickly obtain the estimated capacity of the retired battery. And finally, taking the ohmic internal resistance, the impedance characteristic information and the estimated capacity of the retired lithium ion battery extracted from the electrochemical impedance spectrum data and the relaxation time distribution peak value data as the input of a semi-parameter clustering algorithm, and carrying out rapid sorting on the retired battery. According to the method, the capacity of the decommissioned lithium ion battery can be quickly predicted, outliers of the decommissioned battery are eliminated, and the decommissioned battery is quickly sorted.

Description

Technical Field

[0001] The present invention belongs to the field of sorting retired batteries, and more specifically, relates to a rapid sorting method for retired lithium-ion batteries based on electrochemical impedance spectroscopy and semi-parametric clustering algorithm. Background Art

[0003] Given the excellent energy density, long cycle life, and environmental friendliness of lithium-ion batteries, they have been widely used in the field of electric vehicles. These batteries, as the power core and key energy storage components of electric vehicles, play a decisive role in the durability and safety of the vehicles. However, during the daily operation of electric vehicles, the batteries inevitably undergo the test of gradual aging, fluctuations in environmental temperature, and repeated charge and discharge cycles, which together induce a series of irreversible electrochemical changes. These changes include the depletion of lithium ions, the loss of active substances, and the damage of the electrode structure, resulting in the gradual degradation of battery performance. According to the provisions of the national quality standard GB / T 61484-2015, when the battery capacity drops to 70% to 80% of its initial value, it is considered that the battery has reached the end of its service life, and at this time, the battery is no longer suitable to continue to be used as a power source in the vehicle.

[0004] Although they cannot meet the performance requirements of electric vehicles, retired batteries still have 70% - 80% of their initial capacity and have objective use value. If not properly handled, these batteries may cause environmental pollution and waste of resources. However, if these batteries are used in fields with low performance requirements, such as power batteries for two-wheeled and three-wheeled electric vehicles or other lithium-ion battery-powered vehicles with low requirements for driving range, they can also be used for energy storage, for power supply to households or outdoors. This can not only improve the utilization efficiency of resources but also bring significant environmental and economic benefits.

[0005] However, the inherent time-consuming nature of the traditional charge and discharge test process leads to a long time for obtaining the capacity and impedance of retired batteries, and there is a lack of effective means to quickly predict the capacity and impedance of retired batteries; the differences in the use conditions, cycle loads, and management systems of retired batteries result in diverse aging paths of retired batteries; the multi-dimensional characteristic data of retired batteries have high heterogeneity, which limits the sorting effect based on traditional clustering methods and makes it difficult to achieve precise sorting. Summary of the Invention

[0006] (1) Technical Problems to be Solved

[0007] Based on this, the present invention provides a rapid sorting method for retired lithium-ion batteries based on electrochemical impedance spectroscopy and semi-parametric clustering algorithm to solve the technical problems mentioned in the background art, namely, the long time for sorting retired batteries and the difficulty in achieving precise sorting of retired batteries due to diverse aging paths.

[0008] (2) Technical solution

[0009] To achieve the above object, the present invention provides a rapid sorting method for retired lithium-ion batteries based on electrochemical impedance spectroscopy and semi-parametric clustering algorithm, including:

[0010] Step S1: Perform electrochemical impedance spectroscopy test and relaxation time distribution analysis on the retired battery to obtain the electrochemical impedance spectroscopy data and relaxation time distribution peak data of the retired battery;

[0011] Step S2: Use a convolutional autoencoder to extract features from the electrochemical impedance spectroscopy data to obtain the electrochemical impedance spectroscopy data feature F e , and use a fully connected neural network to extract features from the relaxation time distribution peak data to obtain the relaxation time distribution peak data feature F d ;

[0012] Step S3: Perform feature splicing on the electrochemical impedance spectroscopy data feature F e and the relaxation time distribution peak data feature F d to obtain the spliced feature F c ;

