A Battery Health State Estimation Method Based on Local Fragment Data Extraction

Abstract

This invention provides a battery health state estimation method based on local segment data extraction, belonging to the field of lithium-ion battery technology, including: step S1, constructing an estimation model; step S2, in... m In each charge-discharge cycle, the intervals for feature extraction are determined on the battery charging curve corresponding to each charge-discharge cycle, with the same interval Δ. V Select n Step S3: Obtain the optimal interval Δ for each voltage segment and acquire the input feature matrix. V p And obtain the optimal input feature matrix; Step S4, repeat steps S2 and S3 a set number of times, each repetition at a different interval Δ V Step S5 involves selecting a voltage segment and obtaining the optimal input feature matrix corresponding to this repetition; then, inputting each optimal input feature matrix into the estimation model to determine the final optimal interval, and repeating step S2 with the final optimal interval to obtain the final input feature matrix. This invention balances the complexity and accuracy of the estimation model.

Description

Technical Field

[0001] This invention belongs to the field of lithium-ion battery technology, specifically relating to a battery health state estimation method based on the extraction of local segment data from the battery. Background Technology

[0002] Lithium-ion batteries (LIBs) possess advantages such as high energy power density, low self-discharge rate, and long lifespan, and are widely used in consumer electronics, energy storage systems, and electric vehicles. To improve the reliability of battery-powered systems, a battery management system (BMS) is crucial for real-time estimation and monitoring of battery state. Throughout the LIB's lifespan, due to the interaction of complex physical and chemical reactions, battery capacity gradually degrades with use and aging; a battery is considered to have reached the end of its life when its capacity decays to 80%. As a key component of the BMS, accurate battery capacity estimation is a critical and challenging problem.

[0003] In recent years, research on battery capacity estimation has mainly fallen into two categories: physical model-based methods and data-driven methods. Physical model-based methods link the battery's electrochemical model (EM) and equivalent circuit model (ECM) with its capacity, using adaptive techniques known in control theory, such as Kalman filtering (KF) or particle filtering (PF), to predict battery capacity degradation. However, to achieve accurate capacity estimation, physical model-based methods need to establish a suitable and generalized aging model, which is not always easy.

[0004] Data-driven methods refer to learning the mapping relationship between features and State of Health (SOH) through neural networks and support vector machines. This method has wide applicability, and with the rapid development of artificial intelligence, it has become a research hotspot for SOH estimation. However, feature selection is crucial to this method; therefore, during extraction, we need to select features that are easy to obtain, simple to process, and highly relevant to extracting battery health characteristics. Since the battery discharge process is dynamic, extracting stable features presents a challenge. For example, discharge data varies under different discharge conditions (such as operating temperature, discharge current, and cutoff voltage), making it difficult to obtain consistent and reliable discharge data. To improve the accuracy of SOH estimation, more complex models may be needed for feature extraction and SOH estimation; however, model complexity introduces other problems, and how to balance complexity and accuracy requires further research. Summary of the Invention

[0005] The purpose of this invention is to propose a battery health state estimation method based on the extraction of local battery fragment data, which can balance the complexity and accuracy of the estimation model.

[0006] This invention is achieved through the following technical solution:

[0007] A battery health state estimation method based on local battery fragment data extraction includes the following steps:

[0008] Step S1: Construct an estimation model. The estimation model includes an attention mapping layer that generates feature nodes, an enhancement node layer that enhances feature nodes into enhancement nodes, a dropout layer connected to the enhancement node layer, and an output layer connected to the attention mapping layer and the dropout layer. The output weights of the output layer are calculated by elastic network regression. The attention mapping layer includes a feature attention module, a temporal attention module, and a fusion module that fuses feature attention and temporal attention.

[0009] Step S2, in m During each charge-discharge cycle, based on statistical data, the intervals for feature extraction are determined on the battery charging curves corresponding to each charge-discharge cycle. , at the same interval △ V Select n voltage segment ,in accordance with n The battery charging curve is extracted by segmenting each voltage range to obtain... n Each battery charging curve segment, for m The charging curve segments of each battery obtained from each charge-discharge cycle are subjected to curve fitting. The fitted parameters are used as input features to form an input feature matrix. , ;

[0010] Step S3: Input the input feature matrix into the estimation model, and use the multi-objective particle swarm optimization algorithm to obtain the optimal margin Δ with the objectives of computational cost and computational accuracy. V p and with the optimal interval △ V p Repeat step S2 to obtain the optimal input feature matrix;

[0011] Step S4: Repeat steps S2 and S3 a set number of times, with each repetition occurring at a different interval △. V Select the voltage segment and obtain the optimal input feature matrix corresponding to this repetition;

[0012] Step S5: Input each optimal input feature matrix into the estimation model, and evaluate the estimation results through MAPE and RMSE to determine the final optimal interval. Repeat step S2 with the final optimal interval to obtain the final input feature matrix, and input the final input feature matrix into the estimation model for estimation.

