A Method for Estimating the State of Health of Lithium-Ion Batteries Based on the SSA-SVR Model

Abstract

The present invention discloses a method for estimating the state of health (SOH) of lithium-ion batteries based on the SSA-SVR model. The technical solution adopted by the present invention is as follows: data acquisition; health feature extraction; construction of a feature processing model based on Savitzky-Golay filtering; using the matrix F = [F1, F2, F3, F4] consisting of the extracted and processed health features as input and the battery's state of health (SOH) as output, to construct a training set and a prediction set; step 5, constructing an SSA-SVR model: constructing a baseline SVR model, and based on the above training set, using a sparrow search algorithm to optimize the hyperparameters of the SVR model, constructing an SOH estimation model, and outputting the estimated SOH. The present invention adopts an SVR model that considers four indicators of voltage and current segment data during the charging process, feature processing, and parameter adjustment combined with a sparrow search algorithm to achieve SOH prediction for lithium-ion batteries.

Description

Technical Field

[0001] The present invention relates to the technical field of lithium-ion battery state estimation, in particular to a lithium-ion battery state of health (SOH) estimation method based on an SSA-SVR model. Background Art

[0002] Lithium-ion batteries, with their ability to strike a good balance between performance parameters such as cost, lifespan, safety, and environmental impact, play a vital role in grid energy storage systems and new energy vehicles. However, the performance of lithium-ion batteries gradually degrades over their lifespan, posing safety risks to battery systems. Accurate estimation of the state of health (SOH) of lithium-ion batteries is necessary to improve system resilience to risky failures and reduce maintenance costs. However, parameters directly related to battery SOH are internal variables that are difficult to measure directly with sensors. Estimation requires measurement of characteristic parameters. The SOH of a lithium-ion battery is generally defined as the ratio of measured capacity to nominal capacity. When this ratio drops below a certain level, the battery cell should be replaced and maintained. Existing SOH estimation methods can be categorized into three categories: model-based, data-driven, and fusion-based. Model-based methods require the construction of highly complex models to achieve high-precision SOH estimation results, making them difficult to apply in practice. Fusion-based methods aim to combine the advantages of different methods, but have not made significant progress, suffer from model compatibility issues, and are computationally complex. Data-driven methods do not require analysis of complex reaction mechanisms and can extract hidden mapping relationships from the data. Therefore, in comparison, data-driven methods, which are more flexible and more applicable, are currently a hot topic of research. Data-driven methods are mainly affected by two aspects: input features and machine learning algorithms.

[0003] High-quality input features have the characteristics of stable required data, easy extraction, and high correlation. Some studies have used data from the discharge process for feature extraction, but because the discharge process is greatly affected by working conditions and loads, it is not as stable as the charging process data. Using the features of discharge data is not conducive to the model adapting to complex working conditions. Features with high correlation play an important role in improving the accuracy of modeling. After feature extraction, some processing can be performed to improve the correlation between features and SOH. Existing machine learning models are of various types, with different principles and each has its own advantages; and machine learning methods are more dependent on the setting of model hyperparameters. Inappropriate parameters will lead to insufficient model generalization ability. Therefore, it is urgent to extract features based on fragmented charging data, and combine feature processing to further improve the correlation between the proposed features and the target quantity. At the same time, it is a problem that needs to be solved by using high-performance search algorithms to adjust the parameters of the machine learning model to improve the accuracy of SOH estimation. Summary of the Invention

[0004] In order to overcome the shortcomings of the above-mentioned prior art, the present invention proposes a lithium-ion battery SOH estimation method based on the SSA-SVR model, in order to mine features that are highly correlated with SOH from charging data fragments, and adopts the Savitzky-Golay filtering algorithm to process features with poor correlation to further improve the correlation of existing features. At the same time, the sparrow search algorithm is used to optimize the hyperparameters of SVR, thereby improving the prediction accuracy of the SOH estimation model.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is: a lithium-ion battery SOH estimation method based on the SSA-SVR model, which includes:

[0006] Step 1, data collection: charge and discharge the lithium-ion battery multiple times, and record the current, voltage, time data of the lithium battery during the charge and discharge process, as well as the capacity of each complete discharge;

[0007] Step 2: Extract health features: Extract four health features from the voltage and current data segments during the charging process, namely, the voltage value after charging, the current value after charging, the charging duration of the voltage data segment, and the charging duration of the current data segment;

[0008] Step 3: Construct a feature processing model based on Savitzky-Golay filtering: Select the Savitzky-Golay filtering algorithm to process the extracted health features, eliminate the chaotic fluctuations of the health features caused by the influence of sensor drift noise on the charging data, and construct a feature processing model;

[0009] Step 4: Using the matrix F = [F1, F2, F3, F4] consisting of the extracted and processed health features as input and the battery SOH as output, construct a training set and a prediction set;

[0010] Step 5: Build an SSA-SVR model: Build a baseline SVR model. Based on the above training set, use the sparrow search algorithm to optimize the hyperparameters of the SVR model, build an SOH estimation model, and output the estimated SOH.

