A method, device and medium for screening retired batteries based on fuzzy clustering analysis

Abstract

The application discloses a retired battery screening method and device based on fuzzy clustering analysis, and a medium. The method comprises the following steps: obtaining the charge-discharge data of a retired battery cell to be screened; training a feedforward neural network according to the charge-discharge data; fitting an incremental capacity curve corresponding to the retired battery cell based on the trained feedforward neural network; constructing a multi-dimensional health feature corresponding to the retired battery cell according to the incremental capacity curve; generating a battery dataset carrying a battery classification label according to the distribution relationship between the health state value output by the health state estimation model and the preset health state threshold; determining an initial clustering center based on the probability density of each sample point in the battery dataset; inputting the initial clustering center and the battery dataset into a preset semi-supervised fuzzy C-means clustering model for iterative calculation, so as to classify and screen the retired battery cell according to the output clustering result.

Description

Technical Field

[0001] This application relates to the field of retired battery sorting technology, specifically to a retired battery screening method, equipment, and medium based on fuzzy clustering analysis. Background Technology

[0002] With the rapid development of the electric vehicle and energy storage industries, the performance of lithium-ion power batteries inevitably degrades, necessitating the reuse of large quantities of retired lithium-ion batteries to achieve resource recycling. However, due to significant performance inconsistencies in retired batteries, directly reassembling and reusing individual cells can easily lead to safety issues such as localized overcharging, over-discharging, and even thermal runaway. Therefore, precise sorting is essential before reusing retired batteries.

[0003] Traditional sorting technologies for retired lithium batteries primarily rely on measurements of external, static, and single-point parameters, inherently lacking the ability to simultaneously achieve high efficiency, accuracy, and comprehensiveness. The most commonly used capacity and internal resistance sorting methods typically measure the remaining capacity and DC internal resistance of the battery under fixed conditions. This method only obtains limited static parameters and cannot reflect the dynamic characteristics and aging trajectory of the battery under real-world operating conditions, resulting in insufficient sorting dimensions and potential performance differentiation even within the same group of batteries in actual use. Methods based on open-circuit voltage or simple electrochemical impedance spectroscopy often suffer from low accuracy, suitable only for coarse screening of retired batteries, and are time-consuming and costly, failing to meet the sorting efficiency requirements for large-scale retired batteries. Furthermore, some methods employ threshold grading or K-means clustering algorithms for classification. These methods are insufficient for handling ambiguous samples at performance boundaries, and forced classification is prone to misjudgment, failing to accurately depict the continuous distribution of battery performance and affecting the consistency and reliability of battery sorting. Summary of the Invention

[0004] To address the aforementioned issues, this application proposes a method for screening decommissioned batteries based on fuzzy clustering analysis, comprising: Obtain the charge and discharge data of the retired battery cells to be screened, train the feedforward neural network based on the charge and discharge data, and fit the incremental capacity curve corresponding to the retired battery cells based on the trained feedforward neural network. Based on the incremental capacity curve, construct the multidimensional health characteristics corresponding to the retired battery cell; The multidimensional health features are input into a preset health status estimation model. Based on the distribution relationship between the health status values ​​output by the health status estimation model and the preset health status threshold, a battery dataset carrying battery classification labels is generated. The battery dataset includes the multidimensional health features of each retired battery cell. Based on the probability density of each sample point in the battery dataset, the initial cluster centers are determined; The initial cluster centers and the battery dataset are input into a preset semi-supervised fuzzy C-means clustering model for iterative calculation, so as to classify and screen the retired battery cells based on the output clustering results.

[0005] In one implementation of this application, a feedforward neural network is trained based on the charge / discharge data, and the incremental capacity curve corresponding to the retired battery cell is fitted based on the trained feedforward neural network, specifically including: The feedforward neural network is trained based on the charging and discharging data; wherein the input layer of the feedforward neural network is the battery voltage, the output layer is the battery capacity, and the hidden layer contains two neurons. Based on the neuron parameters of the feedforward neural network obtained through training, the relationship function between battery voltage and battery capacity is determined; The derivative of the relationship function is calculated, and the incremental capacity value obtained from the derivative and the battery voltage are used to fit the incremental capacity curve corresponding to the retired battery cell.

[0006] In one implementation of this application, a battery dataset carrying battery classification labels is generated based on the distribution relationship between the health status values ​​output from the health status estimation model and a preset health status threshold. Specifically, this includes: Based on the target application scenario of the retired battery cell, a preset health status threshold is determined, and a first boundary value and a second boundary value are respectively formed by the upper and lower distribution threshold boundaries of the health status threshold. For each retired battery cell, if the health status value is between the first boundary value and the second boundary value, the retired battery cell is not labeled. Otherwise, based on the relationship between the health status value and the first and second boundary values, a battery classification label corresponding to the retired battery cell is generated. A battery dataset is constructed based on the health status values ​​of all retired battery cells and the battery classification labels.

[0007] In one implementation of this application, the initial cluster centers and the battery dataset are input into a preset semi-supervised fuzzy C-means clustering model for iterative calculation, so as to classify and screen the retired battery cells based on the output clustering results, specifically including: The initial cluster centers and the battery dataset are input into a preset semi-supervised fuzzy C-means clustering model; The initial cluster center is used as the cluster center for the current iteration. A membership matrix is ​​generated based on the distance between each sample point in the battery dataset and the cluster center. The cluster centers are updated according to the membership matrix, and the updated cluster centers are used as the cluster centers for the next iteration. Repeat the above membership calculation and cluster center update steps until the objective function corresponding to the semi-supervised fuzzy C-means clustering model converges or the number of iterations reaches the preset maximum number of iterations, and output the corresponding clustering results; wherein, the clustering results are the membership of each retired battery cell to each cluster center; The retired battery cells are classified and screened based on the battery classification label corresponding to the cluster center with the highest membership degree.