[0013] Furthermore, in step S2, in each charge-discharge cycle, based on statistical data, the range of the extracted feature interval is [3.8V, 4.2V].

[0014] Furthermore, in step S2, a quadratic equation is used in each charge-discharge cycle. Each of the battery charging curve segments is fitted, and the fitted parameters include quadratic parameters, linear parameters, and constant parameters. The correlation between the quadratic parameters, linear parameters, and constant parameters and the average battery capacity under the corresponding battery charging curve segment is compared, and the parameters with the highest correlation are used to form the input feature matrix.

[0015] Furthermore, in step S2, the process of comparing the correlation between the quadratic parameter, the linear parameter, and the constant parameter and the average battery capacity under the corresponding battery charging curve segment includes:

[0016] Step S21, the m In the next charge-discharge cycle, for the first n The parameters obtained by fitting the battery charging curve segments corresponding to each voltage segment are: , , , respectively m The parameters obtained from the second charge-discharge cycle are combined into three matrices. , , ;

[0017] Step S22, according to the formula Calculate the first n The correlation parameter sequence of each voltage segment, and according to the formula , , Calculate the average battery capacity for the quadratic, linear, and constant parameters and the corresponding battery charging curve segments, and take... , , The matrix corresponding to the largest number in the matrix is ​​used as the input feature matrix. ,in, Indicates the first n The time series of quadratic parameters under the battery charging curve segment Indicates the first n A time series of parameters under a single battery charging curve segment Indicates the first n Time series of constant parameters under each battery charging curve segment Indicates the first n A sequence of average battery capacity under each battery charging curve segment E () indicates taking the average. D () indicates taking the variance.

[0018] Furthermore, in step S1, the loss function calculated by the elastic network regression is expressed as: ,in, x i Indicates the first i Characteristics of the estimated observations y i Indicates the first i One actual observation value, This represents the Lasso regression regularization coefficient. This represents the ridge regression regularization coefficient. p Indicates the number of actual observations. Represents the regression coefficient. This represents the Lasso regression regularization term. This represents the ridge regression regularization term.

[0019] Furthermore, in step S1, the feature attention weight matrix in the feature attention module is represented as follows: ,in, Indicates the first i The first charge-discharge cycle k The weights of the input features fitted to the battery charging curve segments corresponding to each voltage segment. i =1,2,…, m , k =1,2,…, n .

[0020] Furthermore, in step S1, the time attention weight matrix in the time attention module is represented as follows: ,in, .

[0021] Furthermore, in step S1, the fusion matrix in the fusion module is represented as follows: .

[0022] The present invention has the following beneficial effects:

[0023] This invention first constructs an estimation model, which includes an attention mapping layer for generating feature nodes, an enhancement node layer for enhancing feature nodes, a dropout layer connected to the enhancement node layer, and an output layer connected to the attention mapping layer and the dropout layer. The output weights of the output layer are calculated by elastic network regression. The attention mapping layer includes a feature attention module, a temporal attention module, and a fusion module for fusing feature attention and temporal attention. Then... m During each charge-discharge cycle, the intervals for feature extraction are determined on the corresponding battery charging curve, with equal intervals Δ. V Select n Each voltage segment, and according to n The battery charging curve is extracted by segmenting each voltage range to obtain... n For each battery charging curve segment, curve fitting is performed. The fitted parameters are used as input features to form an input feature matrix. Then, the input feature matrix is ​​input into the estimation model, and a multi-objective particle swarm optimization algorithm is used to obtain the optimal margin Δ with the objectives of computational cost and accuracy. V p and with the optimal interval △ V p To obtain the optimal input feature matrix, and to further improve accuracy, multiple intervals Δ are set. V For each interval △ V The above process is repeated for each case to obtain the corresponding optimal input feature matrix. Each optimal input feature matrix is ​​then input into the estimation model, and the estimation model is evaluated using MAPE and RMSE to determine the final optimal interval. In actual estimation, the final optimal interval is used to obtain the final input feature matrix, and the final input feature matrix is ​​then input into the estimation model for estimation. In this process, features are extracted based on the battery charging curve. Compared to the discharging process, the charging process is relatively standardized, allowing for the collection of more stable and representative features. Furthermore, selecting intervals on the battery charging curve for input feature extraction better aligns with the daily usage patterns of new energy vehicle users. A multi-objective particle swarm optimization algorithm is used to determine the optimal interval based on computational cost and accuracy. This ensures computational accuracy while avoiding a significant increase in computational load due to excessive voltage segmentation, and also reduces the risk of overfitting in the estimation model. The elastic network regression calculation in the estimation model stabilizes the estimation of the battery charging curve segment fitting coefficients, reduces the variance of the coefficients, and is more suitable for estimating battery health status. The addition of a dropout layer further reduces the risk of overall overfitting in the estimation model. The attention mapping layer in the estimation model calculates attention weights for the time series and feature sequences separately, enabling the estimation model to better focus on key information rather than uniformly processing all inputs, thereby improving the performance of the estimation model and shortening training time. In summary, this invention effectively balances the complexity and accuracy of the estimation model. Attached Figure Description