[0011] Furthermore, the step 2 includes:

[0012] Step 2.1: Extract the voltage data fragment during the charging process [V a ,V b The charging duration of ] is the first health feature F1, and the feature F1 is calculated using formula (1):

[0013] F1=t b -t a (1)

[0014] Where, t a As the voltage rises to V a The corresponding charging time, t b As the voltage rises to V b The corresponding charging time.

[0015] Step 2.2: Extract the current data fragment during the charging process [I c ,I d The charging duration of ] is the second health feature F2, and the feature F2 is calculated using formula (2):

[0016] F2=t d -t c (2)

[0017] Where, t c As the current drops to I c The corresponding charging time, t d As the current drops to I d The corresponding charging time.

[0018] Step 2.3: Extract the constant current charging process from the voltage V e As the starting point of the timing, charging t e The voltage value after the time is taken as the third health feature F3, and the feature F3 is calculated using formula (3):

[0019] F3=V e +ΔV e (3)

[0020] Where, ΔV e t e The voltage value that changes during the charging time.

[0021] Step 2.4: Extract the constant voltage charging process from the current I f As the starting point of the timing, charging t f The current value after the time is taken as the fourth health feature F4, and the feature F4 is calculated using formula (4):

[0022] F4=I f +ΔI f (4)

[0023] Among them, ΔI f is the current value that changes during the charging time tf.

[0024] Furthermore, the step 3 includes:

[0025] Step 3.1: Calculate the correlation between the four extracted health characteristics and the target value SOH using formula (5):

[0026]

[0027] Where X and Y are health characteristics and SOH samples, respectively.

[0028] Step 3.2: Use Savitzky-Golay filter to construct feature processing model, which is determined by polynomial order N and window length L = 2M + 1, where M represents the absolute value of the coordinates of the maximum and minimum measurement points in the measurement window. For data with each measurement point (-M, -M + 1, ..., 0, 1, ..., M-1, M), use equations (6), (7) and (8) for processing:

[0029]

[0030]

[0031]

[0032] Where p(n) is the limiting polynomial, a k represents the coefficient of the polynomial term, ε N is the residual of the least squares fit, x[n] is the true value of the data set in the nth sequence, and the system of equations derived from (8) is solved to finally obtain the optimal polynomial coefficients.

[0033] Step 3.3, set the threshold to 0.97. When the Pearson correlation coefficient between the health feature and SOH is greater than the set threshold, the feature processing model will not be activated, and the original aging feature will be used directly to estimate SOH. When the Pearson correlation coefficient between the health feature and SOH is less than the set threshold, it indicates that the health feature is greatly affected by sensor drift noise. The health feature should be processed to improve the correlation between the health feature and the target quantity.

[0034] Furthermore, the step 5 includes:

[0035] Step 5.1: Build a baseline SVR model

[0036] S={x i ,y i |x i ∈R m ,y i ∈R};i=1,2,…,T (9)

[0037] Where x i is the feature vector of the i-th sample, y i is the corresponding regression value, T is the number of samples, and m is the dimension of the feature vector;

[0038] The SVR function is defined as:

[0039] f(x)=ωΦ(x)+b (10)

[0040] Where f(x) is the output, Φ(x) is the nonlinear mapping function, ω and b are the parameters to be determined, and the following objective function is minimized to solve ω and b:

[0041]

[0042] Where C is the penalty coefficient, f(x i ) is the predicted value of the i-th sample, ε represents the maximum error allowed by regression, which is defined as:

[0043] |yf(x)| ε =max{0,|yf(x)|-ε} (12) Introducing the slack variable ξ i and Then, it is transformed into the following objective function:

[0044]

[0045] constraint:

[0046]

[0047] Transform Equation (10) into solving the dual problem:

[0048]

[0049] Where, β i and is the Lagrangian operator, K(x i ,x j ) is the kernel function, and the RBF kernel function with strong linear approximation ability is selected, which is defined as:

[0050]

[0051] Where σ is the width of the kernel function;

[0052] Step 5.2: Use the sparrow search algorithm to optimize the parameters of the SVR model, build the SOH estimation model, and output the estimated SOH.