[0008] In one implementation of this application, the membership calculation and cluster center update steps described above are repeated until the objective function corresponding to the semi-supervised fuzzy C-means clustering model converges or the number of iterations reaches a preset maximum number of iterations, specifically including: Based on the Euclidean distance between each sample and each cluster center in the battery dataset, and the membership degree of each sample to each cluster center, an unsupervised term is constructed; For the labeled samples carrying battery classification labels in the battery dataset, a supervision term is constructed based on the degree of difference between the current membership degree of the labeled sample and its corresponding membership degree reference value. The unsupervised term and the supervised term are weighted and summed to obtain the objective function corresponding to the semi-supervised fuzzy C-means clustering model. Minimizing the objective function is used as the iteration objective. The above membership calculation and cluster center update steps are repeated until the objective function corresponding to the semi-supervised fuzzy C-means clustering model converges or the number of iterations reaches the preset maximum number of iterations.

[0009] In one implementation of this application, determining the initial cluster centers based on the probability density of each sample point in the battery dataset specifically includes: Using a preset Gaussian kernel function, the probability density corresponding to each sample point in the battery dataset is calculated, and local density peak points in the sample points whose probability density is greater than that of all adjacent sample points are identified. Based on the preset number of clusters, select several local density peaks with the highest probability density from the local density peaks as the initial cluster centers.

[0010] In one implementation of this application, before inputting the multidimensional health features into a preset health status estimation model, the method further includes: Initialize a firefly population; wherein the firefly population consists of several individuals, each of which corresponds to a set of initial connection weights and offsets of a backpropagation neural network; Based on the relative brightness relationship between the individuals, the firefly population is iterated so that individuals with lower brightness move to individuals with higher brightness, until the preset maximum number of iterations is reached. The connection weights and offsets of the brightest individual in the iterated firefly population are used as the optimal initial parameters, and the backpropagation neural network is initialized based on the optimal initial parameters. The backpropagation neural network is trained using a pre-built training sample set to obtain the health status estimation model.

[0011] In one implementation of this application, the relational function is represented as: ; in, v' represents the normalized battery capacity, and v′ represents the normalized battery capacitance. This represents the weights from the input layer to the hidden layer. It is the bias term of the i-th neuron. This represents the weights from the hidden layer to the output layer. It is the bias term of the output layer.

[0012] This application provides a decommissioned battery screening device based on fuzzy clustering analysis, the device comprising: At least one processor; And, a memory communicatively connected to the at least one processor; The memory stores instructions that can be executed by the at least one processor, which, when executed by the at least one processor, enables the at least one processor to perform a decommissioned battery screening method based on fuzzy clustering analysis as described above.

[0013] This application provides a non-volatile computer storage medium storing computer-executable instructions, wherein the computer-executable instructions are configured as follows: A method for screening decommissioned batteries based on fuzzy clustering analysis, as described in any of the preceding items.

[0014] The decommissioned battery screening method based on fuzzy clustering analysis proposed in this application can bring the following benefits: By constructing incremental capacity curves through neural network fitting, the complex aging process inside the battery is transformed into quantifiable curve morphology features. Compared to traditional methods that rely solely on static parameters, this approach comprehensively captures the dynamic aging trajectory of the battery under real-world operating conditions. Initial cluster centers are determined by the probability density of each sample point in the battery dataset, ensuring that clustering begins with dense regions based on the data's own distribution. This significantly reduces the number of iterations required for convergence and ensures consistency in sorting results across different batches, effectively suppressing fluctuations in clustering results caused by initial randomness. A semi-supervised fuzzy C-means clustering model guides the clustering direction using a small amount of labeled data while fully preserving the objective distribution information of unlabeled data. This effectively solves the problem of misclassification caused by forced categorization of performance boundary samples, achieving accurate sorting of retired batteries. Attached Figure Description

[0015] The accompanying drawings, which are included to provide a further understanding of this application and form part of this application, illustrate exemplary embodiments and are used to explain this application, but do not constitute an undue limitation of this application. In the drawings: Figure 1 A flowchart illustrating a method for screening decommissioned batteries based on fuzzy clustering analysis, provided in an embodiment of this application; Figure 2 This is a schematic diagram of a decommissioned battery screening device based on fuzzy clustering analysis, provided as an embodiment of this application. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be clearly and completely described below in conjunction with specific embodiments and corresponding drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0017] The technical solutions provided by the various embodiments of this application are described in detail below with reference to the accompanying drawings.

[0018] like Figure 1 As shown in the embodiments of this application, the decommissioned battery screening method based on fuzzy clustering analysis includes: S101: Obtain the charge and discharge data of the retired battery cells to be screened, train the feedforward neural network based on the charge and discharge data, and fit the incremental capacity curve corresponding to the retired battery cells based on the trained feedforward neural network.

[0019] Standardized charge-discharge tests are conducted on the selected retired lithium-ion battery cells. Constant current charging is performed at a set rate under constant temperature conditions, and the battery voltage and cumulative charging capacity are collected simultaneously during the charging process to form continuous and complete charge-discharge data. Retired battery cells refer to lithium-ion battery cells that have completed their initial design life or have been removed from electric vehicles, energy storage systems, or other equipment. Their performance has degraded to varying degrees, and they need to be sorted and evaluated before they can be reused. After collecting the charge-discharge data, this application proposes a neural network fitting (NNF) method to describe the relationship between battery voltage and capacity. The aim is to train a feedforward neural network based on the charge-discharge data, using battery voltage as input layer data and battery capacity as output layer data. By continuously adjusting the weights and biases within the network, the network accurately learns the nonlinear mapping relationship between voltage and capacity, completing network training and achieving the preset accuracy requirements. Neuron parameters, including connection weights from the input layer to the hidden layer, bias terms of hidden layer neurons, connection weights from the hidden layer to the output layer, and bias terms of the output layer, directly reflect the capacity of the voltage plateau. This establishes a mapping relationship between voltage and capacity, allowing for the fitting of the incremental capacity (IC) curve corresponding to the retired battery cell. This curve clearly presents the voltage plateau characteristics related to battery aging, providing a stable and high-fidelity curve foundation for subsequent health feature extraction.

