A sorting method for the secondary use of retired lithium-ion batteries

Abstract

This invention relates to the technical field of sorting retired lithium-ion batteries, and provides a sorting method for the cascade utilization of retired lithium-ion batteries, comprising the following steps: safety inspection of the retired batteries to be sorted; electrochemical impedance spectroscopy testing under preset conditions to obtain raw impedance data and construct an initial feature vector; parsing the initial feature vector to convert it into relaxation time distribution data to construct a multimodal feature set containing real data and enhanced data; inputting the multimodal feature set into a Transformer-MLSAE network, performing deep feature distillation and EIS-DRT feature fusion to extract a deep fused feature vector; classifying the batteries into performance level labels according to cascade utilization standards; training a regularized logistic regression classification model based on the deep fused feature vector and performance level labels to obtain a retired battery sorting model; obtaining the deep fused feature vector and inputting it into the trained retired battery sorting model to obtain the predicted performance level of the battery for cascade utilization.

Description

Technical Field

[0001] This invention relates to the technical field of sorting retired lithium-ion batteries, and more particularly to a sorting method for the cascade utilization of retired lithium-ion batteries. Background Technology

[0002] Current sorting technologies for retired batteries rely heavily on Remaining Capacity (RUL) or State of Health (SOH) as single criteria. These static indicators fail to reveal complex degradation mechanisms within the battery, such as lithium-ion inventory loss, and cannot effectively predict the nonlinear degradation risks that may occur in future cycles. This leads to the misclassification of "pseudo-healthy" batteries that currently have acceptable capacity but are potentially decaying extremely rapidly. Meanwhile, traditional full charge-discharge testing is time-consuming and energy-intensive, making it difficult to meet the needs of large-scale industrial applications. Although there have been attempts to improve detection speed using electrochemical impedance spectroscopy (EIS), directly using raw impedance data faces problems such as severe overlap of frequency domain features and ambiguity of microscopic mechanisms. Furthermore, the scarcity of high-quality samples due to the highly heterogeneous historical operating conditions of retired batteries makes existing data-driven models prone to overfitting under small sample conditions, resulting in insufficient generalization ability. In addition, EIS is extremely sensitive to temperature: it is difficult to ensure completely consistent temperatures during on-site testing of retired batteries. The impedance spectra of the same battery under different ambient temperatures will exhibit systematic shifts, thus affecting the stability of feature extraction and grade determination. On the other hand, DRT is obtained by impedance spectrum inversion, which is an ill-conditioned inverse problem. Measurement noise and missing frequency points will be amplified by the inversion process. If there is a lack of noise reduction and regularization, it will easily lead to DRT peak distortion and feature drift, reducing sorting reliability. Summary of the Invention

[0003] The purpose of this invention is to provide a sorting method for the cascade utilization of retired lithium-ion batteries, aiming to solve the problems of poor model robustness and insufficient feature extraction under sample data.

[0004] To address the aforementioned technical problems, a sorting method for the secondary utilization of retired lithium-ion batteries is provided, comprising the following steps:

[0005] S1. The retired batteries to be sorted undergo appearance and voltage safety inspections, and abnormally damaged battery cells are removed.

[0006] S2. Battery samples that have passed the safety inspection are subjected to short-time electrochemical impedance spectroscopy under preset conditions to obtain raw impedance data and preprocess it to construct an initial feature vector.

[0007] S3. The initial feature vector is parsed to convert it into relaxation time distribution data, and the data is augmented using a pre-built variational autoencoder generation model to construct a multimodal feature set containing real data and augmented data.

[0008] S4. Input the multimodal feature set into the Transformer-MLSAE network, use the self-attention mechanism to capture the full spectrum dependency, and perform deep feature distillation and EIS-DRT feature fusion to extract a highly robust deep fusion feature vector.

[0009] S5. Select a portion of battery samples for full life cycle testing to obtain their true remaining life value, and classify the batteries into different performance level labels according to the preset tiered utilization standards.

[0010] S6. Based on the extracted deep fusion feature vector and the performance level label, supervise the training of the regularized logistic regression classification model to obtain the trained retired battery sorting model.

[0011] S7. Obtain the deep fusion feature vector of the retired battery to be sorted, input it into the trained retired battery sorting model, and obtain the prediction result of the battery's tiered utilization performance level.

[0012] Further, in step S4, let the Transformer-MLSAE network consist of a Transformer backbone network and a multi-layer sparse autoencoder (MLSAE); let the Transformer backbone network have L layers. For the input enhanced multimodal feature set, in the l-th layer of the Transformer backbone network, Let its output residual flow vector be denoted as The , That is, the dimension is The real vector space; extract the residual flow set of all layers. As input to the MLSAE.