[0024] The present invention will now be described in further detail with reference to the accompanying drawings.

[0025] Figure 1 This is a flowchart of the present invention.

[0026] Figure 2 Simulation results comparing the SoH (Solar Oxygen Optimization) of a univariate K-degree equation with different choices of K (K=1,2,3).

[0027] Figure 3 The simulation graphs show the SOH comparison between fitting linear and quadratic equations in one variable.

[0028] Figure 4 This is a framework diagram of the estimation model of the present invention.

[0029] Figure 5 The image shows a simulation diagram of battery health state estimation using the present invention and existing SOH estimation methods.

[0030] Figure 6 This is a simulation diagram showing the battery health state estimation error between the present invention and existing SOH estimation methods. Detailed Implementation

[0031] The dataset used in this embodiment comes from the NASA Ames Prediction Center of Excellence. The three lithium-ion batteries (labeled 5, 6, and 7) are second-generation 18650 batteries manufactured by Idaho National Laboratory. All batteries were repeatedly operated at room temperature (24°C) in three different operating modes: constant current and constant voltage (CC-CV) charging mode, CC discharging mode, and impedance measurement mode. The CC charging current, upper limit voltage, and CV charging voltage were 1.5A, 4.2V, and 4.2V, respectively, with CV ending when the current dropped to 50mA. CC discharging and CV charging were continuously repeated. The experiment was terminated when the LIB reached its end-of-life (EOL) standard.

[0032] In this embodiment, SOH is defined as capacity decay, i.e. ,in, C i This indicates the maximum capacity for the current period. C O This indicates the initial capacity.

[0033] like Figure 1 As shown, the battery health state estimation method based on local battery fragment data extraction includes the following steps:

[0034] Step S1: Construct an estimation model. The estimation model includes an attention mapping layer that generates feature nodes, an enhancement node layer that enhances feature nodes into enhancement nodes, a dropout layer connected to the enhancement node layer, and an output layer connected to the attention mapping layer and the dropout layer. The output weights of the output layer are calculated by elastic network regression. The attention mapping layer includes a feature attention module, a temporal attention module, and a fusion module that fuses feature attention and temporal attention.

[0035] Step S2, in m During each charge-discharge cycle, based on statistical data, the intervals for feature extraction are determined on the battery charging curves corresponding to each charge-discharge cycle. , at the same interval △ V Select n voltage segment ,in accordance with n The battery charging curve is extracted by segmenting each voltage range to obtain... n Each battery charging curve segment, for m The charging curve segments of each battery obtained from each charge-discharge cycle are subjected to curve fitting. The fitted parameters are used as input features to form an input feature matrix. , ;

[0036] As batteries age, the battery charging curve shows a clear upward trend, which can well explain the battery aging process. Statistical records of 11,000 user charging behaviors in 100 electric vehicles show that most users chose to start charging at a voltage of 3.8V and end charging at a voltage of 4.18V. Therefore, in this embodiment, the voltage range for extracting features is determined to be [3.8V, 4.2V].

