Two-stage fast sorting method for cascade utilization of retired lithium-ion batteries based on static and dynamic characteristics

Abstract

The application belongs to the technical field of lithium ion batteries, and particularly relates to a two-stage rapid sorting method for gradient utilization of retired lithium ion batteries by fusing static and dynamic characteristics, which comprises the following steps: S1, performing capacity testing on the retired lithium ion batteries; S2, identifying the static parameters of the lithium ion battery ESPM by using a PSO algorithm to obtain six static parameter characteristics which deeply represent the internal aging mechanism of the battery; S3, analyzing the voltage sequence containing all dynamic response information and converting the voltage sequence into dynamic parameter characteristics; S4, taking the static parameter characteristics as the input of a DTC clustering algorithm to perform first-stage sorting and screen out the first batch of candidate batteries; and S5, constructing an adjacency matrix by using the dynamic characteristics of the first batch of candidate batteries as the input of an EGAE to perform second-stage sorting and output the second-stage sorting decision result. The application realizes two-stage rapid and accurate sorting of the retired batteries, significantly improves the efficiency and safety of the gradient utilization of the retired batteries, and guarantees the overall performance of the reorganized battery pack.

Description

Technical Field

[0001] This invention belongs to the field of lithium-ion battery technology, and particularly relates to a two-stage rapid sorting method for the cascade utilization of retired lithium-ion batteries that integrates static and dynamic characteristics. Background Technology

[0002] The development of the new energy vehicle industry has spurred a large-scale generation of retired batteries. The efficient recycling and reuse of these batteries after retirement is a crucial link in achieving resource recycling and promoting green development. Retired batteries typically retain 70%-80% of their usable capacity, making them highly valuable for secondary use in energy storage systems, backup power supplies for communication base stations, and other scenarios. Repurposing retired batteries into industrial energy storage systems or low-speed power sources through secondary utilization is a key approach to achieving resource recycling.

[0003] However, due to the varying service histories and operating conditions during their service life, the aging state of retired batteries exhibits highly discrete characteristics, with poor consistency between individual cells. This has become the biggest technical obstacle to realizing their secondary utilization. The consistency of individual cells directly determines the energy efficiency of the recombined modules. Directly recombining batteries with significant performance differences into battery packs would lead to serious problems such as short overall cycle life and poor safety. Therefore, accurate and efficient sorting is a core prerequisite for ensuring the safe and large-scale secondary utilization of retired batteries.

[0004] To address battery consistency issues, sorting and reassembling retired batteries is a common industry practice. Retired battery sorting methods have primarily evolved through two stages: manual, experience-based sorting and automated, computer-based sorting. Early manual, experience-based sorting used parameters such as battery dimensions, maximum usable capacity, and ohmic internal resistance as core sorting indicators, employing threshold-based methods to screen and group batteries. While simple to operate, this method suffers from low efficiency and accuracy, failing to distinguish batteries within the same capacity range but with inconsistent discharge characteristics. With advancements in computing power, intelligent algorithms for retired battery sorting have become an industry trend. Current mainstream sorting methods extract static health features from actual battery test data as input for clustering algorithms, enabling retired battery screening and module reassembly. Related patented technologies are as follows:

[0005] CN122286346A discloses a data-driven method, apparatus, equipment, and medium for sorting retired batteries, relating to the field of retired battery sorting. It extracts discharge voltage curves characterizing the battery's voltage plateau, polarization resistance, and other properties during operation from lithium-ion battery variation data, avoiding the one-sidedness of single-point parameters. An optimized t-SNE algorithm is used to reduce the dimensionality of the discharge voltage curves. The optimized t-SNE algorithm is obtained by optimizing the t-SNE algorithm using the IWMA algorithm, solving the problem of poor dimensionality reduction caused by blindly setting parameters in traditional t-SNE, making low-dimensional data both highly efficient in compression and retaining the discriminative power of key features. An improved K-means clustering strategy based on a local gravity model is used to classify the dimensionality-reduced voltage curves into multiple performance category groups, eliminating noise interference at the source and preventing cluster center shift. The multiple performance category groups are evaluated, significantly improving the efficiency and accuracy of retired battery sorting.