[0013] Step S4: Input the spliced feature F c into a deep neural network to obtain the estimated capacity of the retired battery;

[0014] Step S5: Perform weight processing and normalization on the estimated capacity of the retired battery and the ohmic internal resistance R0, contact resistance R c , SEI film resistance R SEI , charge transfer resistance R ct and diffusion resistance R d extracted from the electrochemical impedance spectroscopy data and the relaxation time distribution peak data, and perform semi-parametric clustering as multi-dimensional input;

[0015] The steps of the semi-parametric clustering algorithm are as follows:

[0016] First, divide the data into a high-density core region and a low-density edge region according to density;

[0017] The data is the estimated capacity, ohmic internal resistance R0, contact resistance R c , SEI film resistance R SEI , charge transfer resistance R ct and diffusion resistance R dMap the data to the feature space, evaluate the local density of the data using cosine similarity, and analyze the aggregation degree of the data in the feature space through the kernel density estimation method; after sorting the density values from low to high, divide the data into a high-density core region and a low-density edge region according to a certain ratio;

[0018] Secondly, process the data in the high-density core region;

[0019] Then, process the data in the low-density edge region;

[0020] Step S6: Analyze the results of the semi-parametric clustering and compare them with other clustering algorithms to verify the sorting results.

[0021] (III) Beneficial effects

[0022] As can be seen from the above technical solutions, a rapid sorting method for retired lithium-ion batteries based on electrochemical impedance spectroscopy and semi-parametric clustering algorithm proposed by the present invention has the following beneficial effects:

[0023] 1. Efficiently extract and splice electrochemical impedance spectroscopy data and relaxation time distribution peak data, and can quickly and effectively estimate the remaining capacity of retired batteries.

[0024] 2. By analyzing the impedance characteristics of retired batteries, key health indicators such as contact impedance, SEI film impedance, charge transfer impedance, and lithium-ion diffusion impedance are extracted, which can reflect the aging mode and state of the battery, and then achieve accurate sorting.

[0025] 3. Adopt the semi-parametric clustering method, evaluate the density of data points through cosine similarity, adaptively adjust the number of clusters and the corresponding clustering results in the high-density core region of the data through the clustering algorithm combining the growing self-organizing mapping network and kernel fuzzy C-means; in the low-density edge region of the data, set an adaptive cluster threshold to screen outliers and assign data. This method has good clustering indicators and a relatively fast sorting speed, can accurately identify batteries in different aging states, and effectively improves the accuracy and efficiency of battery sorting. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] The features and advantages of the present invention will be more clearly understood by referring to the accompanying drawings. The drawings are schematic and should not be construed as limiting the present invention in any way. In the drawings:

[0027] Figure 1 is a schematic flow chart of the rapid sorting of retired lithium-ion batteries based on electrochemical impedance spectroscopy and semi-parametric clustering algorithm of the present invention;

[0028] Figure 2 is the overall architecture diagram of the battery capacity estimation neural network based on electrochemical impedance spectroscopy and relaxation time analysis of the present invention;

[0029] Figure 3 This is the result graph of the rapid sorting of retired lithium-ion batteries based on electrochemical impedance spectroscopy and semi-parametric clustering algorithm in the embodiments of the present invention;

[0030] Figure 4 This is the comparison graph of clustering metrics between the semi-parametric clustering algorithm of the present invention and other common clustering algorithms. Detailed implementation manners

[0031] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Apparently, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0032] The present invention provides a rapid sorting method for retired lithium-ion batteries based on electrochemical impedance spectroscopy and semi-parametric clustering algorithm, as Figure 1 shown, which specifically includes:

[0033] Step S1: Perform electrochemical impedance spectroscopy test and relaxation time distribution analysis on the retired battery to obtain the electrochemical impedance spectroscopy data and relaxation time distribution peak data of the retired battery.

