An intelligent sorting method and device for multi-source heterogeneous retired batteries

Abstract

The application provides a kind of multi-source heterogeneous retired battery-oriented intelligent sorting method and device, first obtain the dimensionless characteristics of retired battery to extract the data representing the original aging characteristics of retired battery;Then construct a double-branch decoupling network model, branch one outputs the material category label of battery, branch two outputs the relative aging state of battery, use the A type retired battery data set with known material attribute but no aging state label to train the model, get the A type aging state grouping rule;The retired battery data with unknown material attribute is input into the model, whether the battery belongs to A type distribution is identified by distance measurement in feature space, if not, then establish the class of the battery as B type, use unsupervised domain adaptation method to align and apply A type aging state grouping rule to B type battery space;Finally, more types of retired battery data are input into the model to optimize the model, to realize the identification and sorting of unknown material retired battery.

Description

Technical Field

[0001] This invention belongs to the field of battery performance testing technology, and in particular relates to an intelligent sorting method and device for multi-source heterogeneous retired batteries. Background Technology

[0002] Due to the diverse sources of retired batteries, their internal electrochemical material systems (such as lithium iron phosphate, ternary lithium, and lithium titanate) vary, and their aging and degradation levels differ significantly. If batteries of different materials are connected in series, during charging and discharging, batteries with high-voltage platforms are prone to overcharging, easily leading to thermal runaway; while batteries with low-voltage platforms are prone to over-discharging, causing the copper current collector to dissolve. Furthermore, mixing severely aged and slightly aged batteries can also easily trigger thermal runaway accidents, as the severely aged batteries experience a sharp increase in internal resistance and a drastic decrease in capacity, generating abnormal Joule heating and polarization heat during high-current charging and discharging. Simultaneously, capacity mismatch can cause the battery management system to frequently trigger undervoltage or overvoltage protection, resulting in the entire battery pack failing prematurely before reaching its designed lifespan. Therefore, precise sorting of retired batteries is essential before their reuse.

[0003] Currently, the sorting of retired batteries mainly employs data-driven methods, such as the retired battery sorting method, device, electronic equipment, and storage medium disclosed in CN121682043A. This method uses multiple sorting feature parameters combined with a clustering algorithm to sort batteries of the same material. However, this method can only handle a single known battery type. Furthermore, the literature HangHu, Xinghao Huang, Chen Liang, et al., "Second..." -level heterogeneous retired battery type identification using pulse-test-enabled federated learning with output- The paper "Level Privacy Preservation", published in *eTransportation*, 2026, Vol. 29, art. no.100607, DOI: 10.1016 / j.etran.2026.100607, discloses a method for identifying retired battery types based on second-level pulse testing. The method is tested on a heterogeneous dataset containing eight battery types, over 600 retired batteries, and 10,184 pulse test records. Model training and evaluation require the prior collection of label data for all battery types, with a training set to test set ratio of 8:2 for each battery type.

[0004] However, actual recycling lines deal with battery packs from unknown sources and with complex electrochemical systems. Existing methods are all based on the closed-set assumption, and the models can only identify the categories they have seen before. This has many shortcomings in the intelligent sorting of retired batteries:

[0005] 1. Risk of missorting due to closed-set identification, lacking the ability to perceive unknown distributions. Most existing sorting models are closed-set classifiers. When new electrochemical system batteries that have not been trained on the model are mixed into the production line (e.g., the model is only trained on lithium iron phosphate batteries, but ternary lithium batteries are actually mixed in), the model lacks the ability to reject unknown categories and will forcibly classify the new system batteries into known categories, leading to serious batch mixing risks.

[0006] 2. Poor model scalability and extremely high retraining costs. When a new type of battery is introduced into the recycling line, traditional methods must mix the new battery data with the old battery data, recollect massive amounts of data, and retrain the model, making it impossible to achieve dynamic model expansion.

[0007] 3. Lack of ability to transfer aging state characteristics across chemical systems. Although different electrochemical materials have different common characteristics, their differences in aging degradation patterns share certain commonalities in the deep data space. Existing methods cannot directly transfer the aging state classification rules learned on type A batteries to newly emerging type B batteries. Summary of the Invention

[0008] In order to effectively identify the material properties and aging properties of multi-source heterogeneous retired batteries and realize the access and adaptive sorting of new retired batteries, this invention provides an intelligent sorting method and device for multi-source heterogeneous retired batteries.

[0009] This invention discloses an intelligent sorting method for multi-source heterogeneous retired batteries, the method comprising:

[0010] Step S1: Obtain the dimensionless or incremental features of the retired battery to extract data characterizing the original aging features of the retired battery.

[0011] Step S2: Construct a dual-branch decoupled network model. Branch 1 outputs the material category label of the battery, and branch 2 outputs the relative aging state of the battery in its corresponding material category. Train the model using a dataset of A-class batteries with known material properties but no aging state labels to obtain the A-class aging state grouping rules. The dataset is the data set obtained in step S1.

[0012] Step S3: Input the retired battery data with unknown material properties into the model trained in step S2. Identify whether the battery belongs to class A distribution through the distance metric in the feature space. If not, establish the battery as class B and proceed to step S4.

[0013] Step S4: Using an unsupervised domain adaptive method, the obtained A-class aging state grouping rules are aligned and applied to the B-class battery space to achieve unsupervised sorting of B-class batteries.

[0014] Step S5: Repeat steps S3 and S4 to input more types of retired battery data into the model and optimize the model in order to achieve the identification and sorting of retired batteries of unknown materials.

[0015] Preferably, further detailed description of the present invention can be found in the following detailed embodiments.

[0016] As another aspect of the present invention, an intelligent sorting device for multi-source heterogeneous retired batteries includes a memory, a processor, and a computer program stored in the memory. The processor executes the computer program to implement the intelligent sorting method for multi-source heterogeneous retired batteries.

[0017] In another aspect of the invention, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the intelligent sorting method for multi-source heterogeneous decommissioned batteries.

