Lithium ion battery remaining service life estimation method combining multi-scale window features and Gaussian process regression

Abstract

The invention relates to a lithium ion battery remaining service life estimation method combining multi-scale window features and Gaussian process regression, and belongs to the technical field of battery management. The method comprises the following steps: calculating multi-dimensional health characteristics related to charging cut-off time, voltage and temperature according to a constant-current charging stage of the lithium ion battery; performing normalization processing on the multi-dimensional health features, and solving a covariance matrix maximization problem of the multi-dimensional health features through a principal component analysis technology to obtain fused health features; identifying capacity regeneration points in the battery aging process according to the statistical indexes fused with the health characteristics, and dividing capacity regeneration intervals; setting the size of a multi-scale window according to the divided capacity regeneration interval, and calculating the characteristics of the multi-scale window; and estimating the remaining service life of the battery in combination with the multi-scale window characteristics and Gaussian process regression. According to the invention, the interference of the capacity regeneration phenomenon on the remaining service life of the battery can be solved, and accurate and reliable lithium ion battery performance evaluation is realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of battery management and relates to a method for estimating the remaining service life of a lithium-ion battery by combining multi-scale window features and Gaussian process regression. Background Art

[0002] Lithium-ion batteries, with their high energy density and strong charge retention, are widely used in new energy vehicles and energy storage systems. However, lithium-ion batteries age with repeated charge and discharge cycles, further impacting overall system performance. Therefore, accurately estimating a battery's remaining useful life (RUL) can help operators manage and maintain the battery promptly, improving resource utilization. Furthermore, RUL is a core evaluation metric for battery management systems. Accurate RUL estimation reduces the risk of battery failure and ensures operational stability.

[0003] The RUL of a lithium-ion battery is typically defined as the number of usable cycles from the current moment to a set failure threshold. The RUL indirectly reflects the degree of battery aging. Degradation in battery performance not only affects the device's battery life but can also lead to safety issues such as overheating, short circuits, and even fires. Therefore, accurately and reliably assessing the RUL of lithium-ion batteries is crucial for decision-making within the battery management system (BMS).

[0004] At present, the RUL estimation of lithium-ion batteries is mainly divided into model-based parameter identification and data-driven methods. Model-based parameter identification mainly studies the capacity degradation process by constructing an equivalent circuit model and an electrochemical model of the battery to estimate the RUL of the battery. The model-based parameter identification method can explain the battery aging process in detail from the physical and chemical principles, but the parameters that the model needs to identify will increase with the changes in battery operating conditions and materials. Therefore, data-driven methods that do not require detailed aging prior knowledge are more suitable for applications in multi-operating battery scenarios. However, most data-driven models do not consider the impact of capacity regeneration. The capacity regeneration phenomenon is manifested as the battery capacity not only does not decrease with the increase in the number of charge and discharge cycles, but instead shows a short-term increase. This phenomenon will break the linear degradation trend of the battery RUL and further affect the reliability of the battery RUL estimation.

[0005] The occurrence of capacity regeneration makes it difficult for traditional RUL estimation methods to accurately reflect the true state of battery aging. For example, health state estimation models based on capacity decay may underestimate the extent of battery aging due to capacity regeneration, leading to incorrect decisions by the battery management system. This not only affects battery efficiency but also increases battery safety risks. Therefore, it is necessary to design a new RUL estimation method for lithium-ion batteries to address the interference caused by capacity regeneration. Summary of the Invention

[0006] In view of this, the purpose of the present invention is to provide a method for estimating the remaining useful life of lithium-ion batteries by combining multi-scale window features and Gaussian process regression, so as to solve the problems of interference from capacity regeneration phenomenon and engineering implementation difficulties in existing technical means, and improve the accuracy and reliability of battery health status assessment.