[0013] Further, in step S4, a multilayer sparse autoencoder (MLSAE) is constructed: a sparse autoencoder with encoder and decoder parameters shared between different layers is constructed; the multilayer sparse autoencoder (MLSAE) generates the output residual flow vector of each layer of the Transformer backbone network. A unified process is performed, using the same set of parameters to map features from different layers to the same latent space.

[0014] Furthermore, the encoding process of the multilayer sparse autoencoder (MLSAE) involves mapping the residual flow vector of the l-th layer to a high-dimensional sparse latent space:

[0015] ;

[0016] The decoding process of the multilayer sparse autoencoder (MLSAE) involves reconstructing the first layer through latent features. Layer input vector:

[0017] ;

[0018] in: It is the first The potential characteristics of the layer; and That is, the dimension of the latent features is much larger than the dimension of the input, which is used to construct an overcomplete representation of the features to capture subtle differences in battery aging. It is a shared encoding weight matrix. It is the encoding bias. It is the decoding bias; It is the first The reconstructed vector of the layer, It is a shared decoding weight matrix.

[0019] Furthermore, the training objective of the multilayer sparse autoencoder (MLSAE) is to minimize the residual flow vector reconstruction error across all layers and maximize the sparsity of the latent features, defining the total loss function. The total loss function is minimized during the training process to update the parameters of the encoder and decoder. The first term of the total loss function forces the model parameters to adjust in a direction that reduces reconstruction error, while the second term, through L1 regularization, forces non-critical elements in the latent feature vector to tend towards zero to achieve sparsity. The total loss function is defined as follows:

[0020] ;

[0021] in: This indicates that the summation is performed over all layers of the Transformer backbone network, and that the model is jointly optimized over the residual flow vectors of all layers. This represents the square of the L2 norm, used to calculate the reconstruction loss; This represents the L1 norm, used to calculate the sparsity penalty; It is a preset sparse regularization coefficient used to control the balance between the activation density of latent features and the reconstruction accuracy.

[0022] Furthermore, the inter-layer distribution variance features are extracted, and the t-th dimension component of the latent feature vector is calculated. The variance of activation values ​​at each layer is used as a metric to measure feature stability.

[0023] ;

[0024] in, This is a cross-layer distribution variance feature used to quantify the consistency of features across different network depths, and it is concatenated into the deep fusion feature vector as input to the classification model. This is the statistical variance operator.

[0025] Furthermore, in step S6, an elastic network regularization mechanism that integrates L1 and L2 norms is introduced, and the basic regression formula of the regularized logistic regression classification model is expressed as:

[0026] ;

[0027] The loss function for training this model combines the log-likelihood loss with the elastic network regularization term, as shown in the following formula:

[0028] ;

[0029] in: These are learnable weight parameters. It is a bias term; It is the first Each sample belongs to the target label. The predicted probability; It is the first The input feature vector of each sample; It is the total number of samples; It is the overall regularization strength coefficient; It is the elastic network hybrid parameter, with a value from 0 to 1, used to adjust the ratio of L1 regularization sparsity to L2 regularization smoothness; It is the feature dimension. It is the weight parameter of the j-th feature.

[0030] Further, in step S2, S21, while performing electrochemical impedance spectroscopy, the battery casing temperature Tc and ambient temperature Ta are collected, and a reference temperature Tr is set.

[0031] S22. Temperature uniformity control during electrochemical impedance spectroscopy testing: allow the battery to stand until the temperature change rate is less than the preset threshold before testing, or use a constant temperature chamber to make Tc fall into the preset window.

[0032] S23. Temperature compensation and normalization of the measured impedance: Based on the pre-calibrated temperature sensitivity coefficient, the impedance at each frequency point under the actual measured temperature T is corrected to the reference temperature Tr to obtain the temperature-compensated impedance Z(ω,Tr).

[0033] The charge transfer-related resistance is compensated using the Arrhenius formula:

[0034] ;

[0035] The compensated output is used for subsequent DRT transformation and construction of initial feature vectors;

[0036] in, The charge transfer resistance is calculated to the reference temperature Tr. The charge transfer resistance is measured at the actual measurement temperature T. Let R be the activation energy constant for the battery reaction, and R be the ideal gas constant. The overall characterization of the intensity of the dynamic parameters’ sensitivity to temperature.