[0018] As another aspect of the present invention, a computer program product includes a computer program that, when executed by a processor, implements the intelligent sorting method for multi-source heterogeneous decommissioned batteries.

[0019] The intelligent sorting method for multi-source heterogeneous retired batteries proposed in this invention first extracts the general dimensionless features of retired batteries as input to a dual-branch decoupled network model, thus avoiding feature confusion caused by differences in the physical decay characteristics of different electrochemical material systems. Then, a dual-branch decoupled network model is constructed and jointly trained using a dataset of batteries without aging state labels and known material properties. This enables the model to minimize the difference between the actual and predicted material categories, and to form stable aging state grouping rules in the latent space while effectively removing interference from material properties. Finally, a chi-square distribution hypothesis test is introduced to calculate statistical significance, enabling out-of-distribution detection of heterogeneous batteries. Furthermore, the maximum mean difference and elastic weight consolidation algorithms driven by new class soft weights are used to unsupervisedly transfer the trained aging state grouping rules across domains to a new group of batteries with unknown material properties. This method overcomes the dependence on battery data and actual aging state labels of different material systems, can automatically learn knowledge of new types of batteries, and sort existing types of retired batteries. It can be directly and efficiently deployed in intelligent cascade utilization sorting production lines for multi-source heterogeneous retired power batteries, giving the production line the ability to identify and sort batteries with unknown materials, and realize the access of new batteries and adaptive sorting. Attached Figure Description

[0020] To more clearly illustrate the technical solutions in this invention or the prior art, the accompanying drawings used in the embodiments will be briefly described below:

[0021] Figure 1 is a flowchart of the intelligent sorting method for multi-source heterogeneous retired batteries provided by the present invention;

[0022] Figure 2 is an architecture diagram of the dual-branch decoupling network model in this invention;

[0023] Figure 3 is a flowchart of the detection and sorting process for retired batteries made of new materials in an embodiment of the present invention;

[0024] Figure 4 The global material identification confusion matrix is ​​established and incrementally updated for different new classes n in the experiments of this embodiment of the invention (where: Figure 4 (a) is the confusion matrix of global material identification after the establishment of the new class with n=30 and the incremental update; Figure 4 (b) is the confusion matrix for global material identification after establishing the new class with n=40 and incrementally updating it; Figure 4 (c) is the confusion matrix of the new class establishment and incremental update of global material identification for n=50).

[0025] Figure 5 The following are the clustering results of the aging state after model solidification classification under different n conditions in the experiments of this invention (where: Figure 5 (a) is the clustering result of the aging state after the model solidification classification with n=30; Figure 5 (b) shows the clustering results of the aging state after the model solidification classification with n=40; Figure 5 (c) is the clustering result of the aging state after the model solidification classification when n=50). Detailed Implementation

[0026] To make the objectives, technical solutions, and advantages of this invention clearer, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of this invention. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention.

[0027] To address the limitations of existing technologies in distinguishing between the material and aging properties of retired batteries, and in spontaneously identifying new types of retired batteries, this invention provides an intelligent sorting method and apparatus for multi-source heterogeneous retired batteries. This method reduces the costs of data collection, processing, and manual labeling when new retired batteries are introduced, while effectively preventing production line sorting errors and safety risks caused by the mixing of heterogeneous batteries. Figure 1As shown, the intelligent sorting method provided by this invention includes acquiring dimensionless or incremental features of retired batteries, constructing a dual-branch decoupled network model, detecting out-of-distribution heterogeneous batteries, identifying unknown groups and transferring unsupervised cross-domain mechanisms, and incremental continuous learning of multi-source batteries. Taking the sequential access of retired batteries of type A (known), type B (unknown), and type C (unknown) as an example, the specific steps of the method proposed in this invention are as follows:

[0028] Step S1: Obtain the dimensionless or incremental features of the retired battery.

[0029] This embodiment acquires the dimensionless characteristics of retired batteries by collecting their charge-discharge dynamic sequences and using differential and integral transforms to extract general aging features that characterize the electrochemical evolution trajectory. These features include the peak position shift and relative change rate of the peak area in the capacity increment curve, as well as the characteristic valley capacity spacing in the differential voltage curve. The extracted general aging features are then globally normalized to obtain the original aging feature vector, which serves as the input to the two-branch decoupling network model. Specifically:

[0030] Discrete dynamic physical quantity sequences of retired batteries during constant current charge-discharge processes are collected using production line testing equipment, including time sequence t, voltage sequence V(t), and current sequence I(t). The charge-discharge capacity Q(t) at the current moment is calculated based on the ampere-hour integral method.

[0031]

[0032] In the formula, Q(t) represents the cumulative capacity of the battery at time t. The original voltage and capacity curves are easily affected by physical noise. To highlight the phase transition characteristics inside the battery, differential and integral transformations are used to reconstruct the incremental capacity (IC) curve and the differential voltage (DV) curve, as shown in the following formula:

[0033]

[0034]

[0035] In the formula, IC(V) represents the unit voltage. From capacity under change to capacity The increment of the curve reflects the phase transition plateau where the main redox reaction occurs in the electrode material; DV(Q) represents the unit capacity. Voltage variation to voltage The peak and valley positions of the curve are highly correlated with the loss of active lithium and the increase in internal resistance of the electrode material.

[0036] On the reconstructed IC and DV curves, a peak-finding algorithm is used to quantify and extract general aging characteristics that can span different material systems. This implementation mainly extracts the following general aging characteristics.

[0037] Feature 1: The peak position offset of the IC curve, including the main peak voltage position V. peak With the height of the main peak H peak This formula is used to directly characterize the main reaction potential, reaction intensity, and polarization resistance evolution of the current battery, as shown below:

[0038]

[0039]

[0040] In the formula, Indicates to make When taking the maximum value The value of ; express The maximum value.