[0037] like Figure 2 The figure shows a comparison of SOH simulations when K (K=1,2,3) is chosen for different degrees of univariate K equations. Figure 2 As can be seen, linear and quadratic functions can still provide accurate predictions when fitting the data. However, when the function exponent reaches the third degree, the overall predicted value fluctuates drastically and the result is very inaccurate, proving that the robustness is poor and it is not suitable for predicting battery health status. Since other functions or functions with better exponents require more time to fit, which deviates from the original intention of our method to achieve fast real-time measurement in vehicles, we will not consider other fitting methods. Next, we will further discuss linear and quadratic fitting functions based on the above to finally determine the fitting function.

[0038] like Figure 3The figure shows a comparison of SOH simulation results obtained by fitting linear equations and quadratic equations. It can be seen that in the latter half of the curve, the results of fitting by linear equations and quadratic equations are comparable, and linear equations have an advantage in fitting speed. However, in the early stage of battery cycling, the fitting by linear equations cannot accurately estimate the SOH of the battery and there are still large errors and fluctuations. Therefore, after weighing the pros and cons, quadratic equations were finally chosen as the fitting function to extract features from the battery health curve and make an estimate.

[0039] In this embodiment, a quadratic equation is used in each charge-discharge cycle. The battery charging curve segments are fitted, and the fitted parameters include quadratic parameters. a Primary parameters b and constant parameters d The correlation between the quadratic parameters, linear parameters, and constant parameters and the average battery capacity under the corresponding battery charging curve segment is compared, and the parameters with the highest correlation are used to form the input feature matrix.

[0040] The process of comparing the correlation between the quadratic parameter, the linear parameter, and the constant parameter and the average battery capacity under the corresponding battery charging curve segment includes:

[0041] Step S21, the m In the next charge-discharge cycle, for the first n The parameters obtained by fitting the battery charging curve segments corresponding to each voltage segment are: , , , respectively m The parameters obtained from the second charge-discharge cycle are combined into three matrices. , , ;

[0042] Step S22, according to the formula Calculate the first n The correlation parameter sequence of each voltage segment, and according to the formula , , Calculate the average battery capacity for the quadratic, linear, and constant parameters and the corresponding battery charging curve segments, and take... , , The largest number The corresponding matrix is ​​used as the input feature matrix. ,in, Indicates the first n The time series of quadratic parameters under the battery charging curve segment Indicates the first nA time series of parameters under a single battery charging curve segment Indicates the first n Time series of constant parameters under each battery charging curve segment Indicates the first n A sequence of average battery capacity under each battery charging curve segment E () indicates taking the average. D () indicates taking the variance.

[0043] Step S3: Input the input feature matrix into the estimation model, and use the multi-objective particle swarm optimization algorithm to obtain the optimal margin Δ with the objectives of computational cost and computational accuracy. V p and with the optimal interval △ V p Repeat step S2 to obtain the optimal input feature matrix;

[0044] Obtain the optimal interval Δ using a multi-objective particle swarm optimization algorithm V p Specifically, it includes the following steps:

[0045] Step S11: Initialize the position of the particles x n0 ,speed v n0 Global optimal solution set G n0 And the individual optimal solution for each particle. P n0 ;

[0046] In step S12, during the k-th iteration, according to the formula... Update speed v nk According to the formula Update particle position x nk ,in, w The inertial weight represents the degree to which a particle's current velocity affects its next velocity. r 1 represents the individual learning factor, which is a random number, typically varying between [0,1], and is used to control how closely a particle moves towards its historical best position. r 2 represents the social learning factor, which is a random number, typically varying between [0,1], and is used to control how much a particle moves closer to the global optimal position in the swarm. e Indicates random perturbation;

[0047] Step S13: After setting the number of iterations G, according to the formula... For the obtained particle position x nGConstraints are applied to ensure that the particle positions are within a reasonable range, according to the formula. y n = f ( x n Calculate the fitness value to evaluate the quality of the current particle position. For non-dominated solution sets, add the new solution and remove dominated solutions. If the new solution is better than the current best solution, then use the new solution as the best solution. enforceConstrans () represents the constraint function. f ( x n )satisfy min (△ V p )=[ f 1( x n ), f 2( x n )], min (△ V p ) indicates that when the condition is met f 1( x n )and f 2( x n Given two objective functions, Δ V p Take the minimum value. x n The optimal interval △ represents V p , f 1( x n )express x n The computational cost when considering the optimal interval. f 2( x n )express x n The Pearson correlation coefficient in the input features obtained at the specified intervals, where G is an integer. k =1,2,…,G. The optimal interval Δ determined in this embodiment. V p It is 0.05V.