[0034] Among them, the method for performing electrochemical impedance spectroscopy test on the retired battery is as follows:

[0035] In a constant temperature environment, place the retired battery after standing in an impedance measurement device. By applying small-amplitude sinusoidal voltage excitation signals with different frequencies to the retired battery respectively, measure the corresponding response current signals of the retired battery. Then perform FTT transformation on the small-amplitude sinusoidal voltage excitation signal with the same frequency and its corresponding response current signal, and calculate the impedance value of the retired battery at the same frequency, so as to obtain the impedance values of the retired battery at different frequencies. The electrochemical impedance spectroscopy data of the retired battery includes the frequency and the impedance at that frequency. In this embodiment, the electrochemical impedance spectroscopy test is performed at 60 different frequencies within the frequency range of 0.02 - 20 kHz.

[0036] The method for performing electrochemical impedance spectroscopy test on the retired battery is as follows:

[0037] Perform relaxation time distribution analysis on the impedances of the retired battery at different frequencies to obtain a time characteristic curve (i.e., DRT curve). The relaxation time distribution peak data includes the peak and peak position of the DRT curve.

[0038] Step S2: Use a convolutional autoencoder to extract features from the electrochemical impedance spectroscopy data to obtain electrochemical impedance spectroscopy data feature F e , and use a fully connected neural network to extract features from the peak data of the relaxation time distribution to obtain peak data feature F of the relaxation time distribution d .

[0039] Among them, step S2 specifically includes the following steps:

[0040] Step S201: Use a convolutional autoencoder to extract features from the electrochemical impedance spectroscopy data to obtain electrochemical impedance spectroscopy data feature F c .

[0041] Input the electrochemical impedance spectroscopy data into the convolutional autoencoder, and the input vector is as follows:

[0042] x = [real1, real2,..., real 60 , imag1, imag2,..., imag 60 (1)

[0043] Among them, the subscript represents the frequency number; real represents the real part of the impedance, and imag represents the imaginary part of the impedance; for example, real1 represents the real part of the impedance obtained by performing electrochemical impedance spectroscopy test at the 1st frequency, and imag1 represents the imaginary part of the impedance obtained by performing electrochemical impedance spectroscopy test at the 1st frequency.

[0044] Then, normalize each component (i.e., input feature) x of the input vector p :

[0045]

[0046] Among them, μ represents the average value of the input feature, and σ represents the standard deviation of the input feature.

[0047] The convolutional autoencoder is a two-dimensional convolutional neural network, and its calculation process is as follows:

[0048]

[0049] Among them, x i+m-1,j+n-1 is the value of the input feature at the position (i + m, j + n), h m,n is the weight of the convolutional kernel at the position (m, n), b is the bias of the convolutional neural network, and σ(·) is the activation function.

[0050] Step S202: Use a fully connected neural network to extract features from the peak data of the relaxation time distribution to obtain peak data feature F of the relaxation time distribution d .

[0051] Input the peak data of the relaxation time distribution into a fully connected neural network. The input vector is as follows:

[0052] y = [τ1,...,τ5,peak1,...,peak5] (4)

[0053] Among them, the subscript represents the number of the peak point of the DRT curve; τ represents the peak position, and peak represents the peak value. For example, τ1 represents the peak position of the 1st peak point, and peak1 represents the peak value of the 1st peak point.

[0054] Encode the peak data of the relaxation time distribution using a fully connected neural network. The calculation process is as follows:

[0055] F d = Wy + b (5)

[0056] Among them, W is the weight matrix of the fully connected layer, and b is the bias vector.

[0057] Step S3: Concatenate the features of the electrochemical impedance spectroscopy data F e and the features of the peak data of the relaxation time distribution F d to obtain the concatenated feature F c .

[0058] Obtain the deep features F e of the electrochemical impedance spectroscopy data through a convolutional autoencoder, and obtain the features F d of the peak data of the relaxation time distribution through a fully connected neural network. F e and F d merge the influence of the two features through feature concatenation, so as to obtain a more compact, more expressive and higher-correlation concatenated feature F c with the capacity of the retired lithium-ion battery.