[0007] In order to achieve the above object, the present invention provides the following technical solutions:

[0008] A method for estimating the remaining useful life of a lithium-ion battery by combining multi-scale window features and Gaussian process regression specifically comprises the following steps:

[0009] S1: Calculate multi-dimensional health characteristics related to charging cut-off time, voltage and temperature according to the constant current charging stage of lithium-ion batteries;

[0010] S2: Normalize the multidimensional health features and solve the covariance matrix maximization problem of the multidimensional health features through principal component analysis technology to further obtain the fused health features;

[0011] S3: Identify the capacity regeneration point (CRP) of the battery aging process based on statistical indicators that integrate health characteristics, and divide the capacity regeneration interval;

[0012] S4: Set the size of the multi-scale window according to the divided capacity regeneration interval and calculate the multi-scale window features (MWF);

[0013] S5: Combine multi-scale window features (MWF) with Gaussian process regression (GPR) to estimate the remaining useful life (RUL) of the battery.

[0014] Furthermore, step S1 specifically includes: directly extracting or indirectly calculating health features according to the constant current charging stage of the battery. The directly extracted health features include the time and temperature at the end of charging, the voltage, time and temperature at the inflection point of the curve, and the indirectly calculated health features include the approximate gradient, average gradient and area of ​​the curve:

[0015] G s1 (G s2 )=(V1(T e1 )-V0(T e0 )) / (T i1 -T i0 )

[0016] G m =mean(gradient(V0:V1,T i0 :T i1 ))

[0017] S V (S Te )=trapz(T i0 :T i1 ,V0(T e0 ):V1(T e1 ))

[0018] Among them, V0 and V1 represent the constant current charging starting voltage and cut-off voltage respectively, T i0 and T e0 They represent the constant current charging starting time and starting temperature, T i1 and T e1 Respectively represent the time and temperature when constant current charging is cut off, G s1 and G s2 Represent the approximate gradients of the voltage curve and temperature curve, G m is the average gradient of the voltage curve, S V and S Te represent the areas of the voltage curve and temperature curve, respectively. gradient(·), mean(·), and trapz(·) represent the gradient function, mean function, and area function in MATLAB software, respectively.

[0019] Furthermore, step S2 specifically includes the following steps:

[0020] S21: Use linear function to normalize the original multidimensional health feature (DF):

[0021]

[0022] Among them, DF norm is the normalized health feature, DF min and DF maxRepresent the minimum and maximum values ​​in the data respectively;

[0023] S22: Calculate the covariance between features and construct the covariance matrix of battery health features:

[0024]

[0025] Among them, Cov(DF1,DF2) represents the covariance between health features DF1 and DF2, and are the i-th battery sample and mean value of the first health feature, Cov(DF) is the covariance matrix, j is the number of features, and n is the total number of samples;

[0026] S23: Construct the covariance matrix maximization problem and transform it into an unconstrained problem through the Lagrangian function D(x):

[0027]

[0028] D(x)=ξ T Cov(DF)ξ-λ(ξ T ξ-1)

[0029] Where ξ is the eigenvector, (·) T is the matrix transpose operation, ξ T ξ=1 is the constraint condition, λ is the eigenvalue;

[0030] S24: Derivative of ξ, and let Solve the characteristic equation and sort the eigenvalues ​​(λ) from largest to smallest:

[0031] |Cov(DF)-λI|=0;λ1≥λ2≥...≥λ j

[0032] Where I is the identity matrix;

[0033] S25: Usually the health feature corresponding to λ>1 is selected as the fusion health feature, so the final selected battery fusion health feature matrix FF d 'for:

[0034]

[0035] Where d is the fusion health feature selection index of the battery.