[0053] Furthermore, the step 5.2 includes:

[0054] Step 5.2.1: Use formula (17) to normalize the sample data. The normalized data is divided into training set and test set.

[0055]

[0056] Where Y and Y' are the sample feature values before and after normalization of the training data, respectively.min and Y max Respectively represent the minimum and maximum values of the sample feature data before normalization;

[0057] Step 5.2.2, set the parameters of the sparrow search algorithm;

[0058] Step 5.2.3, set the value range of the penalty coefficient C and the kernel function parameter σ, and initialize the population;

[0059] Step 5.2.4: Use the above training set to train the SVR model and use formula (18) to calculate the fitness value of each sparrow:

[0060]

[0061] Where, is the predicted value of the nth training sample, Y n is the true value of the nth training sample, N tr is the number of training samples;

[0062] Step 5.2.5: Calculate and update the positions of the discoverer, joiner, and alerter according to equations (19)-(21) respectively;

[0063] The formula for updating the discoverer's position is:

[0064]

[0065] Where, represents the position of a sparrow, i is a positive integer from 1 to P, j is a positive integer from 1 to d, P is the total population, d is the dimension of the feature data, α is a random number between 0 and 1, it max is the maximum number of iterations, R2 is the warning value, ST is the safety value, Q is a random number, and E is a 1×d matrix of all 1s. When R2 ≥ ST, it means that the risk of being preyed on is high, and the sparrow at this location will move quickly; otherwise, the search range can be expanded.

[0066] The update formula for the joiner position is:

[0067]

[0068] Where, and Represent the most unfavorable and most favorable positions globally; A + is the generalized inverse matrix of A, where A represents a 1×d matrix, where the elements are randomly preset to 1 or -1; when i>n / 2, it means that the follower fails to grab the food and needs to move;

[0069] The update formula of the sentinel position is:

[0070]

[0071] Where, represents the most favorable position of the authorities, γ is a standard normal random number, representing the step length control coefficient; k is a random number between -1 and 1, representing the direction of the sparrow; f g represents the maximum fitness of the authority, f w and f g Instead, ε is a small constant to ensure that the denominator is not zero; f i represents the fitness value of the i-th sparrow;

[0072] Step 5.2.6: Get the current updated position and calculate the optimal individual and the best fitness value;

[0073] Step 5.2.7: When the training result reaches the set parameters, stop the calculation and execute the output parameters (C, σ). Otherwise, return to step 5.2.4 to recalculate the fitness value.

[0074] Step 5.2.8. Establish an estimation model based on the parameters (C, σ) and output the estimated SOH.

[0075] Compared with the prior art, the present invention has the following beneficial effects:

[0076] This paper extracts four highly correlated health features based on voltage and current data fragments during the charging process. These health features are highly correlated, efficiently extracted, and in an appropriate quantity, enabling high-precision online estimation of the state of health (SOH) of lithium-ion batteries.

[0077] Aiming at the fact that sensors are susceptible to noise during data acquisition, the present invention adopts Savitzky-Golay filter to process features with relatively low correlation, eliminate the influence of noise, further improve the correlation between the proposed features and the target quantity, and improve the prediction accuracy and stability of the model for SOH estimation.

[0078] In view of the fact that machine learning methods are more dependent on the setting of model hyperparameters and inappropriate parameters will lead to insufficient model generalization ability, the present invention adopts support vector regression, which has fewer data requirements and stronger nonlinear regression ability, as the benchmark model, and optimizes its hyperparameters through the sparrow search algorithm to ensure the effectiveness of the SOH estimation model. BRIEF DESCRIPTION OF THE DRAWINGS

[0079] Figure 1 This is a graph showing the changing trends of the four initially extracted features after normalization in a specific embodiment of the present invention;

[0080] FIG2( a ) is a comparison diagram of the F2 feature change trend before and after Savitzky-Golay filtering in a specific embodiment of the present invention;

[0081] FIG2( b ) is a comparison diagram of the F4 feature change trend before and after Savitzky-Golay filtering in a specific embodiment of the present invention;

[0082] Figure 3 This is a modeling flow chart of the SSA-SVR model in a specific embodiment of the present invention;