[0059] Step S4: Input the concatenated feature F c into a deep neural network to obtain the estimated capacity of the retired battery.

[0060] The concatenated feature F c is a feature vector obtained by concatenating the deep features F e of the electrochemical impedance spectroscopy data and the peak features F d of the relaxation time distribution, as follows:

[0061] F C = Contact(F e ,F d ) (6)

[0062] Input the concatenated feature F cInput a deep neural network with multiple stacked fully connected layers, which is mapped to the battery capacity label. Its calculation formula is as follows:

[0063]

[0064] Among them, is the output of the deep neural network, representing the predicted value of the retired battery capacity. LinearLayers(·) represents multiple stacked fully connected layers. The calculation method of each fully connected layer is as follows:

[0065] y o =σ(W i x i +b i ) (8)

[0066] Among them, x i , W i and b i are the input, weight, and bias of a single fully connected layer respectively, and y o is the output of a single fully connected layer.

[0067] Use the Huber loss function with δ = 1 to calculate the loss between the network output and the training set capacity label Cap label , as follows:

[0068]

[0069] Use the Adam optimization algorithm to perform backpropagation iterative optimization on the parameters of the deep neural network. The initial learning rate is set to 0.03. After every 100 iterations, the learning rate is reduced to half of the current value. After 500 rounds of training, the training process of retired battery capacity prediction is completed. After the deep neural network is trained, fix the model parameters of the deep neural network, input the electrochemical impedance spectrum data and the peak data of the relaxation time distribution of the retired battery, perform retired battery capacity prediction, and obtain the estimated capacity of the retired battery. The overall architecture diagram of the battery capacity estimation neural network based on electrochemical impedance spectrum and relaxation time analysis in this paper is as Figure 2 shown.

[0070] Step S5: Weight process and normalize the estimated capacity of the retired battery and the ohmic internal resistance R0, contact impedance R c , SEI film impedance R SEI , charge transfer impedance R ct , and diffusion impedance R d extracted from the electrochemical impedance spectrum data and the peak data of the relaxation time distribution of the retired battery, and use them as multi-dimensional inputs for semi-parametric clustering.

[0071] Based on the electrochemical impedance spectroscopy data of the retired battery obtained in step S1, a Nyquist plot is drawn; the ohmic internal resistance R0 of the retired battery is calculated from the intersection point of the impedance curve and the horizontal axis in the Nyquist plot.

[0072] Parameter identification is performed on the DRT curve obtained in step S1 to obtain the impedance characteristic information of the retired battery: the leftmost peak in the DRT curve is usually related to the contact resistance R c and reflects the interfacial contact situation between the electrode and the current collector; the second peak is related to the SEI film impedance R SEI and characterizes the formation and stability of the SEI layer; the third and fourth peaks are mainly related to the charge transfer resistance R ct and describe the speed of the electrochemical reaction kinetics on the electrode surface; the rightmost peak is related to the diffusion impedance R d and reflects the diffusion behavior of lithium ions in the bulk of the electrode material.

[0073] Weight processing and normalization are performed on the performance indicators of the retired battery (including the estimated capacity, ohmic internal resistance R0, contact resistance R c , SEI film impedance R SEI , charge transfer resistance R ct and diffusion impedance R d ) and used as the multi-dimensional input of the semi-parametric clustering algorithm. Among them, the weights of the estimated capacity, ohmic internal resistance R0, contact resistance R c , SEI film impedance R SEI , charge transfer resistance R ct and diffusion impedance R d are 0.5, 0.1, 0.1, 0.1, 0.1, 0.1 respectively.