[0041] Feature 2: The relative rate of change of the peak area of ​​the IC curve is the integral area A of the main peak region. peak The absolute content of active substances participating in the reaction of the current electrode material is used to characterize the content of active substances, as shown in the following formula:

[0042]

[0043] In the formula, and The starting and ending voltage boundaries of the main peak of the IC curve are determined using the second derivative zero-finding method.

[0044] Feature 3: Characteristic valley spacing of the DV curve The relative distance between two adjacent phase transition valleys within a single DV curve is used to characterize the capacity matching relationship between the positive and negative electrode materials, as shown in the following formula:

[0045]

[0046] In the formula, and These correspond to the preset capacity search intervals where two adjacent feature valleys on the DV curve are located, respectively. and respectively make When taking the minimum value and The value of .

[0047] The extracted high-dimensional absolute features (i.e., general aging features) are concatenated to form the original aging feature vector. Since the physical dimensions of each feature are different, and the absolute value ranges of different battery materials (such as ternary lithium and lithium iron phosphate) are extremely wide, in order to meet the cross-domain alignment requirements of the subsequent network, the maximum and minimum value normalization is used as follows to map all features dimensionlessly to the range [0, 1], and constructs a normalized original aging feature vector as input to the dual-branch decoupled network model. ,in, They are respectively Normalized dimensionless eigenvalues;

[0048]

[0049] In the formula, The normalized k-th dimensionless eigenvalue, where k = 1, 2, 3 or 4; and These are the k-th dimension original aging feature vectors in the current batch of data being processed. The global minimum and maximum values ​​are found. This step removes the dimensional attributes of the original data, leaving only the relative distribution structure of the features, which lays the foundation for the subsequent model to learn the generalization law across electrochemical materials.

[0050] Step S2: Construct a two-branch decoupled network model.

[0051] like Figure 2 As shown, a deep network containing a shared feature extractor and dual-branch outputs is constructed to form a dual-branch decoupled network model. The role of the shared feature extractor is to map the normalized original aging feature vector from the original feature space to the shared latent space, so as to obtain a shared representation that can be called by both branches. The first branch of the model is the common learning branch (material property classification), which is used to output the material property category of the battery. The second branch is the differential learning branch (aging state classification), which is used to output the relative aging state of the battery in class A.

[0052] In this embodiment, the model is trained using a dataset of retired Class A batteries with known material properties but no aging state labels. This dataset is a normalized original aging feature vector dataset containing N Class A battery samples. The training process is illustrated here using the i-th Class A battery sample as an example. Figure 2 As shown:

[0053] First, normalize the original aging feature vector of the i-th class A battery sample in the dataset. The input is fed into the model, where it is processed by the shared feature extractor. By mapping the original aging feature space to the shared latent space, a shared representation of the i-th class A battery sample that can be invoked by both branches is obtained. As shown in the following formula:

[0054]

[0055] In the formula, lr and L are the set of trainable parameters of the shared feature extractor, the network layer number, and the total number of layers, respectively. These are the shared feature vectors output by the shared feature extractor for the initial layer, the lr-th layer, and the final layer, respectively. For the nonlinear activation function of the shared feature extractor; and These are the weight matrix and bias term of the lr layer, respectively; To share the feature extraction function, it consists of multiple layers of fully connected layers, convolutional layers, or recurrent units. The general evolutionary information in the original aging characteristics is preserved, providing a unified input for subsequent material property identification and aging state classification.

[0056] then, Enter branch one and branch two respectively.

[0057] 1) In branch one, After common mapping The material property latent vector of the i-th type A battery sample is obtained. As shown in the following formula;

[0058]

[0059] In the formula, Trainable parameters for the branch-one commonality mapping; For branch one, the nonlinear activation function is used. and These are the weight matrix and bias term of the nonlinear activation function in branch one, respectively. Commonality mapping. The main focus is on the common discriminative features among different material systems.

[0060] immediately, Upon entering the classification layer, the sample is processed by the Softmax activation function. The predicted probability of the sample being classified as various material categories is calculated according to the following formula. Finally, the material category label corresponding to the highest predicted probability value is output from the classification layer as the material category label of the sample.

[0061]

[0062] In the formula, This represents the predicted probability that the i-th sample is classified as the c-th material property. The activation function for the classification layer; It is a natural exponential function; and These are the weights and biases corresponding to the c-th class in the classification layer, respectively. and These are the weights and biases corresponding to the r-th class in the classification layer, respectively; Let r represent the total number of retired battery material categories. When r = 1, 2, ..., the corresponding material category labels are A, B, ..., respectively; where r = 1 corresponds to category A, r = 2 corresponds to category B; in this training phase, the material category label of the sample is known to be A, and the predicted probability of the i-th sample being classified as a material of category 1 is given. The maximum value should be used; the classification layer will output the material category label for the i-th sample. .

[0063] 2) In branch two, Through difference mapping The aging state latent vector of the i-th type A battery sample is obtained. As shown in the following formula;

[0064]

[0065] In the formula, These are the trainable parameters for the branch-two difference mapping; For branch two, it is a nonlinear activation function; and These are the weight matrix and bias term of the branch-two nonlinear activation function, respectively. (Discretionary mapping) The main focus is on the relative aging differences within the same material system caused by factors such as capacity decay, increased internal resistance, and enhanced polarization.

[0066] immediately, The samples are then grouped unsupervised in the clustering module to obtain their relative aging status.

[0067] Following the above method, the model is trained sequentially using samples from the Class A retired battery dataset. During the entire training process, the first branch of the model learns latent vectors that represent the commonalities of material properties to minimize the difference between the true and predicted material categories. After training, the second branch of the model not only has the ability to extract material information but also forms and records K aging state groups for Class A batteries. , For the Kth aging state group, the aging state group boundary can be determined based on the aging state group. It is worth noting that the aging state group boundary is an implicit discrimination boundary automatically learned by branch two in the aging state latent space, not a manually preset capacity, internal resistance, or voltage threshold. Specifically, after the retired battery samples are processed by the shared feature extractor and difference mapping, aging state latent vectors are obtained. The clustering module, based on the similarity between the aging state latent vectors of different samples, groups samples with similar aging features into the same aging state group, thus separating different aging state groups in the latent space. After training, each aging state group corresponds to a stable latent cluster center or latent space region, and the boundary relationship between adjacent aging state groups constitutes the aging state group boundary. For retired batteries to be sorted, the model assigns them to the best-matching aging state group based on the similarity or soft assignment probability between their aging state latent vectors and each aging state group, thereby achieving relative aging state classification without manual aging labeling.