[0036] Furthermore, step S3 specifically includes the following steps:

[0037] S31: Calculate the fusion health feature change rate (ΔFF) of adjacent time steps in the training dataset t ):

[0038] ΔFF t =FF t -FF t-1

[0039] Among them, FF t and FF t-1 are the battery fusion health characteristics at time t and time t-1 respectively;

[0040] S32: By calculating ΔFF t The median and absolute median difference of the two groups were used to set a robust statistical threshold (θ CRP ):

[0041] θ CRP =median(ΔFF t )+β×mad(ΔFF t )

[0042] Where median(·) and mad(·) are the median function and absolute median difference function of MATLAB software, respectively, which can evaluate the dispersion of the feature change rate; β is the weight coefficient, which is adjusted through testing;

[0043] S33: According to θ CRP Identify the capacity regeneration point (CRP), when ΔFF t >θ CRP , then the point is determined to be the capacity regeneration point (CRP); the area between the previous moment of the capacity regeneration point (CRP) and the next identical capacity point is defined as the capacity regeneration interval. This definition can ensure the global continuity of the subsequent multi-scale window feature (MWF).

[0044] Further, step S4 specifically includes the following steps:

[0045] S41: Calculate the fusion feature change rate (ψFF) at the current moment relative to the initial moment t ):

[0046] ψFF t =(FF t -FF0) / FF0

[0047] Among them, FF0 represents the battery fusion health characteristics at the initial moment, FF t is the battery fusion health characteristics at time t;

[0048] S42: Select the minimum and maximum capacity growth intervals as multi-scale windows and calculate ψFF within the multi-scale windows. t The mean and median of are used as multi-scale window features MWF:

[0049] MWF t=[mean(ψFF t (t-windows:t-1));

[0050] median(ψFF t (t-windows:t-1))]

[0051] Among them, MWF t Represents the multi-scale window features at time t.

[0052] Further, step S5 specifically includes the following steps:

[0053] S51: Assuming that the training set has t observation samples, the corresponding observation values ​​are f(x1) to f(x t ), denoted as f(x 1:t ), which also corresponds to the random variable F(x 1:t ), the maximum probability parameter value based on the known observations Expressed as:

[0054]

[0055] S52: Solve the probability distribution of parameters P(θ|F(x 1:t )=f(x 1:t )), assuming that it satisfies the equal probability distribution, the maximum posterior estimation is degenerated into the maximum likelihood estimation:

[0056]

[0057] Among them, P(F(x 1:t )=f(x 1:t )|θ) is the observed F(x 1:t )=f(x 1:t ), P(θ) is the prior probability distribution of the solution parameters, θ is the parameter vector to be estimated, is the maximum a posteriori estimate, and η is the estimated value of the parameter;

[0058] S53: F(x 1:t ) obeys the multivariate normal distribution, then its mean (μ t ) and covariance (Г t ) are t-dimensional vectors and t×t matrices, respectively, to further solve the probability density function:

[0059]

[0060] Among them, L(x 1:t |θ) represents the likelihood function of the multivariate normal distribution, y t represents F(x1:t ) corresponds to t observations; since x 1:t and f(x 1:t ) is known, the probability density function is converted into a function with parameters as variables. The GPR process is to solve the parameter value that maximizes the probability density function. In order to simplify the functional form, take the log-likelihood function:

[0061]

[0062] The above formula is a function form for solving a known gradient. The optimal value can be found by searching for the minimum point of the objective function through the gradient descent method.

[0063] Furthermore, the method was used to calculate the root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), relative standard deviation (RSD) and coefficient of determination (R 2 ) to evaluate the accuracy.

[0064] The beneficial effects of the present invention are: the present invention constructs an accurate and reliable model for estimating the remaining useful life of lithium-ion batteries. First, direct health features are extracted based on the constant current charging stage of the battery, and fused health features are obtained through principal component analysis technology to solve the multicollinearity problem between multidimensional features; secondly, the capacity regeneration point (CRP) detection threshold is set by calculating the statistical indicators of the fused health features, so as to effectively identify the CRP of the battery degradation process, and determine the capacity regeneration interval based on this; then, a multi-scale window feature (MWF) is constructed through the capacity regeneration interval, and a hybrid model is formed by combining Gaussian process regression (GPR) to achieve accurate estimation of the remaining useful life (RUL) of the battery; finally, the root mean square error (RMSE) and other indicators are calculated to comprehensively evaluate the RUL estimation effect of the proposed model. The hybrid model proposed in the present invention can avoid the influence of the capacity regeneration point and improve the accuracy and reliability of the RUL estimation of lithium-ion batteries.