[0074] The steps of the semi-parametric clustering algorithm are as follows:

[0075] First, the data distribution is divided into a high-density core region and a low-density edge region according to the density;

[0076] The data (i.e., the performance indicators of the retired battery after weight processing and normalization) are mapped to the feature space, the cosine similarity is used to evaluate the local density of the data, and the kernel density estimation method is used to analyze the aggregation degree of the data in the feature space. After sorting the density values from low to high, the data are divided into a high-density core region and a low-density edge region according to a certain ratio. The high-density part represents the main clustering data, while the low-density part may contain noise or transitional state samples.

[0077] Secondly, the data in the high-density core region are processed; specifically, it includes:

[0078] Step 1: Use a growing self-organizing map network (i.e., growing SOM) to train the data in the high-density core region, and use the node positions obtained from the training as the initial clustering centers;

[0079] The growing SOM is a dynamically expanding neural network structure that can gradually increase the neuron nodes according to the distribution of the input data, so as to more flexibly adapt to complex data. The network starts as a small grid and gradually expands the nodes until it covers the data distribution. Adjacent nodes maintain topological continuity during the training process, forming a structure that matches the data morphology. The node positions obtained through the training of the growing SOM are used as the initial clustering centers, providing a basis for subsequent clustering analysis.

[0080] Step 2: Initialize the membership matrix using the initial clustering centers, and use the kernel fuzzy C-means algorithm to iteratively optimize it to obtain the final membership matrix and clustering centers.

[0081] The kernel fuzzy C-means algorithm initializes the membership matrix with the above initial clustering centers, maps the data to a high-dimensional space through the Gaussian kernel function, and adaptively adjusts the clustering centers and memberships until the objective function converges. Specifically, it includes:

[0082] Step 1: Parameter initialization;

[0083] Set the number of clusters C (the number of clusters), select the fuzzy coefficient m, and initialize the membership matrix U = [u ij n×c and satisfy (the sum of the memberships of each data sample is 1). Select the kernel function Gaussian kernel φ(x) to map the data to a high-dimensional space, and the corresponding kernel matrix is where x i , x j represent two data samples in the original input space, σ represents the kernel width, and ‖·‖ is the square of the Euclidean distance in the original input space.

[0084] Step 2: Calculate the clustering centers in the kernel mapping space;

[0085] In the kernel space (i.e., the high-dimensional feature space), calculate the virtual center of each cluster through membership weighting. In the high-dimensional space, the clustering centers cannot be explicitly represented, but the weights can be implicitly calculated through the kernel function as follows:

[0086]

[0087] where, V j represents the implicit representation of the jth clustering center in the kernel space after the tth iteration, u ij is the membership of the data point x i to the jth cluster, m represents the fuzzy coefficient, K(x​i , V j ) represents the kernel function value, and n represents the total number of data samples.

[0088] Step 3: Update the membership matrix;

[0089] According to the current cluster centers, recalculate the membership of each data point to each cluster. The membership update formula calculates the similarity between the data point and the center through the kernel function, as shown below:

[0090]

[0091] where C is the number of clusters.

[0092] Step 4: Iterative optimization;

[0093] Repeat Step 2 and Step 3 until the termination condition is met (such as the membership change is less than the threshold, or the maximum number of iterations is reached).

[0094] Step 5: Output the results to obtain the final membership matrix and cluster centers.

[0095] Then, process the data in the low-density marginal area. Specifically, it includes:

[0096] Step 1: Set the adaptive cluster threshold;

[0097] According to the clustering results of the data in the high-density core area, calculate the radius of each cluster, and dynamically adjust the cluster threshold in combination with the sample density within the cluster (the cluster threshold refers to the threshold for data points in the low-density marginal area to enter each cluster). Clusters with higher density adopt higher thresholds, while clusters with lower density adopt lower thresholds to enhance the adaptability to different distribution characteristics.

[0098] Specifically, for each determined cluster, take the maximum cosine distance from all data points in the cluster to the cluster center as the radius. Then calculate the coverage area of the cluster through the square of the radius; count the ratio of the number of data points within the cluster to the coverage area to obtain the cluster density, that is, cluster density = number of data points within the cluster / coverage area of the cluster.