[0037] Further, in step S3, S31, noise reduction preprocessing is performed on the impedance data, including removing abnormal frequency points, taking the average of repeated measurements, and performing smoothing filtering or robust fitting on the real and imaginary parts respectively, in order to suppress random noise and outliers.

[0038] S32. Establish a DRT discrete inversion model and perform regularized solution: Assume that the impedance measurement data satisfies the linear observation equation z = Ag + ε, where z is the measured impedance vector, A is the forward discretization kernel matrix, g is the true DRT distribution vector to be determined, and ε is the measurement noise error term; In order to estimate g from the noisy observation z, the following constrained regularized optimization problem is constructed: The first item The first term is the fitting residual term, used to ensure that the estimated value approximates the observed data; the second term... For smoothing regularization, used to suppress oscillations caused by noise; L is the smoothing operator matrix, such as a first-order or second-order difference matrix; The regularization coefficient is used to obtain the noise-resistant DRT distribution by solving this optimization problem;

[0039] S33. Adaptive selection of criteria using the L-curve method, generalized cross-validation, or validation set-based methods. The system outputs the inversion residuals and smoothness as quality indicators; when the residuals exceed the threshold, it triggers retesting or weight reduction to improve the stability of DRT features.

[0040] Implementing the embodiments of the present invention will have the following beneficial effects:

[0041] 1. The sorting method for the cascade utilization of retired lithium-ion batteries in this embodiment involves: first, performing appearance and voltage safety checks on the retired batteries to be sorted, removing abnormally damaged battery cells; then, conducting short-time electrochemical impedance spectroscopy tests on the battery samples that pass the safety check under preset conditions to obtain raw impedance data, which is then preprocessed to construct an initial feature vector; finally, parsing the initial feature vector to convert it into relaxation time distribution data, and using a pre-constructed variational autoencoder generation model for data augmentation to construct a multimodal feature set containing both real and augmented data; finally, inputting the multimodal feature set into a Transformer-MLSAE network, using a self-attention mechanism to capture full-spectrum dependencies, and then performing... Deep feature distillation and EIS-DRT feature fusion are used to extract a highly robust deep fusion feature vector. A subset of battery samples are selected for full lifecycle cyclic testing to obtain their true remaining lifespan values. Based on a pre-defined tiered utilization standard, the batteries are classified into different performance level labels. A regularized logistic regression classification model is trained under supervision using the extracted deep fusion feature vectors and performance level labels to obtain a trained retired battery sorting model. Then, the deep fusion feature vector of the retired battery to be sorted is obtained and input into the trained retired battery sorting model to obtain the predicted tiered utilization performance level of the battery. This overcomes the problems of poor model robustness and insufficient feature extraction in existing technologies with sample data.

[0042] 2. The sorting method for the secondary utilization of retired lithium-ion batteries in this embodiment, due to the multilayer sparse autoencoder (MLSAE), can reconstruct the residual flow activation vector of each layer of the Transformer. This design is based on the characteristic of "residual flow retaining information across layers", which can quantify the activation distribution of battery aging characteristics at different network depths, thereby extracting more robust global depth features, and thus solving the problem that the existing layer-by-layer independent training method has severed the information flow between layers and makes it difficult to capture dynamic changes.

[0043] 3. The sorting method for the cascade utilization of retired lithium-ion batteries in this embodiment significantly improves classification performance and achieves both lightweight and high-precision determination of retired battery grades by introducing an elastic network regularization mechanism that integrates L1 and L2 norms through the improved regularized logistic regression classification model.

[0044] 4. The sorting method for the secondary utilization of retired lithium-ion batteries in this embodiment introduces a robust front-end processing flow of "temperature uniformity and compensation and regularization DRT noise-resistant inversion". This allows for the stable extraction of interpretable DRT peak features even under conditions of on-site temperature fluctuations and measurement noise, thereby significantly reducing the feature drift risk caused by EIS temperature sensitivity and DRT ill-conditioned inversion. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0046] Figure 1 This is a flowchart of the sorting method for the cascade utilization of retired lithium-ion batteries according to an embodiment of the present invention;

[0047] Figure 2 This is a schematic diagram of the electrochemical impedance spectroscopy test described in an embodiment of the present invention;

[0048] Figure 3 This is a schematic diagram of the relaxation time distribution data described in an embodiment of the present invention;

[0049] Figure 4 This is a schematic diagram illustrating the generation of raw impedance data samples according to an embodiment of the present invention;

[0050] Figure 5 This is a comparison diagram of the original and generated distributions of the charge transfer resistor described in an embodiment of the present invention;

[0051] Figure 6 This is a schematic diagram of the electrochemical impedance spectroscopy obtained from the post-test data inversion according to an embodiment of the present invention. Detailed Implementation

[0052] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Preferred embodiments of the invention are shown in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided to provide a thorough and complete understanding of the disclosure of the invention.