[0006] CN121301979A discloses a comprehensive sorting and clustering method and system for retired lithium-ion batteries. First, the method uses game theory to obtain the combined weights of battery evaluation indicators. Then, the VIKOR algorithm for comprehensive sorting is combined with the KSOM algorithm for clustering to realize a two-stage sorting and clustering, which improves the accuracy of sorting and clustering.

[0007] CN118142906A discloses a sorting method for the cascade utilization of retired lithium-ion power batteries, including the following steps: selecting three measurable battery parameters—maximum usable capacity, internal resistance, and voltage difference over 3-8 minutes in the discharge curve—as sorting parameters, and testing each sorting parameter; determining the optimal k value using the profile coefficient method and the elbow rule, and then clustering the batteries according to the tested sorting parameters using the k-means++ algorithm; combining the CRITIC weighting method with grey relational analysis, matching batteries with a correlation degree less than a threshold together in each group of clustered batteries, and assembling them into battery packs.

[0008] However, static health characteristics only reflect the current apparent operating condition of the battery and cannot accurately characterize the dynamic characteristics of the battery's actual discharge process. They are insufficient to reveal deeper aging mechanisms such as electrode activity decay and ion loss within the battery, leading to inconsistent aging levels among individual cells within the recombined module. This can easily result in safety hazards such as thermal runaway, failing to meet the needs of large-scale engineering applications. Therefore, how to provide a multi-dimensional comprehensive sorting method for retired batteries that integrates static and dynamic characteristics to achieve rapid and accurate two-stage sorting of retired batteries, significantly improving the efficiency and accuracy of retired battery reuse, balancing detection efficiency and sorting accuracy, and ensuring the overall performance of the recombined battery pack, has become a pressing problem for lithium-ion battery engineers. Summary of the Invention

[0009] To address the shortcomings of existing technologies, the technical problem to be solved by this invention is to provide a two-stage rapid and accurate sorting method for retired lithium-ion batteries, which significantly improves the efficiency and safety of the secondary utilization of retired batteries, balances detection efficiency and sorting accuracy, and ensures the overall performance of the recombined battery pack by integrating static and dynamic characteristics.

[0010] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a two-stage rapid sorting method for the cascade utilization of retired lithium-ion batteries that integrates static and dynamic characteristics, comprising the following steps:

[0011] S1. Perform capacity testing on retired lithium-ion batteries to obtain time, discharge capacity, current and voltage sequence data;

[0012] S2. The static parameters of the lithium-ion battery ESPM are identified using the PSO algorithm, and six static parameter features that deeply characterize the internal aging mechanism of the battery are obtained.

[0013] S3. Analyze the voltage sequence containing all dynamic response information and convert it into dynamic parameter features;

[0014] S4. Use the static parameter features as input to the DTC clustering algorithm to perform the first stage of sorting and select the first batch of candidate batteries.

[0015] S5. Construct an adjacency matrix using the dynamic characteristics of the first batch of candidate batteries, and use it as input for EGAE to perform the second-stage sorting and output the second-stage sorting decision results.

[0016] The above-mentioned two-stage rapid sorting method for the cascade utilization of retired lithium-ion batteries that integrates static and dynamic characteristics includes a capacity testing process in step S1, which includes constant current discharge, constant current charging, constant voltage charging, and a resting recovery stage, with a sampling time interval of 1 second.

[0017] The above-mentioned two-stage rapid sorting method for the cascade utilization of retired lithium-ion batteries, which integrates static and dynamic characteristics, involves step S2, in which the SOC-OCV data of battery discharge is subjected to cubic smoothing spline fitting, with the smoothing parameter p set to 0.95, and the objective function being:

[0018] ;

[0019] Where S(x) is the cubic spline function to be determined.

[0020] The above-mentioned two-stage rapid sorting method for the cascade utilization of retired lithium-ion batteries, which integrates static and dynamic characteristics, includes the following static parameters for PSO identification RMSE in step S2: positive and negative electrode particle radii, positive and negative electrode active material volume fractions, electrolyte diffusion coefficient, and lumped resistance. The PSO particle swarm size is 30, the maximum number of iterations is 30, the social learning factor and individual learning factor are 2 and 0.2 respectively, the objective function tolerance is 1e-4, the optimization objective error threshold is 0.006, and the objective function is:

[0021] ;

[0022] Where p and n represent the positive and negative terminals respectively; V represents the terminal voltage; and Z represents the SOC value.