[0099] The cluster threshold adjustment rule is: if the cluster is a high-density cluster (i.e., cluster density > preset cluster density threshold), then the cluster threshold is adjusted to 0.8 times the radius; if the cluster is a low-density cluster (i.e., cluster density ≤ preset cluster density threshold), then the cluster threshold is adjusted to 1.2 times the radius.

[0100] Step 2: Screen out outliers and assign data.

[0101] Calculate the cosine similarity between the data points in each low-density edge region and the cluster centers of all clusters; map the cosine similarity to the interval [0, 1] (1 indicates completely in the same direction, and 0 indicates orthogonal irrelevance); if the cosine similarity of a data point is lower than the corresponding cluster threshold in all clusters, it is marked as an outlier.

[0102] For non-outliers, select the cluster with the highest cosine similarity as the belonging cluster. Specifically, first calculate the attenuation coefficient, and the formula is as follows:

[0103]

[0104] Among them, the actual similarity is obtained by normalizing the cosine similarity between the data point and a certain cluster; the maximum similarity is 1.

[0105] If the attenuation coefficient > 0.5, assign the data point a complete membership degree with the cluster (i.e., the membership degree is 1.0);

[0106] If the attenuation coefficient ≤ 0.5, proportionally reduce the membership degree of the data point with the cluster (i.e., the membership degree is 0.3 - 0.7).

[0107] Then select the cluster with the highest membership degree as the belonging cluster of the data point. Decaying its membership degree weight according to the similarity ratio can avoid the excessive influence of edge data on the core cluster structure.

[0108] By performing different processes on the data in the high-density core region and the low-density edge region, the number of clusters and the clustering results for the sorting of retired batteries are obtained.

[0109] Step S6: Analyze the results of the semi-parametric clustering and compare them with other clustering algorithms to verify the sorting results.

[0110] In this embodiment, 100 lithium cobalt oxide / graphite button batteries with different aging states are selected, the number of clusters is 6, and the clustering results are as Figure 3 shown. Among them, the red crosses represent the cluster centers of each class, and the black pentagrams represent the battery outliers.

[0111] As Figure 4 shown, four indicators, namely DBI (Davies-Bouldin Index), DI (Dunn Index), SC (Silhouette Coefficient), and CHI (Calinski-Harabasz Index), are used to evaluate the sorting results of the semi-parametric clustering algorithm of the present invention and other clustering algorithms. Specifically, compared with GMM, FCM, DBSCAN, and SOM, the semi-parametric clustering algorithm of the present invention has the lowest DBI value, the highest DI value, a relatively high SC value, and the highest CHI value, proving that the method proposed by the present invention has significant advantages in the overall quality of clustering and can be applied to the rapid sorting of retired lithium-ion batteries.

[0112] As described above, it is only the specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention can easily think of changes or substitutions, which should all be covered within the protection scope of the present invention.

