Rechargeable battery health state estimation method and system

Abstract

The invention provides a rechargeable battery health state estimation method and system, and the method comprises the steps: obtaining the aging data of a reference battery for training, so as to obtain a first data set, and the aging data comprise first feature data and first state data representing the health state; acquiring second feature data of a to-be-detected target battery to obtain a second data set; aligning the first feature data in the first data set with the second feature data in the second data set on statistical characteristic distribution; training a ridge regression model by adopting the aligned first data set to obtain a regression coefficient; and predicting second state data representing the health state of the current target battery based on a plurality of second feature data in the current time window of the target battery and the regression coefficient. According to the estimation method, efficient, stable and high-precision SOH estimation in a cross-battery and few-sample scene is realized through feature distribution alignment and sliding ridge regression dynamic updating.

Description

Technical Field

[0001] The present invention relates to the technical field of battery health detection, and in particular to a method and system for estimating the health status of a rechargeable battery. Background Art

[0002] In the field of lithium-ion battery state of health (SOH) estimation, existing technologies mainly rely on data-driven methods to build prediction models by analyzing the mapping relationship between battery aging characteristics (such as voltage and capacity increment curves) and SOH. However, there are significant defects in practical applications, as follows.

[0003] Insufficient cross-battery migration capability: Different battery batches, usage conditions or material systems lead to differences in the distribution of aging characteristics (such as constant current charging cutoff capacity, IC curve peak offset). Traditional methods use standardization (Z-score) or linear scaling to align feature scales, but ignore high-order statistical characteristics (such as covariance structure, kernel space distribution differences), resulting in a sharp increase in model prediction errors in cross-domain scenarios. For example, the feature distribution of the source domain battery at room temperature cannot be aligned with the high-temperature aging data of the target domain.

[0004] Model failure under small sample data: Early data of target domain batteries is scarce (e.g., only the first 10% of cycle data). Existing transfer learning technologies (e.g., deep domain adaptation networks) require at least 30% of target domain data to train the adaptation layer, which is difficult to meet the needs of online real-time prediction. If the source domain model is used directly to predict the target domain data, the mean absolute error (MAE) is too large due to distribution differences, which cannot meet the requirements of actual applications.

[0005] Dynamic aging characteristics modeling defects: The correlation between battery aging characteristics and SOH changes dynamically with the number of cycles, while traditional static models (such as global ridge regression and support vector machines) cannot capture the timing dynamics. Summary of the invention

[0006] The purpose of the present invention is to solve the above technical problems and to provide a rechargeable battery health status estimation method and system that can adaptively predict scenarios with a small number of samples across battery batches and under different usage conditions.

[0007] In order to achieve the above object, the present invention provides a method for estimating the health status of a rechargeable battery, which comprises:

[0008] Acquire aging data of a reference battery for training during several charge and discharge cycles to obtain a first data set, wherein the aging data includes first characteristic data strongly related to a health state of the battery and first state data representing the health state;

[0009] Acquire second characteristic data strongly related to the health status of the battery during several charging cycles of the target battery to be tested to obtain a second data set;

[0010] aligning the first feature data in the first data set and the second feature data in the second data set in terms of statistical characteristic distribution;

[0011] Using the aligned first data set to train a ridge regression model to obtain regression coefficients;

[0012] Predicting second status data representing the current health status of the target battery based on the plurality of second feature data and the regression coefficient within the current time window of the target battery;

[0013] The ridge regression model is retrained using a plurality of the second feature data and the second state data in the current time window as aging data for training, so as to update the regression coefficient.

[0014] Preferably, the first characteristic data is the same as or different from the second characteristic data.

[0015] Preferably, the first characteristic data includes any one of the constant current charging cut-off power and the IC curve peak value; the second characteristic data includes any one of the constant current charging cut-off power and the IC curve peak value; the IC curve is a curve showing the relationship between the battery's capacity increment and voltage.

[0016] The correlation between the first feature data and the second feature data and the health state of the battery is evaluated according to any one or more of the Pearson correlation coefficient, the Spearman correlation coefficient and the Kendall correlation coefficient, so that the first feature data and the second feature data are strongly correlated with the health state of the battery.