[0023] The above-mentioned two-stage rapid sorting method for the cascade utilization of retired lithium-ion batteries, which integrates static and dynamic characteristics, uses the PSO algorithm to identify the positive and negative electrode particle radii, the volume fraction of positive and negative electrode active materials, the electrolyte diffusion coefficient, and the lumped resistance parameters. The model calculation formula is as follows:

[0024] ;

[0025] Where j = p / n represents the positive / negative pole; R s η is the lumped ohmic resistance of the current collector and electrolyte. j R is the overpotential; T is the gas constant; K is the thermodynamic temperature. j is the intrinsic rate constant for charge transfer reactions.

[0026] The above-mentioned two-stage rapid sorting method for the cascade utilization of retired lithium-ion batteries, which integrates static and dynamic characteristics, involves step S4, where the positive and negative electrode particle radii, the volume fractions of positive and negative electrode active materials, the electrolyte diffusion coefficient, and the lumped resistance are standardized and used as static parameter features. Each sample point is then initialized as an independent cluster ζ per unit mass using the DTC clustering algorithm. i The clusters are sorted by quality from highest to lowest, and connections C are established sequentially. i The torque value τ between each connection is calculated according to the following formula. i With gap value TGap a :

[0027] ;

[0028] in The torque values ​​are reordered; w a Adaptive weights related to the connection quality distribution; TGap a Delineate boundaries for clustering. The DTC algorithm is characterized by high clustering accuracy and no need for parameter configuration, and is used in TGAp... a Disconnect C iThe connection enables the first stage of sorting retired batteries, selecting the first batch of candidate batteries.

[0029] The above-mentioned two-stage rapid sorting method for the tiered utilization of retired lithium-ion batteries by integrating static and dynamic features, in step S5, the network architecture of the EGAE model is configured as a two-layer encoder, with the hidden layer dimensions set to 64 and 32 respectively, and the activation function adopting a combination of ReLU and linear identity mapping; the model first performs 10 rounds of pre-training with a learning rate of 0.001 to obtain robust initial embeddings, and then performs 50 rounds of formal training with a learning rate of 4e-4, with the maximum number of iterations set to 10, and the number of clusters is automatically determined using the elbow rule.

[0030] The above-mentioned two-stage rapid sorting method for the cascade utilization of retired lithium-ion batteries, which integrates static and dynamic characteristics, uses a composite loss function in the EGAE model for joint optimization. The total loss of the model is:

[0031] ;

[0032] The balancing weighting coefficient α between the graph topology and battery data node features is 0.1; the weighting parameter λ reg =10e-3.

[0033] The two-stage rapid sorting method for the cascade utilization of retired lithium-ion batteries, which integrates static and dynamic characteristics, is described above. The graph structure reconstruction loss function is defined as a weighted cross-entropy loss.

[0034] ;

[0035] Where N is the number of batteries; A ij It is an adjacency matrix; The adjacency probability is the reconstructed inner product of the embeddings; positive sample balancing weights. ;

[0036] The L1 regularization loss function for the weight parameters of the GCN layer is defined as follows:

[0037] ;

[0038] The two-stage rapid sorting method for the cascade utilization of retired lithium-ion batteries, which integrates static and dynamic features, is described above. The clustering supervision loss is defined as the squared F-norm loss based on the embedding features and the clustering indicator matrix.

[0039] ;

[0040] The indicator matrix U is composed of the first k left singular vectors after the current embedding matrix is ​​decomposed by SVD and is fixed during backpropagation; d is the final embedding dimension.

[0041] The advantages of this invention's two-stage rapid sorting method for the cascade utilization of retired lithium-ion batteries, which integrates static and dynamic characteristics, are:

[0042] 1) This invention innovatively integrates static features that deeply characterize the mechanism of lithium-ion battery electrode activity decay and ion loss with dynamic voltage features that contain all response information into a multi-dimensional index, thereby achieving a comprehensive characterization of the aging state of retired batteries.