[0020] In one embodiment, the feedforward neural network (FNN) is a multi-layer neural network structure with battery voltage as the input layer, battery capacity as the output layer, and two neurons in the hidden layer, using the sigmoid function as the activation function. The sigmoid function is a monotonically differentiable nonlinear function that can smoothly map the input value to the range of 0 to 1, and its derivative exhibits a single-peak shape. This characteristic allows the neural network to more accurately capture the nonlinear variation characteristics of the voltage plateau region when fitting the voltage-capacity curve.

[0021] A feedforward neural network is trained using collected charge and discharge data, with the training objective of minimizing the mean square error between the predicted and actual capacity. The trained feedforward neural network accurately describes the nonlinear relationship between battery voltage and battery capacity. Based on the neuron parameters of the trained feedforward neural network, the relationship function between battery voltage and battery capacity can be determined. This relationship function is expressed as: ; in, v' represents the normalized battery capacity, and v′ represents the normalized battery capacitance. This represents the weights from the input layer to the hidden layer. It is the bias term of the i-th neuron. This represents the weights from the hidden layer to the output layer. It is the bias term of the output layer.

[0022] IC value, or incremental capacity value, is determined by... The derivative is obtained through differentiation. Since the relational function is composed of a combination of Sigmoid neurons, its derivative can be accurately calculated analytically, avoiding curve oscillations caused by data sampling noise in traditional numerical difference methods. This effectively suppresses data noise and accurately highlights the voltage plateau characteristics, i.e., the peak value of the IC curve. Based on this, the incremental capacity curve corresponding to the retired battery cell is fitted with battery voltage as the x-axis and the incremental capacity value obtained after differentiation as the y-axis. The incremental capacity curve can clearly show the peak value of capacity change corresponding to the voltage plateau region during the charging process, i.e., the peak point of the incremental capacity curve. These peak points directly reflect the phase transition process of the internal electrochemical reaction of the battery and are key features characterizing the aging state of the battery. The incremental capacity curve obtained by fitting and differentiating through a feedforward neural network has good smoothness and the feature points are easy to identify, providing a reliable data foundation for subsequent extraction of multi-dimensional health features.

[0023] S102: Based on the incremental capacity curve, construct the multidimensional health characteristics corresponding to the retired battery cells.

[0024] The aging of retired batteries is a complex electrochemical process. In practical applications, measuring parameters such as capacity and internal resistance of retired batteries often results in lengthy sorting times, making it difficult to apply to the sorting of large quantities of retired batteries. Furthermore, relying on a single parameter cannot fully characterize the battery performance degradation trajectory. Since there is an objective correlation between the State of Health (SOH) of lithium-ion batteries and their charge-discharge curves, ICMAX (the peak point of the IC curve) is extracted based on the fitted incremental capacity curve as a feature for evaluating the aging performance of individual retired battery cells. Simultaneously, to more comprehensively describe the battery's health state, the IC curve peak point is integrated with macroscopic performance parameters such as discharge energy and temperature to construct a multidimensional health feature vector. This multidimensional feature vector comprehensively reflects the battery's internal material phase transitions, polarization characteristics, energy output capacity, and thermal behavior from multiple dimensions, enabling a more comprehensive capture of the battery's true aging state. This avoids the limitations of single-parameter characterization and significantly shortens testing time while ensuring rich information.

[0025] S103: Input the multidimensional health features into the preset health status estimation model, and generate a battery dataset with battery classification labels based on the distribution relationship between the health status values ​​output by the health status estimation model and the preset health status threshold; wherein, the battery dataset includes the multidimensional health features of each retired battery cell.

[0026] The aforementioned multidimensional health features are input into a pre-defined health status estimation model. This model is constructed using a backpropagation (BP) neural network optimized with the firefly algorithm, which accurately outputs the current capacity retention rate, i.e., the health status value, based on the input multidimensional health features. After obtaining the health status estimates of all retired battery cells, these estimates are further analyzed for distribution and labeled according to a pre-defined health status threshold, thereby generating a battery dataset with battery classification labels for subsequent cluster center initialization. The battery dataset includes not only the multidimensional health features of each retired battery cell but also the battery classification labels obtained after thresholding, such as high capacity and low capacity.

[0027] In one embodiment, before inputting multidimensional health features into the health status estimation model, this application embodiment first pre-trains and optimizes the health status estimation model to ensure that the model has high-precision estimation capabilities. The health status estimation model uses a BP neural network as its core architecture, and the selection of its initial connection weights and offsets has a decisive impact on the training convergence speed and final prediction accuracy of the model. To overcome the problems of traditional neural network random initialization methods easily getting trapped in local optima and insufficient model generalization ability, the Glowworm swarm algorithm (GSO) is introduced to globally optimize the initial parameters of the BP neural network.