[0068] In summary, the dual-branch decoupled network model constructed in step S2 directly embeds the clustering mechanism into the differential learning branch. By introducing a soft clustering assignment mechanism, the network optimizes material classification while aggregating K aging state groups in the latent vector of aging state. Each group has a corresponding aging state grouping boundary. At the same time, a multivariate standard normal distribution is introduced as a variational prior constraint in the common learning branch.

[0069] Preferably, in this embodiment, the total loss function for multi-task joint optimization of the dual-branch decoupled network model is expressed as follows:

[0070]

[0071] In the formula, The total loss value for multi-task joint optimization is the ultimate goal of backpropagation in the entire deep network; The loss function for material property classification is set at the classification layer of branch one to constrain the consistency between the predicted probability calculated by the classification layer and the known material category label. Commonality mapping of Class A battery samples through network branch one The resulting implicit vector of material properties; The known material category labels for class A battery samples are given. During the initial training phase, it is known that the material category label for all samples is A. This is an embedded deep clustering loss function, which is set in the clustering module of branch two to constrain branch two to form differential aging state groups in the unlabeled state; Differential mapping of class A battery samples through network branch two The resulting latent vector of aging state; This is a mutual information penalty term between the latent vectors of material properties and the latent vectors of aging state. It is used to quantitatively measure the statistical correlation between the two sets of features. This term needs to be minimized during model training to achieve decoupling of material and aging features. This is a balance coefficient for mutual information constraints, used to adjust the weight ratio between clustering accuracy and feature decoupling degree in the aging state of the model; As variational prior constraints, calculate the posterior distribution of the output of branch one. Compared with the multivariate standard normal distribution The KL divergence between them will standardize the material characteristics of Class A batteries within a unit hypersphere; For Given the condition, the probability distribution of latent features of material properties derived by the network encoder. These are the dimensionless eigenvalues ​​of the normalized Class A batteries; These are the balance coefficients for variational prior constraints.

[0072] Preferably, in this embodiment, the material property classification loss function adopts cross-entropy loss, as shown in the following formula:

[0073]

[0074] In the formula, Let c be the known material category label for the i-th sample. When c = 1, 2, ..., the corresponding material category labels are A, B, ...; represents the cross-entropy loss value; N represents the number of samples. This loss function minimizes the difference between the true material class and the predicted material class, enabling branch-one learning to represent the latent vectors of common material properties.

[0075] Preferably, in this embodiment, the embedded deep clustering loss function adopts KL divergence loss. Based on the aging state latent vector, the soft assignment probability of a sample relative to the K latent cluster centers is calculated, and the KL divergence between the target distribution and the soft assignment probability is used to promote the clustering of samples with similar aging states in the dissimilar latent space. To clarify the unsupervised grouping mechanism within branch two, The core mathematical construction process is as follows:

[0076] First, using the t-distribution as the kernel function, the aging state latent vector of each Class A battery sample is calculated. The soft assignment probability of belonging to the j-th hidden cluster center;

[0077]

[0078] In the formula, The soft assignment probability is that the i-th type A battery sample belongs to the j-th aging state group. The difference mapping of the i-th class A battery sample through model branch two The resulting latent vector of aging state; Let be the optimized latent cluster center vector in the dissimilarity latent space, where K is the number of hidden cluster centers; represents the degrees of freedom parameter of the t-distribution; This represents the square of the Euclidean distance, used to measure the spatial distance between the latent vector and the latent cluster center.

[0079] Next, a target distribution is constructed based on the soft assignment probability to reinforce the grouping tendency of high-confidence samples and normalize the loss;

[0080]

[0081] In the formula, The target probability, calculated based on the soft assignment probability, is used to guide the agglomeration of cluster centers.

[0082] Finally, the embedded deep clustering loss is used. Represented as the KL divergence between the target probability and the current soft assignment probability;

[0083]

[0084] Preferred, and Mutual information penalty item The expression is as follows:

[0085]

[0086] In the formula, Implicit vectors for material properties With aging state latent vector The joint probability distribution, and They are respectively and The marginal probability distribution. During model training, by minimizing this mutual information penalty term, the statistical dependency between the material property latent vector and the aging state latent vector is reduced, thereby achieving decoupling of material property features and aging state features.

[0087] Step S3: External detection of heterogeneous battery distribution.

[0088] Because step S2 is in the material property implicit vector A multivariate standard normal distribution was applied to it. In mathematics, if any sample belongs to class A batteries, the square norm of its latent vector approximately follows a chi-square distribution after training.

[0089] Therefore, when a Class B retired battery B1 with completely unknown material properties first enters the production line, its latent material property vector is extracted first. And calculate the square of its Euclidean distance. ;

[0090]

[0091] Next, based on the chi-square distribution theory, we construct the hypothesis H0 that "this sample belongs to the A-class battery distribution." Under the condition that H0 holds, Strictly follows a chi-square distribution with d degrees of freedom (where d is the implicit vector Z of the material properties) type (Dimension), directly calculate the statistical confidence probability value of the retired battery sample belonging to the known Class A battery distribution, as follows:

[0092]

[0093] In the formula, The statistical confidence probability value for retired battery B1 belonging to the known distribution of class A batteries; The material property latent vector of retired battery B1 after commonality mapping through model branches. It is the cumulative distribution function of the chi-square distribution with d degrees of freedom; For the lower incomplete gamma function, ; For standard gamma function, .