[0053] It should be noted that when a component is said to be "fixed to" another component, it can be directly attached to the other component or there may be an intervening component. When a component is said to be "connected to" another component, it can be directly connected to the other component or there may be an intervening component. The terms "vertical," "horizontal," "left," "right," and similar expressions used in this document are for illustrative purposes only.

[0054] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the description of the invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0055] Please refer to Figures 1-6 This invention provides a sorting method for the secondary use of retired lithium-ion batteries, comprising the following steps:

[0056] S1. The retired batteries to be sorted undergo appearance and voltage safety inspections, and abnormally damaged battery cells are removed.

[0057] S2. Battery samples that have passed the safety inspection are subjected to short-time electrochemical impedance spectroscopy under preset conditions to obtain raw impedance data and preprocess it to construct an initial feature vector.

[0058] S3. The initial feature vector is parsed to convert it into relaxation time distribution data, and the data augmentation is performed using a pre-built variational autoencoder generative model to construct a multimodal feature set containing real data and augmented data.

[0059] S4. Input the multimodal feature set into the Transformer-MLSAE network, use the self-attention mechanism to capture the full spectrum dependency, and perform deep feature distillation and EIS-DRT feature fusion to extract a highly robust deep fusion feature vector.

[0060] S5. Select a portion of battery samples for full life cycle testing to obtain their true remaining life value, and classify the batteries into different performance level labels according to the preset tiered utilization standards.

[0061] S6. Based on the extracted deep fusion feature vectors and performance level labels, supervised training is performed on the regularized logistic regression classification model to obtain the trained retired battery sorting model.

[0062] S7. Obtain the deep fusion feature vector of the retired batteries to be sorted, input it into the trained retired battery sorting model, and obtain the predicted performance level of the battery for future reuse. In specific applications, the retired batteries to be sorted undergo appearance and voltage safety checks to remove abnormally damaged battery cells; battery samples that pass the safety check are subjected to short-time electrochemical impedance spectroscopy testing under preset conditions, such as... Figure 2As shown, the horizontal axis Re(Z) / Ohm represents the real part of the impedance, and the vertical axis Im(Z) / Ohm represents the imaginary part of the impedance. The numbers "0", "20" to "480" in the legend represent the number of charge-discharge cycles the battery has undergone, illustrating the shift and change in the impedance spectrum curve as the number of cycles increases. The raw impedance data is acquired and preprocessed to construct an initial eigenvector; the initial eigenvector is then parsed to convert it into relaxation time distribution data, such as... Figure 3 As shown, the horizontal axis τ / s represents the relaxation time constant on a logarithmic scale, and the vertical axis γ / Ω represents the distribution function value. Similarly, the numbers "0-480" in the legend correspond to the number of charge-discharge cycles of the battery, visually demonstrating the evolution of the time constant distribution of the internal electrochemical processes of the battery at different aging stages. This paper utilizes a pre-built variational autoencoder generative model for data augmentation, constructing a multimodal feature set containing both real and augmented data. This multimodal feature set is then input into a Transformer-MLSAE network, where a self-attention mechanism is used to capture full-spectrum dependencies. Deep feature distillation and EIS-DRT feature fusion are then performed to extract a highly robust deep fusion feature vector. A subset of battery samples undergoes full-lifecycle cyclic testing to obtain their true remaining lifespan values, and the batteries are categorized into different performance levels based on pre-defined tiered utilization standards. A regularized logistic regression classification model is trained under supervised supervision using the extracted deep fusion feature vectors and performance level labels, resulting in a trained retired battery sorting model. Finally, the deep fusion feature vectors of the retired batteries to be sorted are obtained and input into the trained retired battery sorting model to predict the tiered utilization performance level of the battery. This approach addresses the issues of poor model robustness and insufficient feature extraction in existing technologies with limited sample data.

[0063] In addition, by introducing generative data augmentation technology to expand the distribution of training samples, and by using deep feature distillation and EIS-DRT feature fusion technology, frequency domain data reflecting macroscopic impedance characteristics and time domain data revealing microscopic electrochemical mechanisms are deeply combined. This enables rapid and accurate determination of the remaining life and cascade utilization level of retired batteries without relying on long-term charge-discharge tests and complete historical data, while taking into account the economy and safety of the sorting process.

