Battery state of health prediction method, apparatus, device, and storage medium

Abstract

The application provides a battery state of health prediction method, device, equipment and storage medium, including: obtaining charging voltage data of a to-be-tested battery; extracting a health index for indicating the state of health of the battery from the charging voltage data located in a target voltage interval; the health index includes at least two of the following: equal pressure difference charging duration, equal pressure difference charging energy, voltage average, power spectral density value, Euclidean distance and Manhattan distance; the health index is fused in a weighted fusion manner to obtain a fused health index; the fused health index is input into a trained prediction model to obtain a battery state of health prediction result corresponding to the to-be-tested battery; the prediction model is constructed by integrating learning framework and weighting and combining multiple support vector machine models optimized by a swarm intelligence optimization algorithm; thus, the battery state of health can be predicted in real time, with high precision and high efficiency.

Description

Technical Field

[0001] This application relates to the field of battery technology, specifically to a method, apparatus, computing device, and computer-readable storage medium for predicting battery health status. Background Technology

[0002] Accurate prediction of the State of Health (SOH) of lithium-ion batteries is one of the core functions of battery management systems. Due to the complex electrochemical characteristics and nonlinear kinetic processes of batteries, the state of health cannot be directly measured and can only be estimated through external observation signals such as voltage and current. Therefore, state of health prediction has become a key technical challenge in the field of battery management.

[0003] Battery health state prediction methods can be broadly categorized into model-driven and data-driven methods. Model-driven methods include electrochemical models and equivalent circuit models. Electrochemical models use partial differential equations to describe the internal electrochemical processes of the battery, while equivalent circuit models abstract the battery into components such as resistors and capacitors, reassembling them into a circuit to obtain internal and external characteristics consistent with the battery. However, the prediction accuracy of model-driven methods is easily affected by the optimization of model parameters. In recent years, with the rise of artificial intelligence and big data technologies, data-driven methods such as neural networks, support vector machines, and deep learning have been increasingly applied to battery health state prediction. These methods can automatically learn nonlinear mapping relationships from experimental data and are highly adaptable.

[0004] However, existing battery state of health prediction methods suffer from insufficient accuracy and low efficiency due to neglecting the impact of capacity changes in the early stages of battery life on the full-cycle state of health (SOH) prediction. Therefore, there is an urgent need for a battery state of health prediction method that can achieve high accuracy and high efficiency. Summary of the Invention

[0005] To address the aforementioned technical problems, this application provides a battery health state prediction method, apparatus, computing device, and computer-readable storage medium, which can achieve real-time, high-precision, and high-efficiency prediction of battery health state.

[0006] This application provides a method for predicting battery health status, including:

[0007] Step S1: Obtain the charging voltage data of the battery under test;

[0008] Step S2: Extract health indicators to indicate the battery health status from the charging voltage data located in the target voltage range; the health indicators include at least two of the following: equal-voltage charging time, equal-voltage charging energy, average voltage, power spectral density value, Euclidean distance, and Manhattan distance;

[0009] Step S3: The health indicators are fused using a weighted fusion method to obtain fused health indicators; wherein, the weight of each health indicator is determined based on its correlation with the battery health status.

[0010] Step S4: Input the fused health indicators into the trained prediction model to obtain the battery health status prediction result corresponding to the battery under test; the prediction model is constructed by weighted combination of multiple support vector machine models optimized by swarm intelligence optimization algorithm through an ensemble learning framework.

[0011] In one embodiment, step S1 is to acquire the charging voltage data of the battery under test;

[0012] Step S2: Extract health indicators to indicate the battery health status from the charging voltage data located in the target voltage range; the health indicators include at least two of the following: equal-voltage charging time, equal-voltage charging energy, average voltage, power spectral density value, Euclidean distance, and Manhattan distance;

[0013] Step S3: The health indicators are fused using a weighted fusion method to obtain fused health indicators; wherein, the weight of each health indicator is determined based on its correlation with the battery health status.

[0014] Step S4: Input the fused health indicators into the trained prediction model to obtain the battery health status prediction result corresponding to the battery under test; the prediction model is constructed by weighted combination of multiple support vector machine models optimized by swarm intelligence optimization algorithm through an ensemble learning framework.

[0015] In one embodiment, the historical charging voltage data in all training samples are analyzed based on a dual correlation analysis method to determine the target voltage range, including:

[0016] Within a preset charging voltage range, multiple candidate voltage ranges are constructed using different combinations of start and stop voltages;

[0017] For each candidate voltage range, the corresponding equal-voltage charging time sequence is extracted from all training samples, and the Pearson correlation coefficient and Spearman correlation coefficient between the equal-voltage charging time sequence and the historical battery health state sequence corresponding to all training samples are calculated.