[0048] Step S4: Repeat steps S2 and S3 a set number of times, with each repetition occurring at a different interval △. V Select the voltage segment and obtain the optimal input feature matrix corresponding to this repetition;

[0049] In this embodiment, the number of times is set to 10, with an interval of Δ between the 10 times. VThe values ​​are 0.01, 0.02, 0.03, 0.04, 0.05, 0.06, 0.07, 0.08, 0.09, and 0.1, respectively.

[0050] Step S5: Input each optimal input feature matrix into the estimation model, and evaluate the estimation results through MAPE and RMSE to determine the final optimal interval. Repeat step S2 with the final optimal interval to obtain the final input feature matrix, and input the final input feature matrix into the estimation model for estimation.

[0051] In the estimation model, the loss function calculated by elastic network regression is expressed as: ,in, x i Indicates the first i Characteristics of the estimated observations y i Indicates the first i One actual observation value, This represents the Lasso regression regularization coefficient. This represents the ridge regression regularization coefficient. p Indicates the number of actual observations. Represents the regression coefficient. This represents the Lasso regression regularization term. This represents the ridge regression regularization term. Elastic network regression uses both Lasso regression regularization coefficients and ridge regression regularization coefficients in its calculations, thus achieving a trade-off between feature selection and parameter reduction. It can precisely reduce some coefficients to zero, automatically identifying features that are unimportant to the target variable, thereby simplifying the estimation model and improving interpretability. Furthermore, elastic network regression is suitable for handling multicollinearity problems, thus stabilizing coefficient estimates and reducing coefficient variance.

[0052] In the feature attention module, the feature attention weight matrix is ​​represented as follows: ,in, Indicates the first i The first charge-discharge cycle k The weights of the input features fitted to the battery charging curve segments corresponding to each voltage segment. i =1,2,…, m , k =1,2,…, n .

[0053] In the temporal attention module, the temporal attention weight matrix is ​​represented as follows: ,in, .

[0054] In the fusion module, the feature attention weight matrix and the temporal attention matrix obtained above are multiplied element-wise to obtain a matrix of the same size as the input two-dimensional matrix, which is the fusion matrix. C , represented as .

[0055] Feature enhancement layers can be implemented using existing techniques, such as feature enhancement layers using wide learning networks. Dropout layers can also be implemented using existing techniques.

[0056] For multi-feature data like battery data, time-series inputs cannot be effectively processed. Introducing an attention mapping layer enables the estimation model to effectively focus on key information. To avoid overfitting, a dropout layer is added after the boosting nodes in the width learning network. Utilizing the random ignoring characteristic of the dropout layer, the prediction model is prevented from becoming overly dependent on relevant feature values ​​during training. This results in... Figure 4 The framework structure diagram of the estimation model.

[0057] After inputting the optimal input feature matrices into the estimation model, the estimation results are evaluated using MAPE and RMSE. The optimal margin corresponding to the optimal input feature matrix with the best evaluation performance is selected as the final optimal margin. In actual estimation, step S2 is repeated using the final optimal margin to obtain the final input feature matrix. The MAPE formula is as follows: The RMSE formula is: ,in, Indicates the predicted value. Represents the true value. n This indicates the sample size of the test dataset.

[0058] like Figure 5 The image shown is a simulation diagram of battery health state estimation using the present invention (with added attention mechanism and width learning network estimation) and existing SOH estimation methods (convolutional neural network estimation, long and short neural network estimation, and width learning network estimation). Figure 6 The figure shows a simulation diagram of the battery estimation error (Error) of the present invention compared to existing SOH estimation methods. It can be seen that the RMSE of the present invention can be reduced to 0.57%, and the MAPE can be reduced to 0.46%, both lower than other estimation methods, effectively improving the accuracy of the present invention in estimating battery health status.

[0059] The above description is merely a preferred embodiment of the present invention and should not be construed as limiting the scope of the present invention. All equivalent changes and modifications made in accordance with the scope of the patent application and the contents of the specification of the present invention should still fall within the scope of the patent of the present invention.