[0083] Figure 4 This is a comparison chart of prediction results of different methods for B0005 battery in a specific embodiment of the present invention;

[0084] Figure 5 This is a comparison chart of the prediction results of different methods for B0007 battery in a specific embodiment of the present invention. DETAILED DESCRIPTION

[0085] To make the objectives, technical solutions, and advantages of the present invention more clearly understood, the present invention will be further described in detail below in conjunction with specific embodiments and with reference to the accompanying drawings. It should be understood that these descriptions are merely exemplary and are not intended to limit the scope of the present invention. In addition, in the following description, descriptions of well-known structures and technologies are omitted to avoid unnecessary confusion of the concepts of the present invention.

[0086] This embodiment is a lithium-ion battery SOH estimation method based on the SSA-SVR model. It can achieve high feature extraction efficiency and high estimation accuracy for online SOH estimation. It combines the feature extraction of charging process fragment data to mine information related to battery degradation in different data types such as voltage and current. It also optimizes relevant features through feature processing, enhances the correlation between relevant features and target quantities, and optimizes the SVR model through the sparrow search algorithm after obtaining input features, thereby improving the prediction accuracy of the overall framework. Specifically, the estimation method is carried out as follows:

[0087] Step 1: Data collection: Charge and discharge the lithium-ion battery multiple times, and record the current, voltage, time data of the lithium battery during the charge and discharge process, as well as the capacity of each full discharge.

[0088] The lithium-ion battery degradation data used in this example comes from the NASA PCOE Research Center, which includes charge and discharge test data of several commercial lithium-ion 18650 battery packs. The battery data set records detailed data on voltage, current, and temperature during the battery cycle aging experiment. Batteries B05 and B07 were selected as experimental subjects, and their detailed data were recorded in three working modes (charging, discharging, and impedance testing) at a temperature of 24°C. The rated capacity of batteries B05 and B07 is 2Ah. The charging process is carried out in constant current (CC) mode at 1.5A until the battery voltage reaches 4.2V, and then the constant voltage (CV) mode is used for charging until the charging current drops to 20mA. The discharge process is carried out in 2A constant current (CC) mode until the voltage of batteries B05 and B07 drops to 2.7V and 2.2V, respectively.

[0089] Step 2: Feature Extraction

[0090] As the number of cycles increases, the voltage and current curves change regularly. As the battery ages, the duration within the same voltage range gradually shortens; the duration within the same current range gradually lengthens; and after charging for the same period of time starting from a specific voltage and current value, the voltage and current values become higher. To this end, this embodiment selects the following four health characteristics for SOH estimation: the post-charge voltage value, the post-charge current value, the charging duration of the voltage data segment, and the charging duration of the current data segment.

[0091] Step 2.1: Extract the voltage data fragment during the charging process [V a ,V b The charging duration of ] is the first health feature F1. The feature F1 is calculated using formula (1):

[0092] F1=t b -t a (1)

[0093] Where, t a As the voltage rises to V a The corresponding charging time, t b As the voltage rises to V b The corresponding charging time.

[0094] The present invention selects the optimal segment of F1 by traversal method, taking 3.6V as the voltage lower limit, 4.2V as the voltage upper limit, and 0.01V as the change interval for traversal. It can be obtained from the traversal that V a =3.7V, V b =4.2V.

[0095] Step 2.2: Extract the current data fragment during the charging process [I c ,I dThe charging duration of ] is the second health feature F2. The feature F2 is calculated using formula (2):

[0096] F2=t d -t c (2)

[0097] Where, t c As the current drops to I c The corresponding charging time, t d As the current drops to I d The corresponding charging time.

[0098] The present invention selects the optimal segment of F2 by traversal method, taking 0.3A as the lower limit of current, 1.5A as the upper limit of current, and 0.01A as the variation interval for traversal. It can be obtained from the traversal that I c =1.49A, I d =0.42A.

[0099] Step 2.3: Extract the constant current (CC) charging process from the voltage V e As the starting point of the timing, charging t e The voltage value after the time is taken as the third health feature F3. The feature F3 is calculated using formula (3):

[0100] F3=V e +ΔV e (3)

[0101] Where, ΔV e t e The voltage value that changes during the charging time.

[0102] The present invention selects the optimal segment of F3 by traversal method, with 3.6V as the voltage lower limit, 4.2V as the voltage upper limit, 0.01V as the voltage change interval, and 50s as the time change interval. It can be obtained from the traversal that V e =3.72V, t e =400s.