[0017] Preferably, the method for aligning the first data set and the second data set comprises:

[0018] Calculating the distance between the probability distributions of a plurality of the first feature data in the first data set and a plurality of the second feature data sets in the second data set to obtain a difference value;

[0019] Determine whether the difference value reaches a preset minimum value. If not, perform a rotation transformation on the current first feature data in the first data set based on a geometric transformation matrix, and recalculate the difference value. If yes, complete the alignment.

[0020] Preferably, before aligning the first feature data in the first data set and the second feature data in the second data set, the first feature data in the first data set are also normally transformed based on transformation parameters so that a number of the first feature data are close to a normal distribution, and the transformation parameters are fitted into the regression coefficients.

[0021] Preferably, the regression coefficient R is expressed as follows:

[0022] R=(F T F+λkI) -1 Fq

[0023] Wherein, λ is the transformation parameter, F is the vector expression of the first feature data after alignment, and F T is the transpose of the first feature data vector after alignment, I is the unit orthogonal matrix, and k is the ridge parameter.

[0024] The present invention also provides a rechargeable battery health state estimation system, which includes a data processor, wherein the data processor estimates the health state of a target battery based on the rechargeable battery health state estimation method as described above.

[0025] The present invention also provides a rechargeable battery health status estimation system, which includes:

[0026] one or more processors;

[0027] Memory;

[0028] and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the one or more processors, the programs comprising instructions for executing the rechargeable battery health state estimation method as described above.

[0029] The present invention also provides a computer-readable storage medium, which includes a computer program, and the computer program can be executed by a processor to implement the rechargeable battery health state estimation method as described above.

[0030] Compared with the prior art, the rechargeable battery health state estimation method provided by the above technical solution of the present invention aligns the feature data of the reference battery (source domain) and the target battery (target domain) in the statistical characteristic distribution, effectively eliminating the feature distribution offset caused by material differences and usage environment of different batteries, so that the trained ridge regression model can directly adapt to the distribution characteristics of the target battery, without the need to re-collect the target battery's full life cycle data, and significantly improves the generalization ability of the model in cross-battery scenarios. Secondly, only a small amount of early data of the target battery is required for feature alignment, and the model trained with massive data in the source domain can be used for migration prediction, breaking through the strong dependence of traditional methods on the amount of target battery data, reducing data acquisition costs, shortening the model deployment cycle, and being suitable for rapid evaluation of the early health status of the target battery. Furthermore, through sliding window retraining, the model parameters are dynamically adjusted with the battery aging process to capture the non-stationary characteristics of capacity decay and avoid the cumulative errors of static models. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 FIG. 4 is a flow chart of a method for estimating the health status of a rechargeable battery in one embodiment of the present invention.

[0032] Figure 2 This is a flow chart of a method for estimating the health status of a rechargeable battery in another embodiment of the present invention. DETAILED DESCRIPTION

[0033] In order to explain the technical content, structural features, achieved objectives and effects of the present invention in detail, the following is a detailed description in conjunction with the implementation methods and the accompanying drawings.

[0034] This embodiment discloses a rechargeable battery health state estimation method, which is used to perform real-time evaluation of the health state data (SOH) of a rechargeable battery (such as a lithium battery).

[0035] SOH is a core indicator for measuring the degree of performance degradation of rechargeable batteries, and is used to quantify the difference between the current state of the battery and its ideal new battery state.

[0036] SOH is usually expressed as a percentage, reflecting the remaining life or performance degradation of the battery. Its core definition includes the following dimensions:

[0037] Capacity decay: The most common definition is the ratio of the current actual capacity of the battery to the initial nominal capacity. For example, if the new battery capacity is 100Ah and the current capacity is 80Ah, then SOH = 80%. Relevant standards stipulate that when SOH is lower than 80%, the battery may need to be replaced or maintained.

[0038] Internal resistance change: As the battery ages, the internal resistance increases, resulting in increased heat generation and decreased power. SOH can be calculated by the rate of change of internal resistance.

[0039] Number of cycles: In some scenarios, SOH is also defined by the ratio of the remaining number of cycles to the nominal number of cycles.

[0040] The evaluation of SOH requires a combination of experimental data and algorithm analysis, which is mainly divided into two categories, namely direct measurement method and spacing measurement method.