[0064] In one possible implementation, in step S4, let the Transformer-MLSAE network comprise a Transformer backbone network and a multilayer sparse autoencoder (MLSAE); let the Transformer backbone network comprise L layers, and for the input enhanced multimodal feature set, in the l-th layer of the Transformer backbone network, Let its output residual flow vector be denoted as The , That is, the dimension is The real vector space; extract the residual flow set of all layers. As input to the MLSAE. In specific applications, suppose the Transformer backbone network contains L layers. For the input Enhanced Multimodal Feature Sequence (EIS+DRT), in the Transformer's Lth layer... layer, The output vector of its residual flow is denoted as Therefore, it not only extracts the first The layer output, and extract the set. As input to MLSAE.

[0065] In one possible implementation, step S4 involves constructing a multilayer sparse autoencoder (MLSAE): constructing a sparse autoencoder whose encoder and decoder parameters are shared across different layers; the MLSAE generates the output residual flow vector of each layer of the Transformer backbone network. A unified processing method is used to map features from different layers to the same latent space using the same set of parameters. In specific applications, a parameter-shared sparse autoencoder is constructed, consisting of an encoder and a decoder. This MLSAE processes the output of each layer of the Transformer. They will be processed uniformly.

[0066] In one possible implementation, the encoding process of the multilayer sparse autoencoder (MLSAE) involves mapping the residual flow vector of the l-th layer to a high-dimensional sparse latent space:

[0067] ;

[0068] The decoding process of the multilayer sparse autoencoder (MLSAE) involves reconstructing the first layer through latent features. Layer input vector:

[0069] ;

[0070] in: It is the first The potential characteristics of the layer; and That is, the dimension of the latent features is much larger than the dimension of the input, which is used to construct an overcomplete representation of the features to capture subtle differences in battery aging. It is a shared encoding weight matrix. It is the encoding bias. It is the decoding bias; It is the first The reconstructed vector of the layer, It is a shared decoding weight matrix. In practical applications, due to the sparse autoencoder with shared weights, it can reconstruct the residual flow activation vector of each layer of the Transformer. This design is based on the characteristic of "residual flow retaining information across layers", which can quantify the activation distribution of battery aging features at different network depths, thereby extracting more robust global depth features, and thus solving the problem that the layer-by-layer independent training method breaks the information flow between layers and makes it difficult to capture dynamic changes.

[0071] In one possible implementation, the training objective of a multilayer sparse autoencoder (MLSAE) is to minimize the residual flow vector reconstruction error across all layers and maximize the sparsity of the latent features, with the total loss function... The total loss function is minimized during the training process to update the parameters of the encoder and decoder. The first term of the total loss function forces the model parameters to adjust in a direction that reduces reconstruction error, while the second term, through L1 regularization, forces non-critical elements in the latent feature vector to tend towards zero to achieve sparsity. The total loss function is defined as follows:

[0072] ;

[0073] in: This indicates that the summation is performed over all layers of the Transformer backbone network, and that the model is jointly optimized over the residual flow vectors of all layers. This represents the square of the L2 norm, used to calculate the reconstruction loss; This represents the L1 norm, used to calculate the sparsity penalty; This is a preset sparsity regularization coefficient used to control the balance between activation density and reconstruction accuracy of latent features. In practical applications, the training objective of MLSAE is to simultaneously minimize the reconstruction error of all layers and maximize the sparsity of latent features. Total loss function Defined as The total loss function is minimized during the training process to update the parameters of the encoder and decoder. The first term of the total loss function forces the model parameters to adjust in a direction that reduces reconstruction error, while the second term, through L1 regularization, forces non-critical elements in the latent feature vectors to tend towards zero to achieve sparsity. This indicates that the summation is performed over all layers of the Transformer backbone network, and that the model is jointly optimized over the residual flow vectors of all layers. This represents the square of the L2 norm, used to calculate the reconstruction loss; This represents the L1 norm, used to calculate the sparsity penalty; It is a preset sparse regularization coefficient used to control the balance between the activation density of latent features and the reconstruction accuracy.

[0074] In one possible implementation, the inter-layer distribution variance features are extracted, and the t-th dimension component of the latent feature vector is calculated. The variance of activation values ​​at each layer is used as a metric to measure feature stability.