[0103] The changing trend of the four initial extracted features after normalization is as follows: Figure 1 shown.

[0104] Step 2.4: Extract the constant voltage (CV) charging process from the current I f As the starting point of the timing, charging t f The current value after the time is taken as the fourth health feature F4. The feature F4 is calculated using formula (4):

[0105] F4=I f +ΔI f (4)

[0106] Among them, ΔIf t f The current value changes during the charging time.

[0107] The present invention selects the optimal segment of F4 by traversal method, with 0.3A as the current lower limit, 1.5A as the current upper limit, 0.01A as the current change interval, and 50s as the time change interval. It can be obtained from the traversal that I f =1.41V, t f =1800s.

[0108] Step 3: Feature processing

[0109] Step 3.1: Calculate the correlation between the four extracted features and the target value SOH using formula (5):

[0110]

[0111] Where X and Y represent the health feature and SOH samples, respectively. A larger absolute value of the correlation coefficient indicates a stronger correlation between the feature and SOH, and a higher accuracy in estimating SOH using that feature.

[0112] Step 3.2: Use Savitzky-Golay filter to construct feature processing model, which is determined by polynomial order N and window length L = 2M + 1, where M represents the absolute value of the coordinates of the maximum and minimum measurement points in the measurement window. For data with each measurement point (-M, -M + 1, ..., 0, 1, ..., M-1, M), use equations (6), (7) and (8) for processing:

[0113]

[0114]

[0115]

[0116] Where p(n) is the limiting polynomial, a k represents the coefficient of the polynomial term, ε N is the residual of the least squares fit, x[n] is the true value of the data set in the nth sequence, and the system of equations derived from (8) is solved to finally obtain the optimal polynomial coefficients.

[0117] Step 3.3: Set the threshold to 0.97. When the Pearson correlation coefficient between the feature quantity and SOH is greater than the set threshold, the feature processing model will not be activated, and the original aging feature will be used directly to estimate SOH. When the Pearson correlation coefficient between the feature quantity and SOH is less than the set threshold, it indicates that the feature is greatly affected by sensor drift noise. The feature should be processed to improve the correlation between the feature and the target quantity.

[0118] Features F2 and F4 extracted from charging current data have lower correlation than features extracted from voltage data, and their changing trends fluctuate significantly. This is because charging current data is susceptible to current sensor drift noise, and such fluctuations are unavoidable. Therefore, the present invention employs a Savitzky-Golay (SG) filtering algorithm to remove noise from features F2 and F4. To visualize the effects of feature processing, Figures 2(a) and 2(b) show a comparison of feature changing trends before and after Savitzky-Golay filtering.

[0119] Step 4: Use the four health features obtained in the above steps as input and the battery SOH as output to establish a training set and a prediction set.

[0120] Step 5: Construct the SSA-SVR model.

[0121] Step 5.1: Build a baseline SVR model

[0122] S={x i ,y i |x i ∈R n ,y i ∈R};(i=1,2,…T) (9)

[0123] Where x i is the feature vector of the i-th sample, y i is the corresponding regression value, T is the number of samples, and n is the dimension of the feature vector.

[0124] The SVR function is defined as:

[0125] f(x)=ωΦ(x)+b (10)

[0126] Where f(x) is the output, Φ(x) is the nonlinear mapping function, and ω and b are the parameters to be determined. Minimize the following objective function to solve for ω and b:

[0127]

[0128] Where C is the penalty coefficient, f(x i ) is the predicted value of the i-th sample, ε represents the maximum error allowed by regression, which is defined as:

[0129] |yf(x)| ε =max{0,|yf(x)|-ε} (12)

[0130] Introducing slack variables ξ i and After that, it can be transformed into the following objective function:

[0131]

[0132] constraint:

[0133]

[0134] Transform Equation (10) into solving the dual problem:

[0135]

[0136] Where, β i and is the Lagrangian operator, K(x i ,x j ) is the kernel function. The present invention selects the RBF kernel function with strong linear approximation ability, which is defined as:

[0137]

[0138] Where σ is the width of the kernel function.

[0139] Step 5.2: Use the sparrow search algorithm to optimize the parameters of the SVR model.