Claims

1. A method for estimating battery health status based on local segment data extraction, characterized in that: Includes the following steps: Step S1: Construct an estimation model. The estimation model includes an attention mapping layer that generates feature nodes, an enhancement node layer that enhances feature nodes into enhancement nodes, a dropout layer connected to the enhancement node layer, and an output layer connected to the attention mapping layer and the dropout layer. The output weights of the output layer are calculated by elastic network regression. The attention mapping layer includes a feature attention module, a temporal attention module, and a fusion module that fuses feature attention and temporal attention. Step S2, in m During each charge-discharge cycle, based on statistical data, the intervals for feature extraction are determined on the battery charging curves corresponding to each charge-discharge cycle. , at the same interval △ V Select n voltage segment ,in accordance with n The battery charging curve is extracted by segmenting each voltage range to obtain... n Each battery charging curve segment, for m The charging curve segments of each battery obtained from each charge-discharge cycle are subjected to curve fitting. The fitted parameters are used as input features to form an input feature matrix. , ; Step S3: Input the input feature matrix into the estimation model, and use the multi-objective particle swarm optimization algorithm to obtain the optimal margin Δ with the objectives of computational cost and computational accuracy. V p and with the optimal interval △ V p Repeat step S2 to obtain the optimal input feature matrix; Step S4: Repeat steps S2 and S3 a set number of times, with each repetition occurring at a different interval △. V Select the voltage segment and obtain the optimal input feature matrix corresponding to this repetition; Step S5: Input each optimal input feature matrix into the estimation model, and evaluate the estimation results through MAPE and RMSE to determine the final optimal interval. Repeat step S2 with the final optimal interval to obtain the final input feature matrix, and input the final input feature matrix into the estimation model for estimation.

2. The battery health state estimation method based on battery local segment data extraction according to claim 1, characterized in that: In step S2, in each charge-discharge cycle, based on statistical data, the range of the extracted feature interval is [3.8V, 4.2V].

3. The battery health state estimation method based on battery local segment data extraction according to claim 2, characterized in that: In step S2, a quadratic equation is used in each charge-discharge cycle. Each of the battery charging curve segments is fitted, and the fitted parameters include quadratic parameters, linear parameters, and constant parameters. The correlation between the quadratic parameters, linear parameters, and constant parameters and the average battery capacity under the corresponding battery charging curve segment is compared, and the parameters with the highest correlation are used to form the input feature matrix.

4. The battery health state estimation method based on battery local segment data extraction according to claim 3, characterized in that: In step S2, the process of comparing the correlation between the quadratic parameter, the linear parameter, and the constant parameter and the average battery capacity under the corresponding battery charging curve segment includes: Step S21, the m In the next charge-discharge cycle, for the first n The parameters obtained by fitting the battery charging curve segments corresponding to each voltage segment are: , , , respectively m The parameters obtained from the second charge-discharge cycle are combined into three matrices. , , ; Step S22, according to the formula Calculate the first n The correlation parameter sequence of each voltage segment, and according to the formula , , Calculate the average battery capacity for the quadratic, linear, and constant parameters and the corresponding battery charging curve segments, and take... , , The matrix corresponding to the largest number in the matrix is ​​used as the input feature matrix, where, Indicates the first n The time series of quadratic parameters under the battery charging curve segment Indicates the first n A time series of parameters under a single battery charging curve segment Indicates the first n Time series of constant parameters under each battery charging curve segment Indicates the first n A sequence of average battery capacity under each battery charging curve segment E () indicates taking the average. D () indicates taking the variance.

5. The battery health state estimation method based on battery local segment data extraction according to claim 4, characterized in that: In step S1, the loss function calculated by the elastic network regression is expressed as follows: ,in, x i Indicates the first i Characteristics of the estimated observations y i Indicates the first i One actual observation value, This represents the Lasso regression regularization coefficient. This represents the ridge regression regularization coefficient. p Indicates the number of actual observations. Represents the regression coefficient. This represents the Lasso regression regularization term. This represents the ridge regression regularization term.

6. The battery health state estimation method based on battery local segment data extraction according to claim 5, characterized in that: In step S1, the feature attention weight matrix in the feature attention module is represented as follows: ,in, Indicates the first i The first charge-discharge cycle k The weights of the input features fitted to the battery charging curve segments corresponding to each voltage segment. i =1,2,…, m , k =1,2,…, n .

7. The battery health state estimation method based on battery local segment data extraction according to claim 6, characterized in that: In step S1, the time attention weight matrix in the time attention module is represented as follows: ,in, .

8. The battery health state estimation method based on battery local segment data extraction according to claim 7, characterized in that: In step S1, the fusion matrix in the fusion module is represented as follows: .

Images