[0075] ;

[0076] in, This is a cross-layer distribution variance feature used to quantify the consistency of features across different network depths, and it is concatenated into the deep fusion feature vector as input to the classification model. This is a statistical variance operator. In practical applications, for a specific battery state, certain latent features will be "activated" at a specific layer, further calculating the t-th dimension component of the latent feature vector. The variance of activation values ​​at each layer is used as a new metric for measuring feature stability. This metric reflects the degree of focus in the flow of battery aging information within the network, and it is compared with... The concatenated data is used as input to the final classifier.

[0077] In one possible implementation, step S6 introduces an elastic network regularization mechanism that integrates L1 and L2 norms, whereby the basic regression formula of the regularized logistic regression classification model is expressed as:

[0078] ;

[0079] The loss function for training this model combines the log-likelihood loss with the elastic network regularization term, as shown in the following formula:

[0080] ;

[0081] in: These are learnable weight parameters. It is a bias term; It is the first Each sample belongs to the target label. The predicted probability; It is the first The input feature vector of each sample; It is the total number of samples; It is the overall regularization strength coefficient; It is the elastic network hybrid parameter, with a value from 0 to 1, used to adjust the ratio of L1 regularization sparsity to L2 regularization smoothness; It is the feature dimension. is the weight parameter of the j-th feature. In practical applications, this model is not only simpler in classification tasks, but also better suited to high-dimensional and small-sample battery data. The original logistic regression (LR) model can be expressed as... ,in These are learnable weight parameters. It is a bias term; It is the first Each sample belongs to the target label. The predicted probability; It is the first The input feature vector of each sample; the loss function for training the LR model combines log-likelihood loss and elasticity network regularization to prevent overfitting, and can be expressed as: , It is the total number of samples; It is the overall regularization strength coefficient; It is the elastic network hybrid parameter, with a value from 0 to 1, used to adjust the ratio of L1 regularization sparsity to L2 regularization smoothness; It is the feature dimension. It is the weight parameter of the j-th feature.

[0082] In one possible implementation, in step S2,

[0083] S21. While performing electrochemical impedance spectroscopy, collect the battery casing temperature Tc and ambient temperature Ta, and set a reference temperature Tr.

[0084] S22. Temperature uniformity control during electrochemical impedance spectroscopy testing: allow the battery to stand until the temperature change rate is less than the preset threshold before testing, or use a constant temperature chamber to make Tc fall into the preset window.

[0085] S23. Temperature compensation and normalization of the measured impedance: Based on the pre-calibrated temperature sensitivity coefficient, the impedance at each frequency point under the actual measured temperature T is corrected to the reference temperature Tr to obtain the temperature-compensated impedance Z(ω,Tr).

[0086] The charge transfer-related resistance is compensated using the Arrhenius formula:

[0087] ;

[0088] The compensated output is used for subsequent DRT transformation and construction of initial feature vectors;

[0089] in, The charge transfer resistance is calculated to the reference temperature Tr. The charge transfer resistance is measured at the actual measurement temperature T. Let R be the activation energy constant for the battery reaction, and R be the ideal gas constant. The overall characterization assesses the intensity of the kinetic parameters' sensitivity to temperature. In specific applications, the battery casing temperature Tc and ambient temperature Ta are collected simultaneously during EIS testing, and a reference temperature Tr (25℃) is set. Temperature uniformity control is implemented during the EIS testing process: the battery is allowed to stand until the rate of temperature change is less than a preset threshold (<0.1℃ / min) before testing, or a constant-temperature chamber is used to ensure Tc falls within a preset window. Temperature compensation and normalization are performed on the measured impedance: based on a pre-calibrated temperature sensitivity coefficient, the impedance at each frequency point under the actual measured temperature T is mapped to Tr to obtain the temperature-compensated impedance Z(ω,Tr). Arrhenius compensation is used for charge transfer-related resistance. The compensated output is used for subsequent DRT transformation and feature construction. Additionally, further explanation is needed. Let R be the activation energy constant for the battery reaction, and R be the ideal gas constant. The overall characterization of the intensity of the kinetic parameters’ sensitivity to temperature; Ea / R, the overall characterization of the intensity of the kinetic parameters’ sensitivity to temperature, can be understood as a quantity that converts the “energy barrier per mole” into an “equivalent temperature scale” – the intensity parameter of the kinetics’ sensitivity to temperature.

[0090] In one possible implementation, in step S3, S31, noise reduction preprocessing is performed on the impedance data, including removing abnormal frequency points, taking the average of repeated measurements, and performing smoothing filtering or robust fitting on the real and imaginary parts respectively, in order to suppress random noise and outliers.