[0140] The penalty factor C and kernel function parameter σ are key parameters of the SVR model, determining the estimation accuracy and fitting ability of the model. SSA is a novel nature-inspired algorithm based on the foraging and predator avoidance behaviors of sparrows. Sparrows in SSA are able to converge to the current optimal solution by jumping to nearby locations, demonstrating excellent performance in terms of accuracy and convergence speed. Using the sparrow search algorithm to optimize hyperparameters can effectively improve the model's estimation performance. The steps of SSA-SVR are briefly described below:

[0141] Step 5.2.1: Use formula (17) to normalize the sample data. The normalized data can be divided into a training set and a test set.

[0142]

[0143] Where Y and Y' are the sample feature values before and after normalization of the training data, respectively. min and Y max Respectively represent the minimum and maximum values of the sample feature data before normalization.

[0144] Step 5.2.2: Set the parameters of the sparrow search algorithm: maximum number of iterations N = 150, population size n = 50, number of discoverers PD = 0.6, number of sentinels SD = 0.3, safety value ST = 0.5, and upper and lower limits of the independent variable DL = [-7, 7];

[0145] Step 5.2.3, set the value range of the penalty coefficient C and the kernel function parameter σ, and initialize the population;

[0146] Step 5.2.4: Use the above training set to train the SVR model and use formula (18) to calculate the fitness value of each sparrow:

[0147]

[0148] Where, is the predicted value of the nth training sample, Y n is the true value of the nth training sample, N tr is the number of training samples.

[0149] Step 5.2.5: Calculate and update the positions of the discoverer, joiner, and alerter according to equations (19)-(21).

[0150] The formula for updating the discoverer's position is:

[0151]

[0152] Where, represents the unit position of the sparrow, i is a positive integer from 1 to P, j is a positive integer from 1 to d, P is the total population, d is the dimension of the feature data, α is a random number between 0 and 1, it max is the maximum number of iterations, R2 is the warning value, ST is the safety value, Q is a random number, and E is a 1×d matrix of all ones. When R2 ≥ ST, it means that the risk of being preyed upon is high, and the sparrows at this location will move quickly; otherwise, the search range can be expanded.

[0153] When the discoverer finds better food, another joiner will come to grab it. If successful, the discoverer's position will change, otherwise the joiner's position will change.

[0154] The update formula for the joiner position is:

[0155]

[0156] Where, and Represent the most unfavorable and most favorable positions globally, A + It is the generalized inverse matrix of A, where A represents a 1×d matrix, in which the elements are randomly preset to 1 or -1; when i>n / 2, it means that the follower fails to grab food and needs to transfer.

[0157] The update formula of the sentinel position is:

[0158]

[0159] Where, represents the most advantageous position of the authorities; γ is a standard normal random number, representing the step length control coefficient; k is a random number between -1 and 1, representing the direction of the sparrow; f g represents the maximum fitness of the authority, f w and f g Instead, ε is a constant to ensure that the denominator is not zero; f i Represents the fitness value of the i-th sparrow.

[0160] Step 5.2.6: Get the current updated position and calculate the optimal individual and the optimal fitness value.

[0161] Step 5.2.7: When the training result reaches the set parameters, stop the calculation and execute the output parameters (C, σ). Otherwise, return to step 5.2.4 to recalculate the fitness value.

[0162] Step 5.2.8. Establish an estimation model based on the parameters (C, σ) and output the estimated SOH.

[0163] The modeling flow chart of the SSA-SVR model is as follows Figure 3 shown.

[0164] In order to verify the superiority of the method proposed in the present invention, the following three methods are selected for comparison in combination with the actual SOH. Method 1: is the method proposed in the present invention, including SG filtering to optimize features, and using SSA to optimize the hyperparameters of SVR. Method 2: includes SG filtering, but does not use SSA to optimize the hyperparameters of SVR. Method 3: uses SSA to optimize hyperparameters, but does not include SG filtering. The existence of method 2 is to verify the superiority of using SSA to optimize hyperparameters, and the existence of method 3 is to verify the effectiveness of the method of using SG filtering to optimize features. The present invention uses two batteries, B0005 and B0007, in the NASA data set for experiments. 60% of the battery's full life cycle is used as the training set, and the rest is used for testing. The prediction starting point for B0005 and B0007 batteries is 101. The comparison of prediction results of different methods for B0005 and B0007 batteries is shown in the figure below. Figure 4 and Figure 5 shown.