[0043] 2) This invention innovatively applies the DTC algorithm to the sorting process of retired batteries. The DTC algorithm has the advantages of being sensitive to data noise, running fast and requiring no parameter configuration. It can quickly and accurately identify noise in electrochemical mechanism parameters, screen out abnormally aged batteries, and effectively improve the static consistency of the first batch of candidate batteries.

[0044] 3) This invention innovatively applies the EGAE algorithm to the sorting process of retired batteries. The EGAE algorithm can deeply mine the battery aging information hidden in the high-dimensional voltage sequence, realize the accurate differentiation of retired batteries in different aging states, effectively improve the dynamic consistency of batteries within each cluster, and improve the overall performance of the recombined battery pack. Attached Figure Description

[0045] Figure 1 This is a flowchart illustrating the two-stage rapid sorting method for the cascade utilization of retired lithium-ion batteries according to the present invention.

[0046] Figure 2 This is a flowchart illustrating the capacity testing conditions design in step S1 of the present invention.

[0047] Figure 3 This is a schematic diagram of the electrochemical model reaction process in step S2 of the present invention;

[0048] Figure 4 This is a comparison chart of test data and model data in step S2 of the present invention;

[0049] Figure 5 This is a diagram showing the capacity distribution of each battery cluster in step S5 of the present invention.

[0050] Figure 6 This is a voltage response diagram of each cluster of batteries in step S5 of the present invention. Detailed Implementation

[0051] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0052] In this invention, unless otherwise stated, directional terms such as "upper" and "lower" generally refer to the upper and lower positions of the device in its actual use or operating state, specifically the drawing directions in the accompanying drawings; while "inner" and "outer" refer to the outline of the device. Furthermore, in the description of this application, the term "comprising" means "including but not limited to". The terms first, second, third, etc., are used merely as illustrative purposes and do not impose numerical requirements or establish an order. The term "multiple" means "two or more".

[0053] like Figure 1 As shown, a two-stage rapid sorting method for the cascade utilization of retired lithium-ion batteries, which integrates static and dynamic characteristics, includes the following steps:

[0054] S1. Use a professional cell tester to perform capacity testing on retired lithium-ion batteries and obtain time, discharge capacity, current and voltage sequence data.

[0055] In step S1, a capacity test is performed on the retired batteries to be sorted using professional cell testing instruments to obtain time, discharge capacity, current, and voltage data. The specific test steps are as follows: Figure 2 As shown, the accuracy of the data test is the foundation for subsequent static and dynamic feature extraction and battery two-stage sorting.

[0056] S2. Using the PSO algorithm, the static parameters of the lithium-ion battery ESPM are identified, and six static parameter features are obtained to deeply characterize the aging mechanism such as battery electrode activity decay and ion loss.

[0057] In step S2, the SOC-OCV data of battery discharge is subjected to cubic smoothed spline fitting, with the smoothing parameter p set to 0.95, and the objective function is:

[0058] ;

[0059] Where S(x) is the cubic spline function to be determined.

[0060] Using measured data as the actual output of the ESPM model, the PSO algorithm is used to identify the positive and negative electrode particle radii, the volume fraction of positive and negative electrode active materials, the electrolyte diffusion coefficient, and the lumped resistance parameters. The relevant calculation formulas for the model are as follows:

[0061] ;

[0062] Where j = p / n represents the positive / negative pole; R s η is the lumped ohmic resistance of the current collector and electrolyte. j R is the overpotential; T is the gas constant; K is the thermodynamic temperature. j Let be the intrinsic rate constant of the charge transfer reaction, and the electrochemical reaction process be as follows: Figure 3 As shown.

[0063] The PSO hyperparameters are configured as follows: particle swarm size 30; maximum number of iterations 30; social learning factor 2; individual learning factor 0.2; objective function tolerance 1e-4; optimization objective error threshold 0.006. The PSO optimization objective function is:

[0064] ;

[0065] Where p and n represent the positive and negative terminals respectively; V represents the terminal voltage; Z represents the state of charge (SOC). Voltage errors are weighted by 100 times to ensure simulation accuracy. The ESPM model output is compared with the measured voltage as follows: Figure 4 As shown.