[0041] Direct measurement methods include capacitance testing and electrochemical impedance spectroscopy (EIS).

[0042] Capacity test method: Directly measure the available capacity through complete charge and discharge cycles.

[0043] Electrochemical impedance spectroscopy (EIS): By analyzing the battery's response to a sinusoidal voltage, parameters such as internal resistance and polarization impedance are obtained to assess the degree of aging.

[0044] Indirect analysis methods include data-driven methods and feature extraction methods.

[0045] Data-driven approach: Use voltage, current, temperature and other data combined with machine learning to predict SOH.

[0046] Feature extraction: For example, the peak features of the charge and discharge curves are extracted through incremental capacity analysis (ICA) to establish a correlation with SOH.

[0047] The direct measurement method has the advantage of accurate measurement, but it cannot be performed online in real time and cannot meet the user's real-time continuous monitoring of the health status of the rechargeable battery. Therefore, the indirect analysis method has a wider range of applications. The estimation method in this embodiment is the indirect analysis method.

[0048] like Figure 1 , the estimation method includes the following steps:

[0049] S1: Obtain aging data of several charge and discharge cycles (for example, 100 charge and discharge cycles) of a reference battery for training to obtain a first data set belonging to a source domain, wherein the aging data includes first feature data strongly related to the health state of the battery and first state data (ie, SOH) characterizing the health state.

[0050] S2: Acquire second characteristic data (ie, SOH) that is strongly correlated with the health status of the battery during several charging cycles (eg, 10 charging and discharging cycles) of the target battery to be tested, to obtain a second data set.

[0051] S3: Align the first feature data in the first data set and the second feature data in the second data set in terms of statistical characteristic distribution, that is, complete the alignment of source domain data and target domain data.

[0052] S4: Using the aligned first data set to train a ridge regression model to obtain regression coefficients.

[0053] S5: Predicting second status data representing the current health status of the target battery based on the plurality of second feature data and the regression coefficient within the current time window of the target battery.

[0054] In this embodiment, the second characteristic data of the target battery is divided into windows of fixed length (for example, data of the most recent 10 cycles) according to time series. Each time the window slides forward one step (such as adding one cycle data and removing the earliest cycle data), the SOH is evaluated only based on the latest data in the window.

[0055] S6: Retraining the ridge regression model by using a plurality of the second feature data in the current time window and the second state data estimated based on the regression coefficient as aging data for training, so as to update the regression coefficient.

[0056] For the estimation method in this embodiment, first, by aligning the feature data of the reference battery (source domain) and the target battery (target domain) in the statistical characteristic distribution, the feature distribution offset (such as mean shift and covariance structure difference) caused by material differences and usage environment (such as temperature and charging and discharging strategy) of different batteries is effectively eliminated. The trained ridge regression model can directly adapt to the distribution characteristics of the target battery without the need to re-collect the full life cycle data of the target battery, which significantly improves the generalization ability of the model in cross-battery scenarios.

[0057] Secondly, only a small amount of early data of the target battery (such as 10% charge and discharge cycles) is needed for feature alignment, and the model trained with massive data in the source domain can be used for migration prediction. This breaks through the traditional method's strong dependence on the amount of target battery data, reduces data collection costs, shortens the model deployment cycle, and is suitable for rapid assessment of the early health status of the target battery.

[0058] Furthermore, the ridge regression model is used to directly calculate the regression coefficients, avoiding the iterative optimization process of complex models such as neural networks. The time consumption of a single model training is reduced to milliseconds (e.g. <10ms), and with the dynamic update of sliding window technology, real-time SOH estimation can be achieved in embedded devices (such as vehicle-mounted BMS), meeting the low latency requirements of vehicle-mounted systems.

[0059] Furthermore, by retraining with a sliding window, only the latest data in the current time window is retained, and outdated aging patterns in historical data are gradually eliminated. Model parameters are dynamically adjusted as the battery ages, capturing the non-stationary characteristics of capacity decay (such as nonlinear decay inflection points), avoiding the cumulative errors of static models, and effectively improving long-term prediction accuracy.

[0060] In addition, the L2 regularization term of the ridge regression model constrains the magnitude of the regression coefficient, suppressing the risk of overfitting due to feature collinearity or small sample data. In scenarios where the target battery data is small and feature noise is large (such as sensor error), the model output stability is improved.