[0165] In summary, this paper estimates the state of health (SOH) of lithium-ion batteries based on the SSA-SVR model. The four features extracted from the charging voltage and current curve segments show a high correlation with the SOH, laying the foundation for the model's high-precision predictions. SG filtering is used to optimize the two noise-influenced features extracted from the charging current segments. Experimental results show that this method can reduce model error by over 50%. Using SSA to globally optimize the SVR model parameters also reduces error by over 50%, effectively improving the model's accuracy and generalization capabilities.

[0166] It should be understood that the above-described specific embodiments of the present invention are merely illustrative or illustrative of the principles of the present invention and do not constitute limitations of the present invention. Therefore, any modifications, equivalent substitutions, improvements, etc. made without departing from the spirit and scope of the present invention should be included within the scope of protection of the present invention. In addition, the appended claims are intended to cover all variations and modifications that fall within the scope and metes and bounds of the appended claims, or equivalents thereof.

Claims

1. A lithium-ion battery SOH estimation method based on the SSA-SVR model, characterized in that: include: Step 1, data collection: charge and discharge the lithium-ion battery multiple times, and record the current, voltage, time data of the lithium battery during the charge and discharge process, as well as the capacity of each complete discharge; Step 2: Extract health features: Extract four health features from the voltage and current data segments during the charging process, namely, the voltage value after charging, the current value after charging, the charging duration of the voltage data segment, and the charging duration of the current data segment; Step 3: Construct a feature processing model based on Savitzky-Golay filtering: Select the Savitzky-Golay filtering algorithm to process the extracted health features, eliminate the chaotic fluctuations of the health features caused by the influence of sensor drift noise on the charging data, and construct a feature processing model; Step 4: Using the matrix F = [F1, F2, F3, F4] consisting of the extracted and processed health features as input and the battery SOH as output, construct a training set and a prediction set; Step 5: Build an SSA-SVR model: Build a baseline SVR model. Based on the above training set, use the sparrow search algorithm to optimize the hyperparameters of the SVR model, build an SOH estimation model, and output the estimated SOH.

2. The SOH estimation method for lithium-ion batteries based on the SSA-SVR model according to claim 1, characterized in that: The step 2 includes: Step 2.1: Extract the voltage data fragment during the charging process [V a ,V b The charging duration of ] is the first health feature F1, and the feature F1 is calculated using formula (1): F1=t b -t a (1) Where, t a As the voltage rises to V a The corresponding charging time, t b As the voltage rises to V b The corresponding charging time.

3. The SOH estimation method for lithium-ion batteries based on the SSA-SVR model according to claim 2, characterized in that: The step 2 further comprises: Step 2.2: Extract the current data fragment during the charging process [I c ,I d The charging duration of ] is the second health feature F2, and the feature F2 is calculated using formula (2): F2=t d -t c (2) Where, t c As the current drops to I c The corresponding charging time, t d As the current drops to I d The corresponding charging time.

4. The method for estimating SOH of a lithium-ion battery based on the SSA-SVR model according to claim 3, characterized in that: The step 2 further comprises: Step 2.3: Extract the constant current charging process from the voltage V e As the starting point of the timing, charging t e The voltage value after the time is taken as the third health feature F3, and the feature F3 is calculated using formula (3): F3=V e +ΔV e (3) Where, ΔV e t e The voltage value that changes during the charging time.

5. The method for estimating SOH of a lithium-ion battery based on the SSA-SVR model according to claim 4, characterized in that: The step 2 further comprises: Step 2.4: Extract the constant voltage charging process from the current I f As the starting point of the timing, charging t f The current value after the time is taken as the fourth health feature F4, and the feature F4 is calculated using formula (4): F4=I f +ΔI f (4) Among them, ΔI f t f The current value changes during the charging time.

6. The method for estimating SOH of a lithium-ion battery based on the SSA-SVR model according to claim 1, characterized in that: The step 3 includes: Step 3.1: Calculate the correlation between the four extracted health characteristics and the target value SOH using formula (5): Where X and Y are health characteristics and SOH samples, respectively.

7. The method for estimating SOH of a lithium-ion battery based on the SSA-SVR model according to claim 6, characterized in that: The step 3 further comprises: Step 3.2: Use Savitzky-Golay filter to construct feature processing model, which is determined by polynomial order N and window length L = 2M + 1, where M represents the absolute value of the coordinates of the maximum and minimum measurement points in the measurement window. For data with each measurement point (-M, -M + 1, ..., 0, 1, ..., M-1, M), use equations (6), (7) and (8) for processing: Where p(n) is the limiting polynomial, a k represents the coefficient of the polynomial term, ε N is the residual of the least squares fit, x[n] is the true value of the data set in the nth sequence, and the system of equations derived from (8) is solved to finally obtain the optimal polynomial coefficients.