[0065] Other advantages, objects, and features of the present invention will be described in part in the following description and, in part, will be apparent to those skilled in the art upon examination of the following description or may be learned from practice of the present invention. The objects and other advantages of the present invention may be realized and obtained through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0066] In order to make the purpose, technical solutions and advantages of the present invention more clear, the present invention will be described in detail below with reference to the accompanying drawings, in which:

[0067] Figure 1 A flowchart of a method for estimating the remaining useful life of a lithium-ion battery combining multi-scale window features and Gaussian process regression provided by an embodiment of the present invention;

[0068] Figure 2 The capacity degradation process of lithium-ion batteries at different discharge cut-off voltages;

[0069] Figure 3 Results of CRP identification and capacity regeneration interval division for lithium-ion batteries;

[0070] Figure 4 The RUL estimation results of lithium-ion batteries using the proposed model;

[0071] Figure 5 is the evaluation index for the RUL estimation of lithium-ion batteries of the proposed model. DETAILED DESCRIPTION

[0072] The following describes the embodiments of the present invention by means of specific examples, and those skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification. The present invention can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present invention. It should be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present invention, and the following embodiments and features in the embodiments can be combined with each other without conflict.

[0073] See also Figures 1 to 5 , an embodiment of the present invention provides a method for estimating the remaining useful life of a lithium-ion battery by combining multi-scale window features and Gaussian process regression. The simulation software used in this embodiment is MATLAB2023b, and the battery test data of the National Aeronautics and Space Administration (NASA) is used to perform RUL estimation experiments. The battery charging process is first charged at a constant current of 1.5A to a cut-off voltage of 4.2V, and then continues to charge at a constant voltage of 4.2V until the current drops to 20mA. The battery discharge process uses currents of different rates for constant discharge. The battery RUL estimation flow chart of the proposed model is as follows Figure 1 As shown, the method specifically includes the following steps:

[0074] Step S1: Directly extract or indirectly calculate health features according to the constant current charging stage of the battery. The directly extracted health features include the time and temperature at the end of charging, the voltage, time and temperature at the inflection point of the curve. The indirectly calculated health features include the approximate gradient, average gradient and area of ​​the curve:

[0075] G s1 (G s2 )=(V1(T e1 )-V0(T e0 )) / (T i1 -T i0 )

[0076] G m =mean(gradient(V0:V1,T i0 :T i1 ))

[0077] S V (S Te )=trapz(T i0 :T i1 ,V0(T e0 ):V1(T e1 ))

[0078] Among them, V0 and V1 represent the constant current charging starting voltage and cut-off voltage respectively, T i0 and T e0 They represent the constant current charging starting time and starting temperature, T i1 and T e1 Respectively represent the time and temperature when constant current charging is cut off, G s1 and G s2 Represent the approximate gradients of the voltage curve and temperature curve, G m is the average gradient of the voltage curve, S V and S Te represent the areas of the voltage curve and temperature curve, respectively. gradient(·), mean(·), and trapz(·) represent the gradient function, mean function, and area function in MATLAB software, respectively.

[0079] The capacity degradation process of lithium-ion batteries at different discharge cut-off voltages is as follows: Figure 2 As shown in the figure, as the charge and discharge cycle proceeds, the battery capacity degradation process shows a capacity regeneration phenomenon, that is, the capacity suddenly increases at a certain moment, breaking the linear degradation trend of the battery RUL, thereby seriously affecting the accuracy and reliability of the battery RUL estimation.