[0028] Step 1: The firefly algorithm is constructed based on the luminous intensity of individual fireflies and the degree of attraction between them. First, the relative brightness between each firefly is determined, as shown in the following formula: ; in, This represents the Cartesian distance between two fireflies, BP. λ represents the brightness at a distance r=0 (brightest), i.e., its own brightness, which is linked to the objective function value; the better the objective value, the brighter the brightness. λ is the absorption coefficient. Individual light intensity decreases with increasing distance and absorption by the propagation medium. Therefore, the algorithm sets an absorption coefficient to represent this characteristic, which can generally be set to a constant based on experience.

[0029] Step 2: Initially introduce a trained firefly population. Based on the set population size, construct a population with the corresponding number of individuals in the sample set S={( u 1, y 1),( u 2, y 2),( u 3, y 3),…,( u k , y k ),( x 1, x 2,…, x k ) ∈ u} Randomly select m samples as the population size of fireflies. u k and y k For the first k The input feature set and output health status value for each sample. x n It is a multidimensional health characteristic.

[0030] Step 3: The firefly population consists of several individuals. Each individual corresponds to a set of initial connection weights and offsets in a backpropagation neural network. The position of each firefly individual is randomly determined to be within the range of values ​​of the independent variable of the scaling function, as shown in the following formula:

[0031] Where t is the number of iterations. This represents the current location of the sample individual. Let Randt be the brightness of the sample individual, and (Randt, -1 / 2) be a random value. Randt ∈ [0, 1] is a random number uniformly distributed between [0, 1]. Let the step size be , and the expression be . , The initial value is 0.5, and the equation stops at the maximum number of iterations Gmax.

[0032] Step 4: Calculate the brightness of the population, i.e., target evaluation. For maximization problems, the evaluation process directly proportionalizes the target value to the brightness; for minimization problems, the opposite is true, with the target value being the reciprocal of the brightness. After processing, the individuals in the population are sorted according to the solved objective function to determine the optimal individual in the current population. In this embodiment, the objective function is set as the prediction error of the neural network on the validation set. error=|Out(u k )-y k |The minimum value is taken as the optimal solution. The smaller the error, the higher the brightness. Thus, the optimization problem of network parameters is transformed into the problem of maximizing the brightness of individual fireflies in the population.

[0033] Step 5: After population initialization, iterative optimization is performed based on the relative brightness relationships between individuals, causing individuals with lower brightness to move towards those with higher brightness. In each iteration, each firefly determines its movement direction based on the difference in brightness between itself and other individuals. Individuals with lower brightness are attracted to those with higher brightness and move towards them, thus approaching a region with better parameters; while the brightest individuals in the population move randomly to avoid getting trapped in local optima. This movement process is specifically represented as follows:

[0034]

[0035] in, , , These represent low-brightness samples, high-brightness samples, and the current best sample, respectively.

[0036] Step 6: Update the iterator and step size Number of iterations t = t + 1.

[0037] Step 7: Repeat steps 4-6 above until the preset maximum number of iterations is reached. At this point, output the optimal solution, which is the initial connection weights and offsets under the minimum value of the objective function error.

[0038] Step 8: Input the connection weights and offsets of the brightest individual in the firefly population after iteration as the optimal initial parameters into the BP neural network to initialize the BP neural network.

[0039] Step 9: Train the backpropagation neural network using a pre-built training sample set to obtain a health status estimation model.

[0040] Specifically, a BP neural network is a multi-layer feedforward neural network, consisting of an input layer, hidden layers, and an output layer. The processing of a BP neural network mainly consists of two stages: the first stage is the forward propagation of the signal, from the input layer through the hidden layers to the output layer; the second stage is the backward propagation of the error.

[0041] Introducing a pre-trained training sample set L=(x i ,y i (i=1,2,…,n) , x iThese are the input parameters for each training sample, which include multi-dimensional health characteristics such as ICMAX, discharge energy, and temperature of retired battery cells. y i is the actual health status value, and n is the total number of training samples.

[0042] The outputs of the hidden layer and the output layer are calculated as shown in the following formula:

[0043] in, w i,j and w j,k These are the weights for the hidden layer and the output layer, respectively. a j and b k These are the bias values ​​for the hidden layer and the output layer, respectively. The default activation functions for the hidden layer and the output layer are... f(x) The value is taken as the Sigmoid function, and its formula is as follows:

[0044] The error between the expected output and the actual output is calculated using the following formula:

[0045] in, and These represent the expected output and the actual output, respectively.

[0046] Based on the calculation error, the weights are corrected to optimize the BP neural network. Specifically, this is expressed as follows:

[0047] in, The learning rate is determined by subsequent optimization algorithms.

[0048] Furthermore, the network bias is updated, and the offset is updated in a timely manner according to the calculation error. The specific update process is as follows:

[0049] Perform the above steps in a loop until... e k The requirement of setting the expected error value is met. At this point, the optimized health status estimation model is obtained.

[0050] In one embodiment, after outputting the health status value of a retired battery cell through the above-mentioned health status estimation model, it is necessary to further assign a corresponding battery classification label to each retired battery cell based on the distribution relationship between the health status value and the preset status threshold.

[0051] Specifically, based on the target application scenario of the retired battery cells, a preset health status threshold is determined. For example, for energy storage applications, 80% can be set as the sorting baseline. This threshold reflects whether the battery meets the minimum remaining capacity requirements of the target application scenario. After determining the health status threshold, combined with data distribution characteristics, certain floating ranges are set above and below the health status threshold, thus forming a first boundary value and a second boundary value composed of the upper and lower distribution threshold boundaries, respectively. The first boundary value is higher than the health status threshold, and the second boundary value is lower than the health status threshold. The region between the first and second boundary values ​​constitutes a fuzzy interval; samples within this interval are in a transitional performance zone and are difficult to classify accurately using a single threshold.