[0061] Finally, this embodiment does not introduce complex network structures or high-dimensional feature transformations, and only relies on linear operations. The model has low memory usage and is suitable for resource-constrained embedded hardware, reducing the hardware upgrade cost of the battery management system.

[0062] On the other hand, the first characteristic data includes any one of the constant current charging cut-off power and the IC curve peak value. Similarly, the second characteristic data includes any one of the constant current charging cut-off power and the IC curve peak value. The IC curve is a curve showing the relationship between the capacity increment and the voltage of the battery.

[0063] In addition, the first characteristic data is the same as or different from the second characteristic data. For example, the first characteristic data is the constant current charging cut-off power, and the second characteristic data is the IC curve peak. For another example, both are the constant current charging cut-off power or the IC curve peak.

[0064] On the other hand, the correlation between the first feature data and the second feature data and the health state of the battery is evaluated according to any one or more of the Pearson correlation coefficient, the Spearman correlation coefficient and the Kendall correlation coefficient, so that the first feature data and the second feature data are strongly correlated with the health state of the battery.

[0065] On the other hand, a method for aligning the first data set and the second data set includes:

[0066] S31: Calculate the distance between the probability distributions of a number of the first feature data in the first data set and a number of the second feature data sets in the second data set to obtain a difference value.

[0067] S32: Determine whether the difference value reaches a preset minimum value. If not, proceed to S33. If yes, it indicates that the alignment has been completed, and the current alignment process ends.

[0068] S33: Perform a rotation transformation on the current first feature data in the first data set based on the geometric transformation matrix, and return to S31.

[0069] In this embodiment, the maximum mean difference (MMD) is used to calculate the distance between the probability distributions of a number of the first feature data in the first data set and a number of the second feature data sets in the second data set.

[0070] A rotation matrix is ​​used to perform a rotation transformation on the current first feature data in the first data set.

[0071] The rotation matrix is ​​a way to change the distance between two domains. The rotation matrix does not change the direction of the principal components of the data, but only rearranges the directions of these principal components. Therefore, the rotated eigenvectors still retain the initial aging law. The rotation matrix M is as follows:

[0072]

[0073] Where θ is the rotation parameter, and the first characteristic data of the battery before and after rotation are F a and F b , the relationship is as follows:

[0074] F b =M(θ)×F a

[0075] On the other hand, the aging features of the source domain and the target domain may present an asymmetric or skewed distribution (for example, some features are concentrated at one end). At this time, MMD cannot accurately measure the true distribution difference between the two domains. When the rotation matrix adjusts the source domain features, the optimal angle θ cannot be found, resulting in the failure of the alignment of the distributions of the two domains, and the subsequent transfer learning effect is significantly reduced. Moreover, the ridge regression model assumes that if the input features are severely skewed or outliers, the regression coefficient estimation will be biased, resulting in an increase in the prediction error and an increase in the risk of overfitting. In this regard, the estimation method in this embodiment also includes:

[0076] Before aligning the first feature data in the first data set and the second feature data in the second data set, the first feature data in the first data set are also normally transformed based on transformation parameters so that a number of the first feature data are close to a normal distribution, and the transformation parameters are fitted into the regression coefficients.

[0077] The normal transformation in this embodiment can make the extracted aging features more consistent with the normal distribution.

[0078] The transformation formula for normal transformation is as follows:

[0079]

[0080] Wherein, f>0, f represents the aging characteristic of the battery, that is, the first characteristic data, F is the data after f is transformed, and λ is the transformation parameter.

[0081] When performing a normal transformation, a series of λ is randomly generated within a preset range, and then it is determined which λ corresponds to the best normality of the transformed data, which is the desired transformation parameter.

[0082] On the other hand, the regression coefficient R is expressed as follows:

[0083] R=(FT F+λkI) -1 Fq

[0084] Wherein, λ is the transformation parameter, F is the vector expression of the first feature data after alignment, and F T is the transpose of the first feature data vector after alignment, I is the unit orthogonal matrix, and k is the ridge parameter.