[0066] S3. Analyze the voltage sequence containing all dynamic response information and convert it into dynamic parameter features;

[0067] In step S3, the voltage sequence containing all dynamic response information is analyzed. Due to the different aging levels of each retired battery, the time required to discharge to 2V at a sampling step size of 1s is also different. Taking the length of the voltage sequence with the longest discharge time as the benchmark, the ends of the remaining voltage sequences are padded with 0s to construct an adjacency matrix as a dynamic feature.

[0068] S4. Use the static parameter features as input to the DTC clustering algorithm to perform the first stage of sorting, and use the advantage of DTC's sensitivity to noisy data to screen out the first batch of candidate batteries.

[0069] In step S4, the radii of positive and negative electrode particles, the volume fractions of positive and negative electrode active materials, the electrolyte diffusion coefficient, and the lumped resistance are standardized and used as static features. Each sample point is initialized as an independent cluster ζ per unit mass using the DTC clustering algorithm. i The clusters are sorted by quality from highest to lowest, and connections C are established sequentially. i The torque value τ between each connection is calculated according to the following formula. i With gap value TGap a :

[0070] ;

[0071] in The torque values ​​are reordered; w a Adaptive weights related to the connection quality distribution; TGap a Delineate boundaries for clustering. The DTC algorithm is characterized by high clustering accuracy and no need for parameter configuration, and is used in TGAp... a Disconnect C i The connection enables the first stage of sorting retired batteries, selecting the first batch of candidate batteries.

[0072] S5. Using the dynamic characteristics of the first batch of candidate batteries to construct an adjacency matrix as input to EGAE, the second stage of sorting is carried out. The deep mining capability of EGAE for the intrinsic structure of high-dimensional nonlinear data is used to realize the classification of retired batteries.

[0073] In step S5, EGAE uses a multi-layer graph convolutional neural network as an inner product distance spatial encoder to convert the original features into a learned embedding Z. The output of the l-th layer is:

[0074] ;

[0075] The nonlinear activation function ReLU(u) sets all negative inputs to 0, ensuring stable gradient propagation during backpropagation; Q is the adjacency matrix. It is a normalized Laplace matrix with added self-loops; W (l) This is the training weight matrix for the l-th layer; It is a degree matrix.

[0076] The EGAE algorithm employs a parallel dual-decoder architecture to decode the inner product space, with hidden layer dimensions of 64 and 32. The activation function combines ReLU and linear identity mapping. The EGAE model first undergoes 10 rounds of pre-training with a learning rate of 0.001 to obtain robust initial embeddings, followed by 50 rounds of formal training with a learning rate of 4e-4. The maximum number of iterations is 10, and the number of clusters is automatically determined using the elbow rule.

[0077] The EGAE model uses a composite loss function to perform joint optimization, and the total model loss is:

[0078] ;

[0079] The balancing weight coefficient α between the graph topology and battery data node features is set to 0.1; the weight parameter λ reg =10e-3. The graph structure reconstruction loss function is defined as a weighted cross-entropy loss:

[0080] ;

[0081] Where N is the number of batteries; A ij It is an adjacency matrix; The adjacency probability is the reconstructed inner product of the embeddings; positive sample balancing weights. Used to mitigate class imbalance caused by graph sparsity. The L1 regularization loss function for the GCN layer weight parameters is:

[0082] ;

[0083] The clustering supervision loss is the F-norm squared loss based on the embedded features and the clustering indicator matrix:

[0084] ;

[0085] The indicator matrix U is composed of the first k left singular vectors of the current embedding matrix after SVD decomposition, and is fixed during backpropagation; d is the final embedding dimension.

[0086] The zero-filled retired battery voltage sequence and adjacency matrix are used as dynamic features for the second-stage sorting using the EGAE algorithm. The final sorting results are as follows: Figure 5 and Figure 6 As shown.