[0094] Finally, a standard statistical significance level is set. If the calculated Then reject the null hypothesis H0, and use The confidence level is determined as "B1 does not belong to Class A batteries".

[0095] Step S4: Identification of unknown groups and unsupervised cross-domain mechanism migration.

[0096] Step S401, as follows Figure 3 As shown, as the unknown battery continues to be tested, the model is tested up to the nth time (the dataset continues to expand). → →…→ The production line will contain a mixture of known Type A batteries and unknown heterogeneous material batteries. Combined with step S3, for any sample... ( The statistical uncertainty in classifying a battery as an outlier is defined as the new class soft weight. :

[0097]

[0098] in, For the sample The statistical confidence probability value for a battery belonging to the A-class battery distribution obtained in step S3. If a battery belongs to the A-class battery distribution, the statistical confidence probability value is... If the significance level is below the preset threshold, the soft weight of the new class approaches 1. If the value exceeds the significance threshold, the new class soft weights approach 0. The model utilizes the new class soft weights. For all Class B batteries (at least 50) input samples, in the latent vector of material properties Weighted density clustering is performed in the middle, such as using the weighted DBSCAN algorithm, when there are high-new-class soft weights. When the samples aggregate into stable, high-density feature clusters, the cluster is officially confirmed as a Class B battery (i.e., Class B), representing a completely new type of retired battery, and the dataset belonging to this cluster is identified. .

[0099] Step S402: After confirming the B-class group, the model unsupervisedly transfers the aging state grouping rules of class A obtained in step S2 to the B-class batteries. This is because the latent vectors in the aging state... In this model, material physical properties have been removed. At this point, an unsupervised domain adaptation method is used to align the edge distribution of aging state features of class B clusters to the feature distribution of class A. To avoid interference from a small number of noise samples that may be mixed into the class B set, the newly obtained soft weights for the new class are used. As the alignment penalty coefficient, construct the weighted maximum mean difference loss function:

[0100]

[0101] In the formula, The maximum mean difference loss value of cross-domain aging state features is introduced to minimize the distribution distance between class A and class B batteries in the aging state space. The kernel function that maps low-dimensional latent features to the reproducing kernel Hilbert space; and These are the i-th type A batteries. and the j-th type B battery The latent vector of aging state; For the j-th class B battery after normalization . dimensionless eigenvalues.

[0102] In step S403, after the alignment training optimization converges, the aging state feature trajectories of class B batteries have overlapped with the feature space of class A. At this point, the hidden cluster centers learned in step S2 are fixed. For any retired battery belonging to category B, no manual aging state labeling is required; only its latent aging state vector needs to be calculated. Relative to each hidden cluster center soft assignment probability (Same as the formula in step S2), take the cluster number corresponding to the maximum soft assignment probability, and the B-type battery can be finely divided into the corresponding aging state group, thus completing the unsupervised cross-domain migration of aging state rules.

[0103] Step S5: Incremental continuous learning of multi-source batteries.

[0104] As more C-class or higher batteries with novel chemical systems continue to enter the production line, the detection and cross-domain migration operations in steps S3 and S4 are repeated. To prevent the model from erasing previously learned A-class and B-class grouping features during C-class feature alignment, elastic weight consolidation constraints are introduced during incremental model updates.

[0105]

[0106] In the formula, The incremental optimization objective total loss function incorporates a continuous learning mechanism; This is the set of variable parameters for the current model network used for adaptive iteration of Class C batteries; This is a new loss term for new class discovery and neighborhood-adaptive distribution alignment for Class C batteries, as described above. ;m is the battery category index with historically known grouping mechanisms, and this stage only includes categories A and B; The hyperparameter for the memory retention strength of the m-th type of battery; This is an index of the weights or biases of each specific neuron in the deep network parameters of the model; For the m-th type of battery, in terms of parameters Fisher information matrix element at the location; This is an attempt to update the parameters during the current optimization iteration for Class C batteries; After the historical training is completed, the parameters of the m-th type of battery are fixed at constant values ​​when they are in the optimal grouping state.

[0107] This secondary penalty term limits drastic changes in the model's core parameters, allowing the model to maintain the sorting rules for A and B batteries while being compatible with the refined unsupervised classification of C-class batteries.

[0108] In summary, this invention first extracts the general dimensionless aging characteristics of retired batteries as model input, thus avoiding feature confusion caused by differences in the physical decay characteristics of different electrochemical material systems. Then, a dual-branch decoupled network model is constructed and jointly trained using a dataset of batteries without aging state labels and known material properties. This not only enables the model to minimize the difference between the actual and predicted material categories but also effectively isolates material property interference while allowing the model to form stable aging state grouping rules in the latent space. Finally, a chi-square distribution hypothesis test is introduced to calculate statistical significance, enabling out-of-distribution detection of heterogeneous batteries. Soft probability weights are used to drive the maximum mean difference and elastic weights to consolidate constraints, allowing the trained aging state grouping rules to be unsupervisedly transferred across domains to unknown battery groups with novel materials. This method overcomes the dependence on battery data from different material systems and actual aging state labels, and can be efficiently deployed in intelligent cascade utilization and sorting lines for multi-source heterogeneous retired power batteries, achieving new battery access and adaptive sorting.

[0109] On the other hand, based on the same principle as the intelligent sorting method for multi-source heterogeneous retired batteries described in the above embodiments, the present invention also provides an apparatus for an intelligent sorting method for multi-source heterogeneous retired batteries. This apparatus includes a memory, a processor, and a computer program stored in the memory. The processor executes the computer program to implement the intelligent sorting method for multi-source heterogeneous retired batteries described in the above embodiments. Specifically, the apparatus can be an electronic computer or tablet computer, the processor can be a CPU, GPU, etc., and the memory can be RAM, ROM, EEPROM, CDROM, disk storage medium, or any other medium capable of carrying or storing the computer program and capable of being read by a computer; no limitation is made herein.