[0085] The ridge parameter k can be preset or dynamically confirmed in real time. The dynamic confirmation process is as follows:

[0086] First, a series of increasing k values ​​are generated, and then, for each k value, the ridge regression coefficient R is calculated. Next, a curve of the regression coefficient R versus k value is plotted, which is also called a ridge trace plot. In the ridge trace plot, the k value at which the regression coefficient R tends to be stable is the final confirmed ridge parameter.

[0087] In summary, the present invention discloses a method for estimating the health status of a rechargeable battery, such as Figure 2 , now use a specific example to explain its working process in detail.

[0088] First of all, it should be noted that the source domain battery used for training the model is a certain type of lithium-ion battery (battery A), and complete aging data (100 charge and discharge cycles) has been obtained, and the SOH decays from 100% to 70%.

[0089] The target domain battery is a battery of the same model but with different usage conditions (battery B). It only has data for the first 20 cycles, and the SOH of subsequent cycles needs to be predicted.

[0090] 1. Data collection: Obtain the voltage-capacity (QV) curve from the constant current charging stage of battery A, and calculate the cut-off capacity of each cycle (aging feature F1, i.e., the first feature data).

[0091] Calculate the Pearson (r), Spearman (ρ), and Kendall (τ) correlation coefficients for F1 and SOH of battery A:

[0092] F1: r = 0.92, ρ = 0.89, τ = 0.85 (strong negative correlation, F1 decreases when SOH decreases).

[0093] Therefore, F1 is highly correlated with SOH and is suitable for use as an aging feature.

[0094] 2. Feature normalization: Perform normal transformation on F1 of battery A and select the optimal λ parameter (such as λ=0.5) to make the aging feature F1 close to the normal distribution.

[0095] After the transformation, the skewness of F1 dropped from 1.2 to 0.1.

[0096] The same λ=0.5 is applied to the first 20 cycle data of battery B to align its feature (ie, the second feature data) distribution with the source domain.

[0097] 3. Domain adaptation: Calculate the MMD2 values ​​of the aging features of the first 100 cycles of battery A (source domain) and the first 20 cycles of battery B (target domain). Assume that the initial MMD2 = 3.5, indicating that the distribution difference is large.

[0098] Then, the aging features of the source domain are adjusted through the rotation matrix to generate the rotated source domain features.

[0099] Recalculate MMD2. When the rotation parameter θ of the rotation matrix is ​​30°, MMD2 drops to 0.8, reaching the minimum value.

[0100] The source domain aging features are rotated to be consistent with the target domain distribution, and the aligned aging features will be used for subsequent regression.

[0101] 4. Model training: Use the aligned source domain features (100 cycle data of battery A) to train the ridge regression model and obtain the regression coefficient R.

[0102] 5. Sliding Window Prediction:

[0103] Window definition: Window size = 10 periods, sliding 1 period each time.

[0104] First prediction:

[0105] The SOH value of the 11th cycle is calculated based on the aging characteristics of the battery B in the first 10 cycles and the regression coefficient R, for example, the predicted SOH=94.5%.

[0106] Window sliding:

[0107] Use the data from the 2nd to 11th cycles to update the regression coefficient R and predict the SOH for the 12th cycle, and so on.

[0108] As the window slides, the regression coefficient R is dynamically adjusted according to the latest data to adapt to the aging characteristics of battery B.

[0109] Therefore, this estimation method has the advantages of small sample adaptation, dynamic update and efficient calculation through feature alignment and sliding ridge regression.

[0110] In another preferred embodiment of the present invention, a rechargeable battery health state estimation system is also disclosed, which includes a data processor, and the data processor estimates the health state of a target battery based on the rechargeable battery health state estimation method in the above embodiment.

[0111] The present invention also discloses another rechargeable battery health status estimation system, which includes one or more processors, a memory and one or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by the one or more processors, and the program includes instructions for executing the estimation method as described above. The processor can adopt a general central processing unit (CPU), a microprocessor, an application specific integrated circuit (ASIC), or one or more integrated circuits to execute relevant programs to implement the functions required to be executed by the modules in the estimation system of the embodiment of the present application, or to execute the estimation method of the method embodiment of the present application.