Claims

1. A rapid sorting method for retired lithium-ion batteries based on electrochemical impedance spectroscopy and semi-parametric clustering algorithm, characterized in that Including: Step S1: Conduct electrochemical impedance spectroscopy testing and relaxation time distribution analysis on the retired battery to obtain the electrochemical impedance spectroscopy data and relaxation time distribution peak data of the retired battery; Step S2: Extract features from the electrochemical impedance spectroscopy data using a convolutional autoencoder to obtain the electrochemical impedance spectroscopy data feature F e , and extract features from the peak data of the relaxation time distribution using a fully connected neural network to obtain the peak data feature F of the relaxation time distribution d ; Step S3: For the electrochemical impedance spectrum data feature F e and the relaxation time distribution peak data feature F d perform feature splicing to obtain the spliced feature F c ; Step S4: Input the splicing feature F c into a deep neural network to obtain the estimated capacity of the retired battery; Step S5: Weight and normalize the estimated capacity of the retired battery with the ohmic internal resistance R0, contact resistance R c , SEI film resistance R SEI , charge transfer resistance R ct and diffusion resistance R d extracted from the electrochemical impedance spectroscopy data and the peak data of the relaxation time distribution, and perform semi-parametric clustering with them as multi-dimensional inputs; The steps of the semi-parametric clustering algorithm are as follows: First, divide the data into a high-density core region and a low-density edge region according to density; The data are the estimated capacity, ohmic internal resistance R0, contact impedance R c , SEI film impedance R SEI , charge transfer impedance R ct and diffusion impedance R d of retired batteries that have been weighted and normalized; map the data to the feature space, evaluate the local density of the data using cosine similarity, and analyze the degree of aggregation of the data in the feature space through kernel density estimation; after sorting the density values from low to high, divide the data into a high-density core region and a low-density edge region according to a certain proportion; Second, process the data in the high-density core region; Then, process the data in the low-density edge region; Step S6: Analyze the results of the semi-parametric clustering and compare them with other clustering algorithms to verify the sorting results.

2. The method according to claim 1, wherein Processing the data in the high-density core region includes: Step 1: Use a growing self-organizing mapping network to train the data in the high-density core region, and use the obtained node positions as the initial clustering centers; Step 2: Initialize the membership matrix using the initial clustering centers, and use the kernel fuzzy C-means algorithm for iterative optimization to obtain the final membership matrix and clustering centers.

3. The method according to claim 2, characterized in that, Processing the data in the low-density edge region includes: Step 1: Set an adaptive cluster threshold; According to the clustering results of the data in the high-density core region, for each determined cluster, take the maximum cosine distance from all data points in the cluster to the cluster's clustering center as the radius; then calculate the coverage area of the cluster through the square of the radius; count the ratio of the number of data points in the cluster to the coverage area to obtain the cluster density, that is, cluster density = number of data points in the cluster / coverage area of the cluster; The cluster threshold adjustment rule is: if the cluster is a high-density cluster, that is, the cluster density > the preset cluster density threshold, then the cluster threshold is adjusted to 0.8 times the radius; if the cluster is a low-density cluster, that is, the cluster density ≤ the preset cluster density threshold, then the cluster threshold is adjusted to 1.2 times the radius; Step 2: Screen outliers and assign data; Calculate the cosine similarity between each data point in the low-density edge region and the clustering centers of all clusters; map the cosine similarity to the [0, 1] interval, where 1 represents completely in the same direction and 0 represents orthogonally irrelevant; if the cosine similarity of a data point is lower than the corresponding cluster threshold in all clusters, it is marked as an outlier; For non-outliers, select the cluster with the highest cosine similarity as the belonging cluster; specifically, first calculate the attenuation coefficient, and the formula is as follows: where the actual similarity is obtained by normalizing the cosine similarity between the data point and a certain cluster; the maximum similarity is 1; If the attenuation coefficient > 0.5, then assign a complete membership degree between the data point and the cluster, that is, the membership degree is 1.0; If the attenuation coefficient ≤ 0.5, reduce the membership degree between the data point and the cluster proportionally, that is, the membership degree is 0.3 - 0.7; Then select the cluster with the highest membership degree as the belonging cluster of the data point.