[0080] The Pearson correlation analysis between the health characteristics and capacity of different batteries is shown in Table 1.

[0081] Table 1 Pearson correlation coefficient between health characteristics and battery capacity As shown in Table 1, there are differences in the correlation between different health characteristics and capacity of lithium-ion batteries, and the multicollinearity between health characteristics will lead to overfitting problems. For NASA-B30, B31 and B32 batteries, T i1 、T i2 、S V and S Te Both showed strong correlation with capacity, while T e2 , G S1 and G S2 The correlation between the health features is very low. Furthermore, excessive health features increase the computational complexity of the RUL estimation model, making it unsuitable for platforms with limited computing resources. Furthermore, the correlation between the same health features changes when the battery's operating conditions change. Therefore, to balance the comprehensiveness of health features with computational efficiency, principal component analysis will be used to optimize the original health features, further deriving fused health features that incorporate the original information.

[0082] Step S2: normalize the multidimensional health features, solve the covariance matrix maximization problem of the multidimensional health features through principal component analysis technology, and further obtain the fused health features, which specifically includes the following steps:

[0083] S21: Use linear function to normalize the original multidimensional health feature (DF):

[0084]

[0085] Among them, DF norm is the normalized health feature, DF min and DF max Represent the minimum and maximum values ​​in the data respectively.

[0086] S22: Calculate the covariance between features and construct the covariance matrix of battery health features:

[0087]

[0088] Among them, Cov(DF1,DF2) represents the covariance between health features DF1 and DF2, and are the i-th battery sample and the average value of the first health feature, Cov(DF) is the covariance matrix, j is the number of features, and n is the total number of samples.

[0089] S23: Construct the maximization problem of the covariance matrix and transform it into an unconstrained problem through the Lagrangian function D(x):

[0090]

[0091] D(x)=ξ T Cov(DF)ξ-λ(ξ T ξ-1)

[0092] Where ξ is the eigenvector, (·) T is the matrix transpose operation, ξ T ξ=1 is the constraint condition.

[0093] S24: Derivative of ξ and set it equal to Solve the characteristic equation and sort the eigenvalues ​​(λ) from largest to smallest:

[0094] |Cov(DF)-λI|=0;λ1≥λ2≥...≥λ j

[0095] S25: Usually the health feature corresponding to λ>1 is selected as the fusion health feature, so the final selected battery fusion health feature matrix FF d 'for:

[0096]

[0097] Where d is the fusion health feature selection index of the battery.

[0098] Step S3: Identify the capacity regeneration point (CRP) of the battery aging process based on the statistical indicators of the integrated health characteristics and divide the capacity regeneration interval, which specifically includes the following steps:

[0099] S31: Calculate the fusion health feature change rate (ΔFF) of adjacent time steps in the training dataset:

[0100] ΔFF t =FF t -FF t-1

[0101] Among them, FF t and FF t-1 are the fused health characteristics at time t and time t-1 respectively.

[0102] S32: Set robust statistical thresholds (θ) by calculating the median and absolute median difference of ΔFF CRP ):

[0103] θ CRP =median(ΔFF t )+β×mad(ΔFF t )

[0104] Among them, median(·) and mad(·) are the median function and absolute median difference function of MATLAB software, respectively, which can evaluate the dispersion of the feature change rate. β is the weight coefficient, which is adjusted through testing.

[0105] S33: According to θ CRP Identify the Capacity Regeneration Point (CRP), when ΔCap t >θ CRP , then the point is determined to be CRP. The area between the previous moment of CRP and the next point with the same capacity is defined as the capacity regeneration interval. This definition can ensure the global continuity of the subsequent multi-scale window feature (MWF).