[0110] On the other hand, the present invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the intelligent sorting method for multi-source heterogeneous retired batteries described in the above embodiments. Specifically, the computer-readable storage medium may be RAM, ROM, EEPROM, SSD, CDROM, DVD, USB flash drive, or any other medium capable of carrying or storing a computer program and capable of being read by a computer.

[0111] On the other hand, the present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the intelligent sorting method for multi-source heterogeneous retired batteries described in the above embodiments.

[0112] To verify the dynamic sorting capability of the proposed method in a real recycling line, 218 retired batteries containing three different cathode material systems—LMO (lithium manganese oxide), NMC (lithium ternary lithium), and LFP (lithium iron phosphate)—were collected for sorting experiments. Significant differences were observed between the different systems in rated capacity and SOH (State of Health) distribution ranges, as shown in Table 1. In the experimental setup, initially only LMO batteries were used as a known class for static pre-training, and the LMO sets were divided into training and test sets in a 6:4 ratio.

[0113] Table 1

[0114]

[0115] In this experiment, the trigger threshold *n* represents the minimum number of unknown samples required for the system to confirm a new material family. When *n* is small, the buffer pool can trigger the discovery of new classes more quickly, but insufficient sample size may lead to unstable estimation of the new class distribution, making it susceptible to individual outliers or noise points. When *n* is large, the samples in the buffer pool can more fully characterize the statistical distribution of the unknown system. Therefore, *n* essentially reflects the trade-off between discovery speed and the reliability of class confirmation.

[0116] In this experiment, n=30, n=40, and n=50 were set. For the establishment of new classes and the global material identification confusion matrix after incremental updates for different n values, please refer to [link to relevant documentation]. Figure 4 Among them, such as Figure 4 As shown in (a), when n=30, the model correctly identifies both LMO and NMC, with all 38 LMO samples and 67 NMC samples correctly classified. However, for the subsequently accessed LFP samples, only 38 were correctly identified as LFP, while 18 were misclassified as LMO, resulting in an overall accuracy of 88.82%. Figure 4 As shown in (b), when the trigger threshold is increased to n=40, LMO and NMC still maintain completely correct identification, the number of correctly identified LFP increases from 38 to 40, and the number of samples misclassified as LMO decreases from 18 to 16, with an overall accuracy of 90.06%. Figure 4As shown in (c), when n=50, all three material classes were correctly identified, with 38 LMO samples, 67 NMC samples, and 56 LFP samples all located on the diagonal of the confusion matrix, resulting in an overall accuracy of 100%. This result indicates that as the trigger threshold is further increased, the number of samples in the buffer pool for new class confirmation and incremental updates becomes more sufficient, and the material boundary of the LFP new class is more stably solidified. Compared to n=30 and n=40, n=50 completely eliminates the phenomenon of LFP being misclassified as LMO, demonstrating that a sufficient accumulation of unknown samples is crucial for the stable identification of subsequent new classes.

[0117] Based on the identified material system, we will continue to refine the aging state within the model. Figure 5 The results of clustering aging states after model solidification classification are shown under different n values. Figure 5 The aging level classification clusters on the horizontal axis are 0, 1, and 2, which represent the relative aging levels defined within each material system. When the aging level classification cluster is 0, it corresponds to a battery with a low health status; when the aging level classification cluster is 1, it corresponds to a battery with a medium health status; and when the aging level classification cluster is 2, it corresponds to a battery with a high health status. Figure 5 Each dot represents a buffer sample used for buffer pool triggering and model solidification, and each black outlined square represents an inference sample that enters the online inference stage after model solidification.

[0118] Depend on Figure 5 (a) Figure 5 (b) and Figure 5 (c) It can be seen that within the LMO, NMC, and LFP material systems, the aging level classification clusters of 0, 1, and 2 generally exhibit a progressive relationship from low SOH to high SOH. This result indicates that the model in this invention can further stratify the health status of batteries within the same material system after material identification. Furthermore, most inference samples fall within the distribution range of buffer samples corresponding to the material system and the aging level, indicating that the aging grouping rules obtained from the buffer sample solidification can be effectively extended to subsequent online access samples. Therefore, the method proposed in this invention can form a two-level sorting result of material system identification and internal material aging grouping, providing a basis for material isolation and health consistency screening in the cascade utilization of retired batteries.