4. The method according to claim 2, wherein The kernel fuzzy C-means algorithm includes: Step 1: Parameter initialization; Set the number of clusters C, select the fuzzy coefficient m, and initialize the membership matrix U = [u ij n×c , and satisfy Select the kernel function Gaussian kernel φ(x) to map the data to a high-dimensional space, and the corresponding kernel matrix is where x i , x j represent two data samples in the original input space, σ represents the kernel width, and ‖·‖ is the square of the Euclidean distance in the original input space;​ Step 2: Calculate the clustering centers in the kernel mapping space; In the kernel space, calculate the virtual center of each cluster through membership degree weighting; in the high-dimensional space, the clustering centers cannot be explicitly represented, and the weights can be implicitly calculated through the kernel function as follows: Among them, V j represents the implicit representation of the j-th cluster center in the kernel space after the t-th iteration, and u ij is the membership degree of the data point x i to the j-th cluster, m represents the fuzzy coefficient, and K(x i , V j ) represents the kernel function value, and n represents the total number of data samples; Step 3: Update the membership matrix; According to the current clustering centers, recalculate the membership degree of each data point to each cluster; the membership degree update formula needs to calculate the similarity between the data point and the center through the kernel function as follows: Among them, C is the number of clusters; Step 4: Iterative optimization; Repeat Step 2 and Step 3 until the termination condition is met, such as the membership degree change is less than the threshold, or the maximum number of iterations is reached; Step 5: Output the results to obtain the final membership degree matrix and cluster centers.

5. The method according to claim 1, wherein The electrochemical impedance spectroscopy test is carried out at 60 different frequencies within the frequency range of 0.02 - 20 kHz.

6. The method according to claim 5, wherein Step S2 includes the following steps: Step S201: Extract features from the electrochemical impedance spectroscopy data using a convolutional autoencoder to obtain the electrochemical impedance spectroscopy data feature F c ; Input the electrochemical impedance spectroscopy data into the convolutional autoencoder, and the input vector is as follows: x = [real1, real2,..., real 60 , imag1, imag2,..., imag 60 ​ Among them, the subscript represents the frequency number; real represents the real part of the impedance, and imag represents the imaginary part of the impedance; for example, real1 represents the real part of the impedance obtained by the electrochemical impedance spectroscopy test at the 1st frequency, and imag1 represents the imaginary part of the impedance obtained by the electrochemical impedance spectroscopy test at the 1st frequency; Then, normalize each component x of the input vector x p as follows: Among them, μ represents the average value of the input features, and σ represents the standard deviation of the input features; The convolutional autoencoder is a two-dimensional convolutional neural network, and its calculation process is as follows: where x i+m-1,j+n-1 is the value of the input feature at position (i + m, j + n), h m,n is the weight of the convolutional kernel at position (m, n), b is the bias of the convolutional neural network, and σ(·) is the activation function; Step S202: Use a fully connected neural network to extract features from the peak data of the relaxation time distribution to obtain the peak data feature F of the relaxation time distribution d ; Input the relaxation time distribution peak data into the fully connected neural network, and the input vector is as follows: y = [τ1,..., τ5, peak1,..., peak5] Among them, the subscript represents the number of the peak point of the DRT curve; τ represents the peak position, and peak represents the peak value; for example, τ1 represents the peak position of the 1st peak point, and peak1 represents the peak value of the 1st peak point; Encode the relaxation time distribution peak data with the fully connected neural network, and the calculation process is as follows: F d = Wy + b Among them, W is the weight matrix of the fully connected layer, and b is the bias vector.

7. The method according to claim 1, wherein In step S4, the deep neural network uses the Huber loss function with δ = 1 to calculate the network output and the loss with the training set capacity label Cap label is as follows: Adopt the Adam optimization algorithm to perform backpropagation iterative optimization on the parameters of the deep neural network. The initial learning rate is set to 0.

03. Every 100 iterations, the learning rate is reduced to half of the current value. After 500 rounds of training, the training process of retired battery capacity prediction is completed.

8. The method according to claim 1, wherein The estimated capacity, ohmic internal resistance R0, contact impedance R c , SEI film impedance R SEI , charge transfer impedance R ct and diffusion impedance R d have weights of 0.5, 0.1, 0.1, 0.1, 0.1, and 0.1 respectively.

9. The method according to claim 1, characterized in that, In Step S6, four clustering evaluation indicators, namely the DBI (Davies-Bouldin Index), DI (Dunn Index), SC (Silhouette Coefficient), and CHI (Calinski-Harabasz Index), are used to analyze the sorting results of the clustering algorithm.

Images