[0106] Figure 3 The CRP recognition results of different lithium-ion batteries are shown. In the figure, SP1 and SP2 represent different estimation starting points, SP1=17 and SP2=25. S1~S4 are capacity regeneration intervals divided according to CRP. The smaller interval is used to capture the degradation trend of short-term capacity regeneration phenomenon. This definition ensures the global continuity of the training data set. The CRP recognition results of NASA-B30, B31 and B32 batteries under different SP are shown in Table 2. Figure 3 As shown in Table 2, the proposed fusion threshold can accurately realize the CRP identification of the battery at different discharge cut-off voltages and different estimation starting points. Under different SPs, the number of incorrectly identified CRPs does not exceed 1.

[0107] Table 2 Analysis of CRP recognition results of different SPs for NASA-B30, B31 and B32

[0108]

[0109] Step S4: setting the size of the multi-scale window according to the divided capacity regeneration interval and calculating the multi-scale window feature (MWF), specifically including the following steps:

[0110] S41: Calculate the fusion feature change rate (ψFF) at the current moment relative to the initial moment:

[0111] ψFF t =(FF t -FF0) / FF0

[0112] Among them, FF0 represents the battery fusion health characteristics at the initial moment.

[0113] S42: Select the minimum and maximum capacity growth intervals as multi-scale windows and calculate ψFF within the multi-scale windows. t The mean and median of as MWF:

[0114] MWFt =[mean(ψFF t (t-windows:t-1));

[0115] median(ψFF t (t-windows:t-1))]

[0116] Step S5: Combine MWF with Gaussian process regression (GPR) to estimate the battery RUL, and calculate indicators such as root mean square error (RMSE) to evaluate the battery RUL estimation effect of the proposed model. Specifically, the following steps are included:

[0117] S51: Assuming that the training set has t observation samples, the corresponding observation values ​​are f(x1) to f(x t ), denoted as f(x 1:t ), which also corresponds to the random variable F(x 1:t ), the maximum probability parameter value based on the known observations Expressed as:

[0118]

[0119] S52: Solve the probability distribution of parameters P(θ|F(x 1:t )=f(x 1:t )), assuming that it satisfies the equal probability distribution, the maximum posterior estimation is degenerated into the maximum likelihood estimation:

[0120]

[0121] Among them, P(F(x 1:t )=f(x 1:t )|θ) is the observed F(x 1:t )=f(x 1:t ), P(θ) is the prior probability distribution of the solution parameters, θ is the parameter vector to be estimated, is the maximum a posteriori estimate, and η is the estimated value of the parameter.

[0122] S53: F(x 1:t ) obeys the multivariate normal distribution, then its mean (μ t ) and covariance (Г t ) are t-dimensional vectors and t×t matrices, respectively, to further solve the probability density function:

[0123]

[0124] Among them, L(x 1:t|θ) represents the likelihood function of the multivariate normal distribution, y t represents F(x 1:t ) corresponds to t observations. Since x 1:t and f(x 1:t ) is known, the probability density function is converted into a function with parameters as variables. The GPR process is to solve the parameter value that maximizes the probability density function. In order to simplify the functional form, take the log-likelihood function:

[0125]

[0126] S54: The above formula is a function form for solving a problem with a known gradient. The optimal value can be found by searching for the minimum point of the objective function using the gradient descent method.

[0127] S55: Calculate the root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE), relative standard deviation (RSD) and coefficient of determination (R 2 ) to evaluate the RUL estimation effect of the proposed model:

[0128]

[0129] RSD=std(X true -X est ) / (max(X true )-min(X true ))×100%

[0130]

[0131] Among them, X true and X est represents the true data and the estimated data, respectively. K represents the number of samples in the test set. std(·), max(·), and min(·) represent the standard deviation function, the maximum function, and the minimum function, respectively. RMSE reflects the difference between the estimated value and the true value. RSD reflects the degree of dispersion of the estimation error.