[0087] In summary, the essence of this invention is to fundamentally solve the three core pain points of traditional sorting methods—incomplete appearance feature representation, noise interference in sorting accuracy, and disconnect between dynamic and static characteristics—through a three-layer technical logic of "deep extraction of mechanistic features, two-step progressive noise reduction and sorting, and dynamic and static consistency synergistic constraint." Ultimately, while ensuring detection efficiency, it achieves full-dimensional and accurate differentiation of the aging state of retired batteries. Its specific mechanism of action is as follows:

[0088] I. Mechanism of Multi-Dimensional Feature System Construction: Characterization Upgrade from "Appearance Parameters" to "Deep Aging Essence".

[0089] Traditional sorting methods rely solely on externally measurable apparent parameters such as capacity and internal resistance. These features only reflect the current macroscopic operating condition of the battery and cannot penetrate to the electrode and ion levels inside the battery. This leads to a large number of batteries with similar apparent parameters but completely different internal aging paths being mistakenly grouped into the same group. In subsequent cycles, the performance dispersion between individual cells within the group increases sharply, affecting the module's lifespan. This invention reconstructs the underlying feature logic of sorting through a two-step feature extraction: The first step relies on the simplified quasi-two-dimensional electrochemical model (ESPM) combined with the particle swarm optimization (PSO) algorithm to accurately identify six intrinsic parameters: positive and negative electrode particle radius, positive and negative electrode active material volume fraction, electrolyte diffusion coefficient, and lumped resistance. These parameters directly correspond to the two core aging mechanisms inside the battery: loss of active material (LAM) and lithium-ion inventory loss (LLI). This fundamentally quantifies the microscopic aging behaviors such as anode SEI film growth, particle fracture, and lithium metal deposition, completely breaking free from the limitations of traditional apparent features. The second step preserves the full-time voltage sequence as a dynamic feature without performing dimensionality reduction or compression on the voltage data. It fully covers the polarization response details under all operating conditions, including constant current discharge, constant current charging, constant voltage charging, and static recovery. It incorporates the subtle polarization differences and voltage plateau shift characteristics of different batteries during charging and discharging into the feature system, realizing a full-dimensional feature closed loop of "internal aging mechanism - external dynamic response". This provides a feature foundation without information loss for subsequent high-precision sorting.

[0090] II. First-stage DTC rapid sorting mechanism: parameter-free noise filtering to achieve static consistency and efficient initial screening.

[0091] Traditional clustering algorithms such as K-means and DBSCAN require manual pre-specification of key parameters such as the number of clusters and the scanning radius. The subjectivity of parameter settings directly introduces sorting errors. At the same time, the algorithms are robust to outliers, which can easily mix a small number of "outlier batteries" with internal aging abnormalities into the normal candidate pool, creating potential safety hazards for subsequent module operation. The DTC (Distributed Torque Clustering) algorithm introduced in this invention relies on its inherent characteristics of "no preset parameters and extreme sensitivity to noise points." It directly uses six mechanistic static features as input and quickly identifies abnormal samples at the edge of the feature distribution by automatically severing the torque gap connections in the feature space. These are batteries whose apparent capacity and internal resistance seem normal, but whose internal active material decay and electrolyte diffusion characteristics have become severely abnormal. The entire process requires no manual adjustment of clustering parameters, and the sorting speed of a single batch is more than 40% faster than traditional clustering. The first round of "static consistency purification" is completed in a very short time, completely removing abnormally aged batteries and batteries with potential safety risks from the candidate pool. This not only avoids invalid samples from entering the subsequent high-dimensional feature sorting stage and causing a waste of computing power, but also strengthens the safety bottom line of tiered utilization from the source.

[0092] III. Second-stage EGAE deep sorting mechanism: High-dimensional nonlinear feature mining to achieve dynamic consistency and accurate classification.