[0119] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A smart sorting method for multi-source heterogeneous retired batteries, characterized in that, Includes the following steps: Step S1: Obtain the dimensionless or incremental features of the retired battery to extract data characterizing the original aging features of the retired battery. Step S2: Construct a deep network containing a shared feature extractor and dual-branch outputs to form a dual-branch decoupled network model. The shared feature extractor maps the normalized original aging feature vector from the original feature space to the shared latent space, obtaining a shared representation that can be called by both branches. The model has two branches: a commonality learning branch, which outputs the battery's material category label, and a difference learning branch, which outputs the battery's relative aging state in its corresponding material category. The model is trained using a dataset of retired batteries of type A with known material properties but no aging state labels to obtain the grouping rules for type A aging states. The dataset is the data set obtained in step S1. When constructing the dual-branch decoupled network model, the total loss function for multi-task joint optimization of the model is jointly constructed by the material property classification loss function and the embedded deep clustering loss function. In the first branch of the model, a multivariate standard normal distribution is introduced as a variational prior constraint. The material property classification loss function is set in the first branch to constrain the consistency between the predicted probability calculated by the first branch and the known material category label. The embedded deep clustering loss function is set in the second branch to constrain the formation of differential aging state groupings in the unlabeled state. The model is trained using a dataset of retired Class A batteries with known material properties but no aging status labels. This dataset is the set of normalized raw aging feature vectors obtained using the method in step S1. The Class A retired battery dataset contains N Class A battery samples. During training: S201, first input the first sample into the model, and after processing by the shared feature extractor, obtain the shared representation of the sample that can be called by both branches simultaneously, as shown in the following formula: In the formula, lr and L are the set of trainable parameters of the shared feature extractor, the network layer number, and the total number of layers, respectively. Let be the normalized original aging feature vector of the i-th battery sample; Let i be the shared representation of the i-th sample, i = 1, 2, ..., N; These are the shared feature vectors output by the shared feature extractor for the initial layer, the lr-th layer, and the final layer, respectively. For the nonlinear activation function of the shared feature extractor; and These are the weight matrix and bias term of the lr layer, respectively; To share the feature extraction function, it consists of multiple layers of fully connected layers, convolutional layers, or recurrent units; S202, the shared representation of the obtained samples is simultaneously sent to branch one and branch two; 1) In branch one, the shared representation of samples is mapped through commonality. The material property latent vector of the sample is obtained as follows; In the formula, Let i be the latent vector of the material properties of the i-th sample; Trainable parameters for the branch-one commonality mapping; For branch one, the nonlinear activation function is used. and These are the weight matrix and bias term of the nonlinear activation function in branch 1, respectively; Subsequently, the latent vector of the material properties of the sample enters the classification layer, is processed by the Softmax activation function, and the predicted probability of the sample being classified as various material categories is calculated according to the following formula. Finally, the material category label corresponding to the highest predicted probability value is output from the classification layer as the material category label of the sample. In the formula, This represents the predicted probability that the i-th sample is classified as the c-th material property. The activation function for the classification layer; It is a natural exponential function; and These are the weights and biases corresponding to the c-th class in the classification layer, respectively. and These are the weights and biases corresponding to the r-th class in the classification layer, respectively; Let r represent the total number of retired battery material categories. When r = 1, 2, ..., the corresponding material category labels are A, B, ..., respectively; where r = 1 corresponds to category A, r = 2 corresponds to category B; in this training phase, the material category label of the sample is known to be A, and the predicted probability of the i-th sample being classified as a material of category 1 is given. The maximum value should be used; the classification layer will output the material category label for the i-th sample. ; 2) In branch two, the shared representation of samples is mapped via dissimilarity mapping. The aging state latent vector of the sample is obtained as follows; In the formula, Let be the latent vector of the aging state of the i-th sample; These are the trainable parameters for the branch-two difference mapping; For branch two, it is a nonlinear activation function; and These are the weight matrix and bias term of the branch two nonlinear activation function, respectively; Subsequently, the latent vector of the aging state of the sample is entered into the clustering module for unsupervised grouping to obtain the relative aging state of the sample; S203, input the remaining samples of the Class A retired battery dataset into the model in sequence, and repeat S201 and S202 to train the model. When all samples have been trained, the first branch of the model has the ability to minimize the difference between the real material category and the predicted material category, and the second branch forms and records several aging state grouping boundaries of Class A batteries, thereby obtaining the Class A aging state grouping rules. Step S3: Input the retired battery data with unknown material properties into the model trained in step S2. Identify whether the battery belongs to class A distribution through the distance metric in the feature space. If not, establish the battery as class B and proceed to step S4. The process for identifying whether retired batteries with unknown materials belong to type A distribution is as follows: First, the latent vectors of material properties of retired batteries with unknown material properties are extracted, and the squared value of their Euclidean distance is calculated. ; Next, we construct the hypothesis H0 that "this sample belongs to the A-class battery distribution." Under the condition that H0 is true, Calculate the statistical confidence probability that the retired battery belongs to the known class A battery distribution, strictly following a chi-square distribution. As shown in the following formula: In the formula, The statistical confidence probability value of retired battery B1, made of unknown materials, belonging to the known distribution of type A batteries; It is the cumulative distribution function of the chi-square distribution with d degrees of freedom; The lower incomplete gamma function; The standard gamma function; Finally, a standard statistical significance level is set. If the calculated Then reject the null hypothesis H0, and use The confidence level is determined to be "the retired battery does not belong to Class A batteries"; Step S4: Using an unsupervised domain adaptive method, the obtained A-class aging state grouping rules are aligned and applied to the B-class battery space to achieve unsupervised sorting of B-class batteries. Step S401: Calculate the new class soft weight of the retired battery based on the statistical confidence probability value of the retired battery belonging to the known Class A battery distribution with unknown material properties: If the statistical confidence probability value of a retired battery belonging to the Class A battery distribution is... If the significance level is below the preset threshold, the soft weight of the new class approaches 1. If the new class soft weight is higher than the significance threshold, its new class soft weight approaches 0; the model uses the new class soft weight to perform weighted density clustering on the material property latent vector of no less than 50 retired battery samples of the unknown material. When the samples with high new class soft weights gather into a stable high-density feature cluster, the cluster is finally confirmed as a B-class battery, representing a new type of retired battery, and the dataset of the cluster is locked. In step S402, the model unsupervisedly transfers the aging state grouping rules of class A obtained in step S2 to class B batteries, aligning the marginal distribution of aging state features of class B battery clusters with the feature distribution of class A batteries. The new class soft weights obtained in step S401 are used as alignment penalty coefficients to construct a weighted maximum mean difference loss function, as shown in the following equation: In the formula, The maximum mean difference loss value of cross-domain aging state features is introduced to minimize the distribution distance between class A and class B batteries in the aging state space. The kernel function that maps low-dimensional latent features to the reproducing kernel Hilbert space; and These are the i-th type A batteries. and the j-th type B battery The latent vector of aging state; For the j-th class B battery after normalization The original aging feature vector; For the j-th type B battery A new type of soft weight; This is a dataset for battery clusters of type B. Step S403: After the alignment training optimization converges, fix the hidden cluster centers learned in step S2. For any retired battery belonging to class B, without manual aging state labeling, calculate the soft assignment probability of its aging state latent vector relative to each hidden cluster center. Take the cluster number corresponding to the maximum soft assignment probability to refine the class B battery into the corresponding aging state group, completing the unsupervised cross-domain transfer of aging state rules. Step S5: Repeat steps S3 and S4 to input more classes of retired battery data into the model to optimize the model, so as to realize the identification and sorting of retired batteries of unknown materials.