8. The method for estimating SOH of a lithium-ion battery based on the SSA-SVR model according to claim 7, characterized in that: The step 3 further comprises: Step 3.3, set the threshold to 0.

97. When the Pearson correlation coefficient between the health feature and SOH is greater than the set threshold, the feature processing model will not be activated, and the original aging feature will be used directly to estimate SOH. When the Pearson correlation coefficient between the health feature and SOH is less than the set threshold, it indicates that the health feature is greatly affected by sensor drift noise. The health feature should be processed to improve the correlation between the health feature and the target quantity.

9. The method for estimating SOH of a lithium-ion battery based on the SSA-SVR model according to claim 1, characterized in that: The step 5 comprises: Step 5.1: Build a baseline SVR model S={x i ,y i |x i ∈R m ,y i ∈R};i=1,2,…,T (9) Where x i is the feature vector of the i-th sample, y i is the corresponding regression value, T is the number of samples, and m is the dimension of the feature vector; The SVR function is defined as: f(x)=ωΦ(x)+b (10) Where f(x) is the output, Φ(x) is the nonlinear mapping function, ω and b are the parameters to be determined, and the following objective function is minimized to solve ω and b: Where C is the penalty coefficient, f(x i ) is the predicted value of the i-th sample, ε represents the maximum error allowed by regression, which is defined as: |y-f(x)| ε =max{0,|y-f(x)|-ε} (12) Introducing slack variables ξ i and Then, it is transformed into the following objective function: constraint: Transform Equation (10) into solving the dual problem: Where, β i and is the Lagrangian operator, K(x i ,x j ) is the kernel function, and the RBF kernel function with strong linear approximation ability is selected, which is defined as: Where σ is the width of the kernel function; Step 5.2: Use the sparrow search algorithm to optimize the parameters of the SVR model, build the SOH estimation model, and output the estimated SOH.

10. The method for estimating SOH of a lithium-ion battery based on the SSA-SVR model according to claim 9, characterized in that: Said step 5.2 comprises: Step 5.2.1: Use formula (17) to normalize the sample data. The normalized data is divided into training set and test set. Where Y and Y' are the sample feature values before and after normalization of the training data, respectively. min and Y max Respectively represent the minimum and maximum values of the sample feature data before normalization; Step 5.2.2, set the parameters of the sparrow search algorithm; Step 5.2.3, set the value range of the penalty coefficient C and the kernel function parameter σ, and initialize the population; Step 5.2.4: Use the above training set to train the SVR model and use formula (18) to calculate the fitness value of each sparrow: Where, is the predicted value of the nth training sample, Y n is the true value of the nth training sample, N tr is the number of training samples; Step 5.2.5: Calculate and update the positions of the discoverer, joiner, and alerter according to equations (19)-(21) respectively; The formula for updating the discoverer's position is: Where, represents the position of a sparrow, i is a positive integer from 1 to P, j is a positive integer from 1 to d, P is the total population, d is the dimension of the feature data, α is a random number between 0 and 1, it max is the maximum number of iterations, R2 is the warning value, ST is the safety value, Q is a random number, and E is a 1×d matrix of all 1s. When R2 ≥ ST, it means that the risk of being preyed on is high, and the sparrow at this location will move quickly; otherwise, the search range can be expanded. The update formula for the joiner position is: Where, and Represent the most unfavorable and most favorable positions globally, A + is the generalized inverse matrix of A, where A represents a 1×d matrix, where the elements are randomly preset to 1 or -1; when i>n / 2, it means that the follower fails to grab the food and needs to move; The update formula of the sentinel position is: Where, represents the most advantageous position of the authorities; γ is a standard normal random number, representing the step length control coefficient; k is a random number between -1 and 1, representing the direction of the sparrow; f g represents the maximum fitness of the authority, f w and f g Instead, ε is a constant to ensure that the denominator is not zero; f i represents the fitness value of the i-th sparrow; Step 5.2.6: Get the current updated position and calculate the optimal individual and the best fitness value; Step 5.2.7: When the training result reaches the set parameters, stop the calculation and execute the output parameters (C, σ). Otherwise, return to step 5.2.4 to recalculate the fitness value. Step 5.2.

8. Establish an estimation model based on the parameters (C, σ) and output the estimated SOH.

Images