[0052] Furthermore, for each retired battery cell, a label is assigned based on the relationship between its estimated health status and the two boundary values ​​mentioned above. If the health status value of a retired battery cell falls between the first and second boundary values, i.e., it falls into a fuzzy region, its true category is difficult to determine accurately using a single threshold. Forcibly assigning a classification label could easily lead to misjudgment. Therefore, no classification label is assigned to this retired battery cell, and it is retained as unlabeled data. If the health status value is higher than the first boundary value, it indicates that the battery performance is significantly better than the threshold requirement, belonging to a high-performance sample, and a corresponding high-capacity category label is generated for it. If the health status value is lower than the second boundary value, it indicates that the battery performance is significantly worse than the threshold requirement, belonging to a poor-performance sample, and a corresponding low-capacity category label is generated for it. In this way, only battery cells with clear performance characteristics are labeled, while battery cells in the fuzzy boundary region retain their objective status.

[0053] After labeling all retired battery cells, the health status values ​​and corresponding classification labels of all retired battery cells are summarized to construct a battery dataset. In the battery dataset, some samples carry explicit battery classification labels, which can serve as supervisory information to guide the subsequent clustering process. Some samples are unlabeled, preserving the original form of the data distribution and providing the possibility for subsequent semi-supervised clustering models to fully utilize the information from unlabeled data.

[0054] S104: Determine the initial cluster centers based on the probability density of each sample point in the battery dataset.

[0055] After constructing the battery dataset, cluster analysis is performed to achieve rapid sorting of retired battery cells. The core of the clustering algorithm lies in grouping samples with similar distributions in the feature space into the same category, and the initial position of the cluster centers directly determines the convergence speed of the algorithm and the stability of the final classification results. Traditional clustering methods usually use random initialization to select initial cluster centers, which can easily lead to large fluctuations in clustering results. Moreover, due to the deviation between the initial centers and the true distribution centers of the data, a large number of iterations are often required for convergence, and it may even get stuck in local optima, affecting the consistency and reliability of sorting. Therefore, this embodiment analyzes the probability density distribution of each sample point in the battery dataset and adaptively determines the initial cluster centers from the dense regions of the data distribution itself. This can adaptively capture the true distribution characteristics of the data samples, and is especially suitable for irregular and unevenly distributed data. By identifying the peak of the density distribution to initialize the cluster centers, a stable, reliable, and near-global optimal starting point is provided for the subsequent clustering iteration process.

[0056] In one embodiment, a Gaussian kernel function is used as the core function for kernel density estimation to calculate the probability density corresponding to each sample point in the battery dataset. The core of kernel density estimation lies in the nonparametric estimation of the probability density function, given the battery dataset. At any sample point x i probability density estimate at The specific calculation process is as follows: ; in, is the kernel function, usually a Gaussian kernel function that satisfies symmetry, non-negativity and has an integral of 1; h is the bandwidth parameter, which controls the smoothness of the density estimation, and its selection is optimized through empirical rules or cross-validation; d is the dimension of the health factor vector.

[0057] The physical meaning of the above formula is that by placing a kernel function at each sample point and superimposing its contributions, a continuous probability density distribution surface of the data in the entire space can be fitted. The peak region of the density surface corresponds to the location of the prototype point where the data samples are most concentrated and most likely represent different cluster cores.

[0058] Based on the probability density calculation method described above, local density peaks are identified where the probability density of a sample point is greater than that of all its neighboring sample points. In other words, for each sample point, its probability density is compared with the probability densities of all its neighboring sample points. If the probability density of the current sample point is greater than that of all its neighboring sample points, then the sample point is marked as a local density peak. These local density peaks correspond to the apex positions of the probability density distribution surface, representing the most concentrated areas of data samples and serving as ideal candidate points for cluster centers. This identification process does not require a preset density threshold and is completed automatically based entirely on the distribution characteristics of the data itself, ensuring the objectivity and adaptability of the initial center selection.

[0059] After identifying all local density peaks, based on a preset number of clusters, the few local density peaks with the highest probability densities are selected as the final initial cluster centers. For example, if the preset number of clusters is k, then the k points with the highest probability densities from all local density peaks are selected as the initial cluster centers. If the number of candidate cluster centers is too large, a minimum Euclidean distance constraint is further introduced for filtering to ensure that the selected initial cluster centers have sufficient discriminative power in the feature space.

[0060] The initial cluster centers determined by the kernel density estimation optimization strategy described above ensure that the selection of initial cluster centers has high stability and repeatability. The sorting results of different batches of batteries will not fluctuate due to initial randomness. At the same time, since the initial cluster centers are close to the optimal cluster centers, the subsequent iteration process of the semi-supervised fuzzy C-means clustering algorithm greatly reduces the number of unnecessary iterations and significantly improves the overall convergence efficiency of the screening process.

[0061] S105: Input the initial cluster centers and battery dataset into a pre-defined semi-supervised fuzzy C-means clustering model for iterative calculation, so as to classify and screen retired battery cells based on the output clustering results.

[0062] The initial cluster centers obtained through kernel density estimation, along with the pre-constructed battery dataset containing both labeled and unlabeled data, are input into a pre-defined semi-supervised fuzzy C-means clustering model. The clustering results calculated iteratively are used to achieve the final classification and screening of retired battery cells.

[0063] Semi-supervised fuzzy C-means clustering model (SSFCM) is an improved clustering method based on the traditional fuzzy C-means clustering algorithm, which introduces a supervision mechanism. Its core idea is to maintain the flexibility of fuzzy clustering in handling boundary samples, while using the category information of a small amount of labeled data to guide the clustering process to converge in a more reasonable direction, thereby improving the accuracy and stability of the clustering results.