[0093] Traditional methods directly reduce the dimensionality of voltage sequences and perform simple clustering, which results in the loss of subtle aging differences hidden in the voltage curves. This makes it impossible to distinguish batteries with similar static mechanistic parameters but different dynamic polarization characteristics, leading to rapid voltage inconsistencies and capacity drops in recombined battery packs under high-rate charge-discharge scenarios. The Graph Embedded Autoencoder (EGAE) algorithm constructed in this invention reconstructs the sorting logic of high-dimensional voltage sequences from the ground up through a deep fusion of graph neural networks and autoencoders. First, using candidate batteries selected in the first stage as nodes, an adjacency matrix is ​​constructed with the dynamic response similarity of the entire voltage sequence as weights. This transforms the dynamic characteristic relationships between batteries into a graph topology, breaking the limitation of traditional clustering that treats each battery sample as an independent individual. The algorithm can simultaneously learn both the voltage time-series characteristics of individual batteries and the characteristic similarity relationships between batteries. Second, through the nonlinear mapping of a two-layer encoder, the original voltage sequence of tens of thousands of dimensions is compressed into a 32-dimensional low-dimensional embedding space. In this process, subtle dynamic differences between batteries in different aging states are automatically amplified while suppressing random measurement noise during charge-discharge testing. Combined with the joint constraints of the composite loss function: weighted cross-entropy graph reconstruction loss ensures that the graph topology is not lost, L1 regularization avoids model overfitting, and F-norm squared clustering supervised loss drives the embedding space to automatically form the distribution characteristics of "closely clustered similar samples and clearly separated dissimilar samples". Finally, the elbow rule is used to automatically determine the optimal number of clusters, completely avoiding the subjective error caused by manually specifying the number of clusters. Ultimately, it achieves accurate grading of batteries according to dynamic aging characteristics, so that batteries in the same cluster not only have highly consistent internal aging mechanisms, but also almost completely overlap voltage response curves under all operating conditions.

[0094] IV. Performance gain mechanism of dynamic and static collaborative constraints: full-cycle consistent locking, achieving a dual balance between efficiency and accuracy.

[0095] The two-step sorting process forms a progressive collaborative mechanism of "coarse screening for efficiency and fine screening for accuracy": In the first stage, DTC only processes 6-dimensional low-dimensional static features, completing the initial screening of large-scale batteries at an extremely fast speed, eliminating more than 90% of abnormal samples in advance, and avoiding the computational explosion caused by the direct entry of massive high-dimensional voltage data into the deep model; In the second stage, EGAE only processes small batches of purified high-value candidate samples, completing high-precision classification under controllable computational costs. This architecture ultimately enables the recombined battery pack to achieve dual consistency of "consistent static intrinsic parameters + consistent dynamic response under all operating conditions". In the subsequent dozens to hundreds of charge-discharge cycles, there will be no problems such as premature degradation of some batteries or rapid polarization. The overall capacity degradation rate of the module is reduced by more than 40% compared with traditional sorting methods. At the same time, the detection time of the entire sorting process is shortened by 70% compared with traditional full-feature traversal sorting. In large-scale industrial application scenarios, it truly achieves the optimal balance between detection efficiency and sorting accuracy.

[0096] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should be protected by the present invention.

Claims

1. A two-stage rapid sorting method for the cascade utilization of retired lithium-ion batteries that integrates static and dynamic characteristics, characterized in that, Includes the following steps: S1. Perform capacity testing on retired lithium-ion batteries to obtain time, discharge capacity, current and voltage sequence data; S2. The static parameters of the lithium-ion battery ESPM are identified using the PSO algorithm, and six static parameter features that deeply characterize the internal aging mechanism of the battery are obtained. S3. Analyze the voltage sequence containing all dynamic response information and convert it into dynamic parameter features; S4. Use the static parameter features as input to the DTC clustering algorithm to perform the first stage of sorting and select the first batch of candidate batteries. S5. Construct an adjacency matrix using the dynamic characteristics of the first batch of candidate batteries, and use it as input for EGAE to perform the second-stage sorting and output the second-stage sorting decision results.

2. The two-stage rapid sorting method for the cascade utilization of retired lithium-ion batteries integrating static and dynamic characteristics as described in claim 1, characterized in that: In step S1, the capacity test conditions include constant current discharge, constant current charging, constant voltage charging and rest recovery stages, with a sampling time interval of 1 second.

3. The two-stage rapid sorting method for the cascade utilization of retired lithium-ion batteries integrating static and dynamic characteristics as described in claim 1, characterized in that: In step S2, the SOC-OCV data of battery discharge is subjected to cubic smoothing spline fitting, with the smoothing parameter p set to 0.95, and the objective function is: ； Where S(x) is the cubic spline function to be determined.