[0112] The present invention also discloses a computer-readable storage medium, which includes a computer program, and the computer program can be executed by a processor to complete the estimation method as described above. The computer-readable storage medium can be any available medium that can be accessed by a computer or a data storage device such as a server or a data center that includes one or more available media integrated. The available medium can be a read-only memory (ROM), or a random access memory (RAM), or a magnetic medium, such as a floppy disk, a hard disk, a tape, a magnetic disk, or an optical medium, such as a digital versatile disc (DVD), or a semiconductor medium, such as a solid state disk (SSD).

[0113] The embodiment of the present application also discloses a computer program product or a computer program, which includes a computer instruction stored in a computer-readable storage medium. A processor of an electronic device reads the computer instruction from the computer-readable storage medium, and the processor executes the computer instruction, so that the electronic device executes the above-mentioned rechargeable battery health state estimation method.

[0114] The above disclosure is only the preferred embodiment of the present invention, which certainly cannot be used to limit the scope of rights of the present invention. Therefore, equivalent changes made according to the scope of the patent application of the present invention are still within the scope covered by the present invention.

Claims

1. A method for estimating the health status of a rechargeable battery, characterized in that: include: Acquire aging data of a reference battery for training during several charge and discharge cycles to obtain a first data set, wherein the aging data includes first characteristic data strongly related to a health state of the battery and first state data representing the health state; Acquire second characteristic data strongly related to the health status of the battery during several charging cycles of the target battery to be tested to obtain a second data set; aligning the first feature data in the first data set and the second feature data in the second data set in terms of statistical characteristic distribution; Using the aligned first data set to train a ridge regression model to obtain regression coefficients; Predicting second status data representing the current health status of the target battery based on the plurality of second feature data and the regression coefficient within the current time window of the target battery; The ridge regression model is retrained using a plurality of the second feature data and the second state data in the current time window as aging data for training, so as to update the regression coefficient.

2. The method for estimating the health status of a rechargeable battery according to claim 1, characterized in that: The first feature data is the same as or different from the second feature data.

3. The method for estimating the health status of a rechargeable battery according to claim 1, characterized in that: The first characteristic data includes any one of the constant current charging cut-off power and the IC curve peak value; the second characteristic data includes any one of the constant current charging cut-off power and the IC curve peak value; the IC curve is a curve showing the relationship between the capacity increment and the voltage of the battery.

4. The method for estimating the health status of a rechargeable battery according to claim 1, characterized in that: The correlation between the first feature data and the second feature data and the health state of the battery is evaluated according to any one or more of the Pearson correlation coefficient, the Spearman correlation coefficient and the Kendall correlation coefficient, so that the first feature data and the second feature data are strongly correlated with the health state of the battery.

5. The method for estimating the health status of a rechargeable battery according to claim 1, characterized in that: The method for aligning the first data set and the second data set includes: Calculating the distance between the probability distributions of a plurality of the first feature data in the first data set and a plurality of the second feature data sets in the second data set to obtain a difference value; Determine whether the difference value reaches a preset minimum value. If not, perform a rotation transformation on the current first feature data in the first data set based on a geometric transformation matrix, and recalculate the difference value. If yes, complete the alignment.

6. The method for estimating the health status of a rechargeable battery according to claim 1, characterized in that: Before aligning the first feature data in the first data set and the second feature data in the second data set, the first feature data in the first data set are also normally transformed based on transformation parameters so that a number of the first feature data are close to a normal distribution, and the transformation parameters are fitted into the regression coefficients.

7. The method for estimating the health status of a rechargeable battery according to claim 6, characterized in that: The regression coefficient R is expressed as follows: R=(F T F+λkI) -1 Fq Wherein, λ is the transformation parameter, F is the vector expression of the first feature data after alignment, and F T is the transpose of the first feature data vector after alignment, I is the unit orthogonal matrix, and k is the ridge parameter.

8. A rechargeable battery health status estimation system, characterized in that: The method comprises a data processor, wherein the data processor estimates the health state of a target battery based on the rechargeable battery health state estimation method according to any one of claims 1 to 7.

9. A rechargeable battery health status estimation system, characterized in that: include: one or more processors; Memory; and one or more programs, wherein the one or more programs are stored in the memory and are configured to be executed by the one or more processors, the programs comprising instructions for executing the rechargeable battery health status estimation method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The method comprises a computer program which can be executed by a processor to implement the method for estimating the health state of a rechargeable battery as claimed in any one of claims 1 to 7.

Images