[0064] The initial cluster centers and the battery dataset are input into a pre-defined semi-supervised fuzzy C-means clustering model. During model initialization, the initial cluster centers are used as the cluster centers for the current iteration, and iterative calculations begin based on these centers. In each iteration, the Euclidean distance between all sample points in the battery dataset and each cluster center is calculated based on the cluster centers for the current iteration. A membership matrix is ​​then generated, where each element represents the degree to which a sample belongs to a particular category. The value ranges from 0 to 1, reflecting the fuzzy nature of cluster analysis and preventing misclassification of samples at category boundaries due to rigid partitioning. The generation of the membership matrix follows the basic principles of fuzzy C-means clustering: the closer a sample is to a cluster center, the higher its membership to that category; and for any sample, the sum of its memberships to all cluster centers is always equal to 1.

[0065] Based on the membership matrix calculated in the current iteration, all cluster centers are recalculated. Specifically, the updated value of each cluster center is determined by the weighted average position of all samples in its corresponding category, where the weights are the membership degrees of each sample to that category. This update mechanism causes the cluster centers to gradually move towards denser sample areas, thereby continuously optimizing the representative positions of the categories. After the update is complete, the newly obtained cluster centers are used as the cluster centers for the next iteration, preparing for subsequent iterations.

[0066] To achieve precise control of the iterative process, this application constructs a semi-supervised fuzzy C-means clustering model objective function, which is a weighted sum of unsupervised and supervised terms. The unsupervised term is constructed based on the Euclidean distance between each sample and each cluster center in the battery dataset, as well as the sample's membership degree to each cluster center. This reflects the overall compactness of the clustering results, with the core idea of ​​minimizing intra-cluster distances to ensure samples within the same category are clustered as closely as possible. The supervised term is constructed for labeled samples in the battery dataset carrying battery classification labels. It is calculated based on the difference between the current membership degree of these labeled samples and their corresponding membership reference values, forcing the model to follow the constraints of known label information during the iteration process. The introduction of the supervised term allows a small amount of labeled data to play a guiding role in the clustering process, helping the model overcome the directional bias that may arise from unsupervised clustering. The objective function is specifically expressed as follows:

[0067] in, U For membership function, u ij Represents the first [item] in the battery dataset S. j Sample s j With the i Cluster centers C i Membership degree between them d ij =|s j -C i | This indicates the battery dataset. S The Middle j Sample s j With the i Cluster centers C i The Euclidean distance between them. m ∈[1,+) is the weighting exponent, used to control the fuzziness of clustering. m The larger the value, the higher the degree of ambiguity. α This represents the balance factor, used to balance the supervised and unsupervised information in the formula. α The value of is directly proportional to the ratio of the total number of samples q to the number of labeled samples qL. To label the membership reference value of the sample, b j This is a boolean binary vector, whose values ​​can be used to indicate whether it is labeled information, specifically as follows:

[0068] By setting a balance factor in the objective function, the unsupervised and supervised information are optimized in the clustering process. The value of the balance factor is proportional to the ratio of the total number of samples to the number of labeled samples, ensuring that the supervised information can still play an effective guiding role when there are few sample labels.

[0069] Under constraints The optimal solution to the objective function is obtained using the Lagrange multiplier method, where the optimal solution is the minimum value. (Cluster centers) C i and membership degree u ij The iterative expression is shown in the following formula: ; ; in, Indicates the first j Sample s j With the k Cluster centers C k The Euclidean distance between them.

[0070] Using the minimization of the objective function as the iterative objective, the membership calculation and cluster center update steps described above are repeated. After each iteration, the objective function value for the current round is calculated and compared with the result of the previous iteration. The iterative calculation process continues until the objective function corresponding to the semi-supervised fuzzy C-means clustering model converges, i.e., the change in the objective function between two consecutive iterations is less than a preset convergence threshold, indicating that the model has found the optimal solution, or the number of iterations has reached the preset maximum number of iterations. The iterative process terminates as soon as either of these conditions is met.

[0071] After the iteration terminates, the model outputs the corresponding clustering results, presented as a membership matrix that records the membership degree of each retired battery cell to each cluster center. During the classification and screening phase, based on the distribution of each battery cell in the membership matrix, the cluster center with the highest membership degree is determined, and the battery classification label corresponding to that cluster center is assigned to the retired battery cell. Based on the battery classification label, retired battery cells can be quickly classified and screened, thereby achieving accurate matching for different application scenarios.

[0072] In one implementation, 76 retired batteries and 4 lithium-ion batteries undergoing full-life cycle testing were used as the research subjects, and the above-mentioned retired battery screening process was performed.

[0073] S1: Perform standardized charging tests on retired lithium-ion battery cells (e.g., NCM lithium batteries with a nominal capacity of 2.5Ah). Charge the battery at a constant current of 0.5C (i.e., 1.25 A) under a constant temperature environment of 25 ℃ until the voltage reaches the cutoff voltage (e.g., 4.2 V). Simultaneously, collect the battery's terminal voltage V and cumulative charging capacity Q at a frequency of 1 Hz to form the original (V, Q) data sequence.

[0074] A high-fidelity IC curve is constructed using a fitting method based on a three-layer feedforward neural network (FNN). The FNN's input layer is voltage V, its output layer is capacity Q, and its hidden layer contains 20 neurons. The sigmoid function is used as the activation function. The network is trained using collected (V, Q) data, with the goal of minimizing the mean square error between the predicted and actual capacities. The trained network accurately describes the nonlinear relationship between V and Q. Numerical differentiation (dQ / dV) is performed on the trained neural network model in the voltage domain to directly obtain a smoothed and denoised IC curve. Compared to the traditional difference method, this method effectively suppresses data noise and accurately highlights voltage plateau characteristics. The curve obtained by the NNF method is smoother, with more obvious peak features, which is beneficial for the stable extraction of subsequent features.