[0132] Figure 4 and Figure 5 They are the RUL estimation results and evaluation indicators under different batteries and different estimation starting points. Figure 4As can be seen, the RUL estimation results of the proposed model are generally controlled within ±2 cycles, very close to the actual RUL of the battery, and the overall estimation process has small fluctuations. To qualitatively analyze the battery RUL estimation effect of the proposed model, Table 3 shows the RUL evaluation index values ​​of three battery types at different SPs.

[0133] Table 3 RUL evaluation indicators of different SPs for NASA-B30, B31 and B32 As shown in Table 3, the model proposed in this embodiment has a RMSE of no more than 1.1796, a MAE of no more than 0.9356, a MAPE of no more than 14.96%, an RSD of no more than 5.12%, and an R 2 No more than 0.9685. In the case of battery SP2, RMSE and MAE are both less than 1 cycle, indicating that the accuracy of RUL estimation is high. At the same time, due to the change in data scale, the MAPE of the three batteries has a slight increase, but still maintains good stability. 2 They are all greater than 0.96, indicating that the estimation results can well fit the real RUL of the battery. The data of the proposed model is highly interpretable, and further accurate battery RUL estimation is achieved, which is of great significance for designing a safe and reliable battery management system.

[0134] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not limiting. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention can be modified or replaced by equivalents without departing from the purpose and scope of the technical solutions, which should all be included in the scope of the claims of the present invention.

Claims

1. A method for estimating the remaining useful life of a lithium-ion battery by combining multi-scale window features and Gaussian process regression, characterized in that: The method specifically comprises the following steps: S1: Calculate multi-dimensional health characteristics related to charging cut-off time, voltage and temperature according to the constant current charging stage of lithium-ion batteries; S2: Normalize the multidimensional health features and solve the covariance matrix maximization problem of the multidimensional health features through principal component analysis technology to further obtain the fused health features; S3: Identify the capacity regeneration points during the battery aging process based on statistical indicators that integrate health characteristics, and divide the capacity regeneration intervals; S4: setting the size of the multi-scale window according to the divided capacity regeneration interval and calculating the multi-scale window features; S5: Combine multi-scale window features with Gaussian process regression to achieve battery remaining service life estimation.

2. The method for estimating the remaining useful life of a lithium-ion battery according to claim 1, wherein: Step S1 specifically includes: directly extracting or indirectly calculating health features according to the constant current charging stage of the battery. The directly extracted health features include the time and temperature at the end of charging, the voltage, time and temperature at the inflection point of the curve, and the indirectly calculated health features include the approximate gradient, average gradient and area of ​​the curve: G s1 (G s2 )=(V1(T e1 )-V0(T e0 )) / (T i1 -T i0 ) G m =mean(gradient(V0:V1,T i0 :T i1 )) S V (S Te )=trapz(T i0 :T i1 ,V0(T e0 ):V1(T e1 )) Among them, V0 and V1 represent the constant current charging starting voltage and cut-off voltage respectively, T i0 and T e0 They represent the constant current charging starting time and starting temperature, T i1 and T e1 Respectively represent the time and temperature when constant current charging is cut off, G s1 and G s2 Represent the approximate gradients of the voltage curve and temperature curve, G m is the average gradient of the voltage curve, S V and S Te represent the areas of the voltage curve and temperature curve, respectively. gradient(·), mean(·), and trapz(·) represent the gradient function, mean function, and area function in MATLAB software, respectively.