[0091] S32. Establish a DRT discrete inversion model and perform regularized solution: Assume that the impedance measurement data satisfies the linear observation equation z = Ag + ε, where z is the measured impedance vector, A is the forward discretization kernel matrix, g is the true DRT distribution vector to be determined, and ε is the measurement noise error term; In order to estimate g from the noisy observation z, the following constrained regularized optimization problem is constructed: The first item The first term is the fitting residual term, used to ensure that the estimated value approximates the observed data; the second term... The smoothing regularization term is used to suppress oscillations caused by noise; L is the smoothing operator matrix (such as a first-order or second-order difference matrix). The regularization coefficient is used to obtain the noise-resistant DRT distribution by solving this optimization problem;

[0092] S33. Adaptive selection of criteria using the L-curve method, generalized cross-validation, or validation set-based methods. The system outputs the inversion residuals and smoothness as quality indicators. When the residuals exceed a threshold, retesting or weight reduction is triggered to improve the stability of the DRT features. In specific applications, noise-resistant preprocessing is performed on the impedance data, including outlier removal, averaging of repeated measurements, and smoothing or robust fitting of the real and imaginary parts to suppress random noise and outliers. A DRT discrete inversion model is established and solved using regularization: the DRT inversion is written as a linear system z = Ag + ε, where z is the measured impedance vector and g is the true DRT distribution vector to be determined. The noise-resistant DRT is obtained by solving a constrained regularization optimization problem. The first item The first term is the fitting residual term, used to ensure that the estimated value approximates the observed data; the second term... For smoothing regularization, used to suppress oscillations caused by noise, where L is the smoothing operator matrix, such as a first-order or second-order difference matrix; The regularization coefficient is used to obtain the noise-resistant DRT distribution by solving this optimization problem, and the inversion residual and smoothness are output as quality indicators. When the residual exceeds the threshold, retesting or weight reduction is triggered to improve the stability of DRT features.

[0093] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these all fall within the scope of protection of the present invention. Therefore, the scope of protection of this patent should be determined by the appended claims.

Claims

1. A sorting method for the cascade utilization of retired lithium-ion batteries, characterized in that, Including the following steps: S1. The retired batteries to be sorted undergo appearance and voltage safety inspections, and abnormally damaged battery cells are removed. S2. Battery samples that have passed the safety inspection are subjected to short-time electrochemical impedance spectroscopy under preset conditions to obtain raw impedance data and preprocess it to construct an initial feature vector. S3. The initial feature vector is parsed to convert it into relaxation time distribution data, and the data is augmented using a pre-built variational autoencoder generation model to construct a multimodal feature set containing real data and augmented data. S4. Input the multimodal feature set into the Transformer-MLSAE network, use the self-attention mechanism to capture the full spectrum dependency, and perform deep feature distillation and EIS-DRT feature fusion to extract a highly robust deep fusion feature vector. S5. Select a portion of battery samples for full life cycle testing to obtain their true remaining life value, and classify the batteries into different performance level labels according to the preset tiered utilization standards. S6. Based on the extracted deep fusion feature vector and the performance level label, supervise the training of the regularized logistic regression classification model to obtain the trained retired battery sorting model. S7. Obtain the deep fusion feature vector of the retired battery to be sorted, input it into the trained retired battery sorting model, and obtain the prediction result of the battery's tiered utilization performance level. In step S4, let the Transformer-MLSAE network consist of a Transformer backbone network and a multi-layer sparse autoencoder (MLSAE). Assume the Transformer backbone network has L layers. For the input enhanced multimodal feature set, in the l-th layer of the Transformer backbone network... Let its output residual flow vector be denoted as The , That is, the dimension is The real vector space; extract the residual flow set of all layers. As input to the MLSAE; Constructing a Multilayer Sparse Autoencoder (MLSAE): A sparse autoencoder with shared encoder and decoder parameters is constructed across different layers; the MLSAE generates the output residual flow vector for each layer of the Transformer backbone network. A unified process is performed, using the same set of parameters to map features from different layers to the same latent space; The encoding process of the multilayer sparse autoencoder (MLSAE) is to map the residual flow vector of the l-th layer to a high-dimensional sparse latent space: ； The decoding process of the multilayer sparse autoencoder (MLSAE) involves reconstructing the input vector of the l-th layer using latent features: ； in: It is a latent feature of layer l; and That is, the dimension of the latent features is much larger than the dimension of the input, which is used to construct an overcomplete representation of the features to capture subtle differences in battery aging. It is a shared encoding weight matrix. It is the encoding bias. It is the decoding bias; It is the reconstruction vector of the l-th layer. It is a shared decoding weight matrix.