[0075] S2: Peak detection is performed on the IC curve obtained in S1 to extract three key features: the peak point of the IC curve (ICMAX), discharge energy, and temperature, etc., to construct a multi-dimensional health feature vector. A backpropagation neural network is established with the multi-dimensional health feature vector as input and the SOH value as output. The initial network structure is set empirically. To improve the model's accuracy and generalization ability, the firefly algorithm is used to globally optimize the initial weights, biases, and number of hidden layer neurons of the BP neural network. The root mean square error of the model on the validation set is minimized. Each firefly represents a set of network parameter configurations. The brightness of the firefly is related to the objective function value (the reciprocal of RMSE). By simulating the behavior of fireflies attracting each other and moving towards brighter individuals, the parameters are iteratively updated to find the optimal combination of network parameters. The neural network lithium-ion battery SOH estimation model based on the firefly optimization algorithm is trained using 63 randomly selected battery cells. The model's accuracy under different operating conditions shows a root mean square error of less than 4%, which can effectively guide the initial parameter confirmation of the subsequent clustering algorithm.

[0076] S3: Based on the target tiered utilization scenario, for example, batteries used in energy storage systems require SOH > 80%, set the SOH threshold to 80%. Battery cells with an estimated SOH > 85% from S2 are labeled "Class A" (high capacity group). Battery cells with SOH < 75% are labeled "Class C" (low capacity group). For battery cells whose estimated SOH falls within the ambiguous range of 75% to 85%, their true category is difficult to determine accurately using a single threshold; therefore, they are not labeled in this step and are treated as unlabeled data.

[0077] S4: Construct a battery dataset based on the SOH estimates and battery distribution labels for all batteries (including labeled and unlabeled). Use a Gaussian kernel function to estimate the probability density of the battery dataset's distribution. For any point x in the two-dimensional feature space, its probability density estimate is calculated from the combined contribution of all data points. Calculate the probability density of all sample points in the battery dataset and identify the points with the highest probability density. For example, local density peaks are located near (SOH=88%, V_peak=3.45V) and (SOH=72%, V_peak=3.38V), respectively. These two local density peaks are directly determined as the two initial cluster centers v1 and v2 for subsequent fuzzy C-means clustering. This method ensures that the initial centers are located at the densest natural cluster centers and that the calculation results are consistent each time, exhibiting high repeatability.

[0078] S5: The initial smart centers (v1, v2) obtained in S4, the data of tagged batteries and their labels from S3, and the feature data of all untagged batteries are input into the SSFCM model. SSFCM adds a label constraint term to the standard FCM objective function, forcing tagged data to belong to their specified categories as much as possible while minimizing intra-class distance. The membership matrix and cluster centers are iteratively updated using the Lagrange multiplier method until the objective function converges or the maximum number of iterations is reached. Based on the converged membership matrix, a final category is assigned to each battery cell.

[0079] The sorting method proposed in this application has advantages in computational complexity compared with conventional methods, mainly in the optimization of the initialization strategy and the essential improvement of the algorithm's convergence efficiency. Its core is not to significantly increase the amount of computation per step, but to greatly reduce the total number of iterations and computational resources required to achieve the goal by introducing intelligent preprocessing and guidance mechanisms, thereby achieving higher overall efficiency in practical applications.

[0080] The above are embodiments of the methods proposed in this application. Based on the same idea, some embodiments of this application also provide devices and non-volatile computer storage media corresponding to the above methods.

[0081] Figure 2This is a schematic diagram of a decommissioned battery screening device based on fuzzy clustering analysis, provided as an embodiment of this application. Figure 2 As shown, it includes: At least one processor; and, At least one processor-communication-connected memory; wherein, The memory stores instructions that can be executed by at least one processor to enable the at least one processor to perform a fuzzy clustering analysis-based method for screening decommissioned batteries as described in any of the preceding claims.

[0082] This application provides a non-volatile computer storage medium storing computer-executable instructions, wherein the computer-executable instructions are configured as follows: A method for screening decommissioned batteries based on fuzzy clustering analysis, as described in any of the preceding items.

[0083] The various embodiments in this application are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the device and medium embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the description of the method embodiments.

[0084] The devices and media provided in this application are one-to-one with the methods. Therefore, the devices and media also have similar beneficial technical effects as their corresponding methods. Since the beneficial technical effects of the methods have been described in detail above, the beneficial technical effects of the devices and media will not be repeated here.

[0085] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0086] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0087] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0088] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0089] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0090] Memory may include non-persistent storage in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0091] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

[0092] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0093] The above description is merely an embodiment of this application and is not intended to limit the scope of this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A method for screening decommissioned batteries based on fuzzy clustering analysis, characterized in that, The method includes: Obtain the charge and discharge data of the retired battery cells to be screened, train the feedforward neural network based on the charge and discharge data, and fit the incremental capacity curve corresponding to the retired battery cells based on the trained feedforward neural network. Based on the incremental capacity curve, construct the multidimensional health characteristics corresponding to the retired battery cell; The multidimensional health features are input into a preset health status estimation model. Based on the distribution relationship between the health status values ​​output by the health status estimation model and the preset health status threshold, a battery dataset carrying battery classification labels is generated. The battery dataset includes the multidimensional health features of each retired battery cell. Based on the probability density of each sample point in the battery dataset, the initial cluster centers are determined; The initial cluster centers and the battery dataset are input into a preset semi-supervised fuzzy C-means clustering model for iterative calculation, so as to classify and screen the retired battery cells based on the output clustering results.