2. The intelligent sorting method for multi-source heterogeneous retired batteries according to claim 1, characterized in that, In step S1, the charging and discharging dynamic sequence of retired batteries is collected, and the general aging features that can characterize the electrochemical evolution trajectory are extracted using differential and integral transformations. These features include the peak position shift of the capacity increment curve, the relative change rate of the peak area, and the characteristic valley capacity spacing of the differential voltage curve. The extracted general aging features are then globally normalized to obtain the original aging feature vector, which is used as the input to the dual-branch decoupling network model. 1) The peak position shift of the capacity increment curve, including the main peak voltage position V. peak With the height of the main peak H peak As shown in the following formula: In the formula, Indicates to make Voltage at maximum value The possible values ​​of ; express The maximum value; It represents the capacity increment under unit voltage change, and the peak value of its curve reflects the phase transition plateau where the electrode material undergoes the main redox reaction. 2) The relative rate of change of the peak area of ​​the capacity increment curve is the integral area A of the main peak region. peak As shown in the following formula: In the formula, and The starting and ending voltage boundaries of the main peak of the capacity increment curve are determined using the second derivative zero-finding method. 3) Characteristic valley spacing of differential voltage curves As shown in the following formula: In the formula, and These correspond to the preset capacity search intervals where two adjacent characteristic valleys on the differential voltage curve are located, respectively. and respectively make When taking the minimum value and The value of ; The voltage change represents the voltage change per unit capacity, and the peak and valley positions of its curve are highly correlated with the loss of active lithium and the increase in internal resistance of the electrode material.

3. The intelligent sorting method for multi-source heterogeneous retired batteries according to claim 2, characterized in that, The total loss function of the dual-branch decoupled network model for multi-task joint optimization is expressed as follows: In the formula, The total loss value for multi-task joint optimization; The loss function for classifying material properties; Commonality mapping of Class A battery samples through network branch one The resulting implicit vector of material properties; The known material category labels for class A battery samples are given. During the initial training phase, it is known that the material category label for all samples is A. For embedded deep clustering loss function; Differential mapping of class A battery samples through network branch two The resulting latent vector of aging state; for and The mutual information penalty term is used to quantitatively measure the statistical correlation between the two sets of features. This term needs to be minimized during model training to achieve decoupling between material and aging features. This is a balance coefficient for mutual information constraints, used to adjust the weight ratio between clustering accuracy and feature decoupling degree in the aging state of the model; As variational prior constraints, calculate the posterior distribution of the output of branch one. Compared with the multivariate standard normal distribution The KL divergence between them will standardize the material characteristics of Class A batteries within a unit hypersphere; For Given the condition, the probability distribution of latent features of material properties derived by the network encoder. This is the normalized original aging feature vector of the Class A battery; The equilibrium coefficients for variational prior constraints; The material property classification loss function uses cross-entropy loss, as shown in the following formula: In the formula, Let c be the known material category label for the i-th sample. When c=1,2,..., the corresponding material category labels are A,B,..., respectively; where c=1 corresponds to class A, and c=2 corresponds to class B. This represents the cross-entropy loss value; N represents the number of samples. The embedded deep clustering loss function adopts KL divergence loss, and its construction process is as follows: First, calculate the soft assignment probability that the aging state latent vector of each battery sample belongs to a certain latent cluster center; In the formula, The soft assignment probability is that the i-th type A battery sample belongs to the j-th aging state group. The difference mapping of the i-th class A battery sample through model branch two The resulting latent vector of aging state; Let be the optimized latent cluster center vector in the dissimilarity latent space, where K is the number of hidden cluster centers; represents the degrees of freedom parameter of the t-distribution; This represents the square of the Euclidean distance, used to measure the spatial distance between the latent vector and the latent cluster center; Next, a target distribution is constructed based on the soft assignment probability to reinforce the grouping tendency of high-confidence samples and normalize the loss; In the formula, The target probability, calculated based on the soft assignment probability, is used to guide the agglomeration of cluster centers; Finally, the embedded deep clustering loss is used. Represented as the KL divergence between the target probability and the current soft assignment probability; The and Mutual information penalty item The expression is as follows: In the formula, Implicit vectors for material properties Latent vectors of aging state The joint probability distribution, and They are respectively and The marginal probability distribution.

4. The intelligent sorting method for multi-source heterogeneous retired batteries according to claim 3, characterized in that, Step S5 combines memory replay and regularization mechanisms to ensure continuous learning and continuous model optimization. Elastic weights are introduced to reinforce constraints during incremental model updates. In the formula, The incremental optimization objective total loss function incorporates a continuous learning mechanism; This is the set of variable parameters for the model network currently being prepared for adaptive iteration targeting the new type of battery; A new loss term is added for discovering new battery types and aligning them with the domain-adaptive distribution; m is the index of battery categories with historically known grouping mechanisms; The hyperparameter for the memory retention strength of the m-th type of battery; This is an index of the weights or biases of each specific neuron in the deep network parameters of the model; For the m-th type of battery, in terms of parameters Fisher information matrix element at the location; This refers to the attempted updated values ​​of parameters during the current optimization iteration for the new type of battery; After the historical training is completed, the parameters of the m-th type of battery are fixed at constant values ​​when they are in the optimal grouping state.

5. An intelligent sorting device for multi-source heterogeneous retired batteries, comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: The processor executes the computer program to implement the intelligent sorting method for multi-source heterogeneous retired batteries as described in any one of claims 1-4.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that: When executed by a processor, the computer program implements the intelligent sorting method for multi-source heterogeneous retired batteries as described in any one of claims 1-4.

7. A computer program product, comprising a computer program, characterized in that: When executed by a processor, the computer program implements the intelligent sorting method for multi-source heterogeneous retired batteries as described in any one of claims 1-4.

Images