2. The sorting method for the cascade utilization of retired lithium-ion batteries according to claim 1, characterized in that, The training objective of a multilayer sparse autoencoder (MLSAE) is to minimize the residual flow vector reconstruction error across all layers and maximize the sparsity of the latent features. The overall loss function is defined as follows: for: ； in: This indicates that the summation is performed over all layers of the Transformer backbone network, and that the model is jointly optimized over the residual flow vectors of all layers. This represents the square of the L2 norm, used to calculate the reconstruction loss; This represents the L1 norm, used to calculate the sparsity penalty; It is a preset sparse regularization coefficient used to control the balance between the activation density of latent features and the reconstruction accuracy.

3. The sorting method for the cascade utilization of retired lithium-ion batteries according to claim 2, characterized in that, Extract the inter-layer distribution variance features and calculate the t-th dimension component of the latent feature vector. The variance of activation values ​​at each layer is used as a metric to measure feature stability. ； in, This is a cross-layer distribution variance feature used to quantify the consistency of features across different network depths, and it is concatenated into the deep fusion feature vector as input to the classification model. This is the statistical variance operator.

4. The sorting method for the cascade utilization of retired lithium-ion batteries according to claim 1, characterized in that, In step S6, an elastic network regularization mechanism that integrates L1 and L2 norms is introduced. The basic regression formula of the regularized logistic regression classification model is expressed as: ； The loss function for training this model combines the log-likelihood loss with the elastic network regularization term, as shown in the following formula: ； in: These are learnable weight parameters. It is a bias term; It is the first Each sample belongs to the target label. The predicted probability; It is the first The input feature vector of each sample; It is the total number of samples; It is the overall regularization strength coefficient; It is the elastic network hybrid parameter, with a value from 0 to 1, used to adjust the ratio of L1 regularization sparsity to L2 regularization smoothness; It is the feature dimension. It is the weight parameter of the j-th feature.

5. The sorting method for the cascade utilization of retired lithium-ion batteries according to claim 1, characterized in that, In step S2, S21. While performing electrochemical impedance spectroscopy, collect the battery casing temperature Tc and ambient temperature Ta, and set a reference temperature Tr. S22. Temperature uniformity control during electrochemical impedance spectroscopy testing: allow the battery to stand until the temperature change rate is less than the preset threshold before testing, or use a constant temperature chamber to make Tc fall into the preset window. S23. Temperature compensation and normalization of the measured impedance: Based on the pre-calibrated temperature sensitivity coefficient, the impedance at each frequency point under the actual measured temperature T is corrected to the reference temperature Tr to obtain the temperature-compensated impedance Z(ω,Tr). The charge transfer-related resistance is compensated using the Arrhenius formula: ； The compensated output is used for subsequent DRT transformation and construction of initial feature vectors; in, The charge transfer resistance is calculated to the reference temperature Tr. The charge transfer resistance is measured at the actual measurement temperature T. Let R be the activation energy constant for the battery reaction, and R be the ideal gas constant. The overall characterization of the intensity of the dynamic parameters’ sensitivity to temperature.

6. The sorting method for the cascade utilization of retired lithium-ion batteries according to claim 1, characterized in that, In step S3, S31. Noise-resistant preprocessing of impedance data: including outlier frequency point removal, averaging of repeated measurements, and smoothing or robust fitting of the real and imaginary parts to suppress random noise and outliers. S32. Establish a DRT discrete inversion model and perform regularized solution: Assume that the impedance measurement data satisfies the linear observation equation z = Ag + ε, where z is the measured impedance vector, A is the forward discretization kernel matrix, g is the true DRT distribution vector to be determined, and ε is the measurement noise error term; In order to estimate g from the noisy observation z, the following constrained regularized optimization problem is constructed: The first item The first term is the fitting residual term, used to ensure that the estimated value approximates the observed data; the second term... For smoothing regularization, used to suppress oscillations caused by noise; L is the smoothing operator matrix, such as a first-order or second-order difference matrix; The regularization coefficient is used to obtain the noise-resistant DRT distribution by solving this optimization problem; S33. Adaptive selection of criteria using the L-curve method, generalized cross-validation, or validation set-based methods. The system outputs the inversion residuals and smoothness as quality indicators; when the residuals exceed the threshold, it triggers retesting or weight reduction to improve the stability of DRT features.

Images