2. The method for screening decommissioned batteries based on fuzzy clustering analysis according to claim 1, characterized in that, The feedforward neural network is trained based on the charge and discharge data, and the incremental capacity curve corresponding to the retired battery cell is fitted based on the trained feedforward neural network, specifically including: The feedforward neural network is trained based on the charging and discharging data; wherein the input layer of the feedforward neural network is the battery voltage, the output layer is the battery capacity, and the hidden layer contains two neurons. Based on the neuron parameters of the feedforward neural network obtained through training, the relationship function between battery voltage and battery capacity is determined; The derivative of the relationship function is calculated, and the incremental capacity value obtained from the derivative and the battery voltage are used to fit the incremental capacity curve corresponding to the retired battery cell.

3. The method for screening decommissioned batteries based on fuzzy clustering analysis according to claim 1, characterized in that, Based on the distribution relationship between the health status values ​​output by the health status estimation model and the preset health status threshold, a battery dataset carrying battery classification labels is generated, specifically including: Based on the target application scenario of the retired battery cell, a preset health status threshold is determined, and a first boundary value and a second boundary value are respectively formed by the upper and lower distribution threshold boundaries of the health status threshold. For each retired battery cell, if the health status value is between the first boundary value and the second boundary value, the retired battery cell is not labeled. Otherwise, based on the relationship between the health status value and the first and second boundary values, a battery classification label corresponding to the retired battery cell is generated. A battery dataset is constructed based on the health status values ​​of all retired battery cells and the battery classification labels.

4. The method for screening decommissioned batteries based on fuzzy clustering analysis according to claim 1, characterized in that, The initial cluster centers and the battery dataset are input into a pre-defined semi-supervised fuzzy C-means clustering model for iterative calculation. Based on the output clustering results, the retired battery cells are classified and screened. Specifically, this includes: The initial cluster centers and the battery dataset are input into a preset semi-supervised fuzzy C-means clustering model; The initial cluster center is used as the cluster center for the current iteration. A membership matrix is ​​generated based on the distance between each sample point in the battery dataset and the cluster center. The cluster centers are updated according to the membership matrix, and the updated cluster centers are used as the cluster centers for the next iteration. Repeat the above membership calculation and cluster center update steps until the objective function corresponding to the semi-supervised fuzzy C-means clustering model converges or the number of iterations reaches the preset maximum number of iterations, and output the corresponding clustering results; wherein, the clustering results are the membership of each retired battery cell to each cluster center; The retired battery cells are classified and screened based on the battery classification label corresponding to the cluster center with the highest membership degree.

5. The method for screening decommissioned batteries based on fuzzy clustering analysis according to claim 4, characterized in that, Repeat the above membership calculation and cluster center update steps until the objective function of the semi-supervised fuzzy C-means clustering model converges or the number of iterations reaches the preset maximum number of iterations, specifically including: Based on the Euclidean distance between each sample and each cluster center in the battery dataset, and the membership degree of each sample to each cluster center, an unsupervised term is constructed; For the labeled samples carrying battery classification labels in the battery dataset, a supervision term is constructed based on the degree of difference between the current membership degree of the labeled sample and its corresponding membership degree reference value. The unsupervised term and the supervised term are weighted and summed to obtain the objective function corresponding to the semi-supervised fuzzy C-means clustering model. Minimizing the objective function is used as the iteration objective. The above membership calculation and cluster center update steps are repeated until the objective function corresponding to the semi-supervised fuzzy C-means clustering model converges or the number of iterations reaches the preset maximum number of iterations.

6. The method for screening decommissioned batteries based on fuzzy clustering analysis according to claim 1, characterized in that, Based on the probability density of each sample point in the battery dataset, the initial cluster centers are determined, specifically including: Using a preset Gaussian kernel function, the probability density corresponding to each sample point in the battery dataset is calculated, and local density peak points in the sample points whose probability density is greater than that of all adjacent sample points are identified. Based on the preset number of clusters, select several local density peaks with the highest probability density from the local density peaks as the initial cluster centers.

7. The method for screening decommissioned batteries based on fuzzy clustering analysis according to claim 1, characterized in that, Before inputting the multidimensional health features into the preset health status estimation model, the method further includes: Initialize a firefly population; wherein the firefly population consists of several individuals, each of which corresponds to a set of initial connection weights and offsets of a backpropagation neural network; Based on the relative brightness relationship between the individuals, the firefly population is iterated so that individuals with lower brightness move to individuals with higher brightness, until the preset maximum number of iterations is reached. The connection weights and offsets of the brightest individual in the iterated firefly population are used as the optimal initial parameters, and the backpropagation neural network is initialized based on the optimal initial parameters. The backpropagation neural network is trained using a pre-built training sample set to obtain the health status estimation model.

8. The method for screening decommissioned batteries based on fuzzy clustering analysis according to claim 2, characterized in that, The relational function is expressed as follows: ； in, v' represents the normalized battery capacity, and v′ represents the normalized battery capacitance. This represents the weights from the input layer to the hidden layer. It is the bias term of the i-th neuron. This represents the weights from the hidden layer to the output layer. It is the bias term of the output layer.

9. A decommissioned battery screening device based on fuzzy clustering analysis, characterized in that, The device includes: At least one processor; And, a memory communicatively connected to the at least one processor; The memory stores instructions that can be executed by the at least one processor, which are executed by the at least one processor to enable the at least one processor to perform a decommissioned battery screening method based on fuzzy clustering analysis as described in any one of claims 1-8.

10. A non-volatile computer storage medium storing computer-executable instructions, characterized in that, The computer-executable instructions are set as follows: A method for screening decommissioned batteries based on fuzzy clustering analysis as described in any one of claims 1-8.

Images