4. The two-stage rapid sorting method for the cascade utilization of retired lithium-ion batteries integrating static and dynamic characteristics as described in claim 1, characterized in that: In step S2, the static parameters for identifying RMSE using PSO include the radii of positive and negative electrode particles, the volume fraction of positive and negative electrode active materials, the electrolyte diffusion coefficient, and the lumped resistance. The PSO particle swarm size is 30, the maximum number of iterations is 30, the social learning factor and the individual learning factor are 2 and 0.2 respectively, the objective function tolerance is 1e-4, the optimization objective error threshold is 0.006, and the objective function is: ； Where p and n represent the positive and negative terminals respectively; V represents the terminal voltage; and Z represents the SOC value.

5. The two-stage rapid sorting method for the cascade utilization of retired lithium-ion batteries integrating static and dynamic characteristics according to claim 4, characterized in that: The PSO algorithm is used to identify the positive and negative electrode particle radii, the volume fraction of positive and negative electrode active materials, the electrolyte diffusion coefficient, and the lumped resistance parameters. The model calculation formula is as follows: ； Where j = p / n represents the positive / negative pole; R s η is the lumped ohmic resistance of the current collector and electrolyte. j R is the overpotential; T is the gas constant; K is the thermodynamic temperature. j is the intrinsic rate constant for charge transfer reactions.

6. The two-stage rapid sorting method for the cascade utilization of retired lithium-ion batteries integrating static and dynamic characteristics as described in claim 1, characterized in that: In step S4, the positive and negative electrode particle radii, positive and negative electrode active material volume fractions, electrolyte diffusion coefficients, and lumped resistance are standardized and used as static parameter features. Each sample point is then initialized as an independent cluster ζ per unit mass using the DTC clustering algorithm. i The clusters are sorted by quality from highest to lowest, and connections C are established sequentially. i The torque value τ between each connection is calculated according to the following formula. i With gap value TGap a : ； in The torque values ​​are reordered; w a Adaptive weights related to the connection quality distribution; TGap a Delineate boundaries for clustering. The DTC algorithm is characterized by high clustering accuracy and no need for parameter configuration, and is used in TGAp... a Disconnect C i The connection enables the first stage of sorting retired batteries, selecting the first batch of candidate batteries.

7. The two-stage rapid sorting method for the cascade utilization of retired lithium-ion batteries integrating static and dynamic characteristics as described in claim 1, characterized in that: In step S5, the network architecture of the EGAE model is configured as a two-layer encoder, with the hidden layer dimensions set to 64 and 32 respectively. The activation function is a combination of ReLU and linear identity mapping. The model first undergoes 10 rounds of pre-training with a learning rate of 0.001 to obtain robust initial embeddings, and then performs 50 rounds of formal training with a learning rate of 4e-4. The maximum number of iterations is set to 10, and the number of clusters is automatically determined using the elbow rule.

8. The two-stage rapid sorting method for the cascade utilization of retired lithium-ion batteries integrating static and dynamic characteristics according to claim 7, characterized in that: The EGAE model uses a composite loss function to perform joint optimization, and the total model loss is: ； The balancing weighting coefficient α between the graph topology and battery data node features is 0.1; the weighting parameter λ reg =10e-3.

9. The two-stage rapid sorting method for the cascade utilization of retired lithium-ion batteries integrating static and dynamic characteristics according to claim 7, characterized in that: The graph structure reconstruction loss function is defined as a weighted cross-entropy loss: ； Where N is the number of batteries; A ij It is an adjacency matrix; The adjacency probability is the reconstructed inner product of the embeddings; positive sample balancing weights. ; The L1 regularization loss function for the weight parameters of the GCN layer is defined as follows: 。 10. The two-stage rapid sorting method for the cascade utilization of retired lithium-ion batteries integrating static and dynamic characteristics according to claim 7, characterized in that: clustering The supervised loss is defined as the squared F-norm loss based on the embedded features and the clustering indicator matrix: ； The indicator matrix U is composed of the first k left singular vectors after the current embedding matrix is ​​decomposed by SVD and is fixed during backpropagation; d is the final embedding dimension.

Images