3. The method for estimating the remaining service life of a lithium-ion battery according to claim 1, wherein: Step S2 specifically includes the following steps: S21: Use linear function to normalize the original multidimensional health feature DF: Among them, DF norm is the normalized health feature, DF min and DF max Represent the minimum and maximum values ​​in the data respectively; S22: Calculate the covariance between features and construct the covariance matrix of battery health features: Among them, Cov(DF1,DF2) represents the covariance between health features DF1 and DF2, DF1 i and are the i-th battery sample and mean value of the first health feature, Cov(DF) is the covariance matrix, j is the number of features, and n is the total number of samples; S23: Construct the maximization problem of the covariance matrix and transform it into an unconstrained problem through the Lagrangian function D(x): D(x)=ξ T Cov(DF)ξ-λ(ξ T x-1) Where ξ is the eigenvector, (·) T is the matrix transpose operation, ξ T ξ=1 is the constraint condition, λ is the eigenvalue; S24: Derivative of ξ, and let Solve the characteristic equation and sort the eigenvalues ​​λ from large to small: |Cov(DF)-λI|=0;λ1≥λ2≥...≥λ j Where I is the identity matrix; S25: Select the health feature corresponding to λ>1 as the fusion health feature, and the final selected battery fusion health feature matrix FF d 'for: Where d is the fusion health feature selection index of the battery.

4. The method for estimating the remaining service life of a lithium-ion battery according to claim 1, wherein: Step S3 specifically includes the following steps: S31: Calculate the fusion health feature change rate ΔFF of adjacent time steps in the training dataset t : ΔFF t =FF t -FF t-1 Among them, FF t and FF t-1 are the battery fusion health characteristics at time t and time t-1 respectively; S32: By calculating ΔFF t The median and absolute median difference were used to set robust statistical thresholds. Wherein, median(·) and mad(·) are the median function and absolute median difference function of MATLAB software, respectively, and β is the weight coefficient; S33: According to Identify the capacity regeneration point when The point is determined to be a capacity regeneration point; the area between the moment before the capacity regeneration point and the next same capacity point is defined as the capacity regeneration interval.

5. The method for estimating the remaining service life of a lithium-ion battery according to claim 1, wherein: Step S4 specifically includes the following steps: S41: Calculate the fusion feature change rate ψFF at the current moment relative to the initial moment t : ψFF t =(FF t -FF0) / FF0 Among them, FF0 represents the battery fusion health characteristics at the initial moment, FF t is the battery fusion health characteristics at time t; S42: Select the minimum and maximum capacity growth intervals as multi-scale windows respectively, and calculate ψFF within the multi-scale windows t The mean and median of are used as multi-scale window features MWF: MWF t =[mean(ψFF t (t-windows:t-1)); median(ψFF t (t-windows:t-1))] Among them, MWF t Represents the multi-scale window features at time t.

6. The method for estimating the remaining service life of a lithium-ion battery according to claim 1, wherein: Step S5 specifically includes the following steps: S51: Assuming that the training set has t observation samples, the corresponding observation values ​​are f(x1) to f(x t ), denoted as f(x 1:t ), which also corresponds to the random variable F(x 1:t ), the maximum probability parameter value based on the known observations Expressed as: S52: Solve the probability distribution of parameters P(θ|F(x 1:t )=f(x 1:t )), assuming that it satisfies the equal probability distribution, the maximum posterior estimation is degenerated into the maximum likelihood estimation: Among them, P(F(x 1:t )=f(x 1:t )|θ) is the observed F(x 1:t )=f(x 1:t ) is the probability density function; P(θ) is the prior probability distribution of the solution parameters, θ is the parameter vector to be estimated, is the maximum a posteriori estimate, and η is the estimated value of the parameter; S53: F(x 1:t ) obeys the multivariate normal distribution, then its mean μ t and covariance Г t They are t-dimensional vectors and t×t matrices, respectively, to further solve the probability density function: Among them, L(x 1:t |θ) represents the likelihood function of the multivariate normal distribution, y t represents F(x 1:t ) corresponds to t observations; take the log-likelihood function: The above formula is a function form for solving a known gradient, and the optimal value is found by searching for the minimum point value of the objective function through the gradient descent method.

7. The method for estimating the remaining service life of a lithium-ion battery according to claim 1, wherein: The method evaluates accuracy by calculating the root mean square error, mean absolute error, mean absolute percentage error, relative standard deviation, and coefficient of determination.

Images