[0018] The correlation coefficient and data sampling duration of each candidate voltage range are comprehensively evaluated, and the candidate voltage range that meets the preset performance conditions is selected as the target voltage range. The data sampling duration of each candidate voltage range is determined based on the equal voltage difference charging duration sequence corresponding to that candidate voltage range.

[0019] In one implementation, a method combining entropy weights and correlation factors is used to fuse the health indicators corresponding to each training sample, resulting in a fused health indicator for each training sample, including:

[0020] For each training sample, calculate the health indicator set X = [X1, X2, ... X]. i …,X m The i-th health indicator X in ] i The Pearson correlation coefficient γ between the battery health status sample and the corresponding battery health status sample i The correlation coefficient ρ with Spearman i ; m represents the number of health indicators;

[0021] According to the formula Calculate health index X respectively i Information entropy E i ; where X' ij ε is the normalized value of the i-th health indicator in the j-th training sample, where n is the number of training samples and ε approaches zero.

[0022] According to the formula and Determine health indicator X i weight w i ;

[0023] According to the formula Calculate the Fusion Health Index (IHF) for each training sample.

[0024] In one embodiment, a method combining entropy weights and correlation factors is used to fuse the health indicators corresponding to each training sample, resulting in a fused health indicator for each training sample, including:

[0025] For each training sample, calculate the health indicator set X = [X1, X2, ... X]. i …,X m The i-th health indicator X in ] i The Pearson correlation coefficient γ between the battery health status sample and the corresponding battery health status sample i The correlation coefficient ρ with Spearman i ; m represents the number of health indicators;

[0026] According to the formula Calculate health index X respectively i Information entropy E i ; where X' ij ε is the normalized value of the i-th health indicator in the j-th training sample, where n is the number of training samples and ε approaches zero.

[0027] According to the formula and Determine health indicator X i weight w i ;

[0028] According to the formula Calculate the Fusion Health Index (IHF) for each training sample.

[0029] In one embodiment, the swarm intelligence optimization algorithm is the Grey Wolf Optimization Algorithm, the support vector machine model is the Least Squares Support Vector Machine Model, and the ensemble learning algorithm is the AdaBoost Algorithm.

[0030] In one embodiment, the formulas for calculating the equal-pressure differential charging duration (HF1) and the equal-pressure differential charging energy (HF2) are as follows:

[0031] HF1=T CDVC (k) = t2(k) - t1(k)

[0032]

[0033] Where k is the number of cycles, t1(k) and t2(k) correspond to the start and end times of the selected voltage segment in each cycle, respectively, and U is the voltage.

[0034] The formulas for calculating the voltage mean characteristic (HF3) and power spectral density characteristic (HF4) are as follows:

[0035]

[0036] Where w is the number of voltage sampling points collected in each cycle within the voltage range, and V p (k) represents the voltage mean characteristic of the k-th cycle, V i S is the voltage value at the i-th sampling point. V (f) is the power spectral density at frequency f;

[0037] The formulas for calculating Euclidean distance (HF5) and Manhattan distance (HF6) are as follows:

[0038]

[0039] HF6=D M (k)=|a1-b1|+|a2-b2|+|a i -b i |

[0040] Among them, a i and b i These represent the voltage data points of the reference voltage curve and the voltage curve of the kth cycle, respectively.

[0041] This application also provides a battery health status prediction device, including:

[0042] The acquisition module is used to acquire the charging voltage data of the battery under test;

[0043] The extraction module is used to extract health indicators that indicate the health status of the battery from the charging voltage data located in the target voltage range; the health indicators include at least two of the following: equal-voltage charging time, equal-voltage charging energy, average voltage, power spectral density value, Euclidean distance, and Manhattan distance;

[0044] The fusion module is used to fuse health indicators using a weighted fusion method to obtain fused health indicators; wherein, the weight of each health indicator is determined based on its correlation with the battery health status;

[0045] The processing module is used to input the fused health indicators into the trained prediction model to obtain the health status prediction result of the battery to be tested; the prediction model is constructed by weighted combination of multiple support vector machine models optimized by swarm intelligence optimization algorithm through an ensemble learning framework.

[0046] This application also provides a computing device, including: a memory and a processor, wherein the memory stores computer program instructions for execution on the processor, and when the processor executes the computer program instructions, it implements the battery health state prediction method as described above.

[0047] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the battery health state prediction method as described above.

[0048] As described above, the battery health state prediction method, apparatus, computing device, and computer-readable storage medium of this application organically combine the selection of charging voltage data within a preset voltage range, the extraction and fusion of multiple health indicators, and a high-performance hybrid prediction model. This enables collaborative prediction of battery health state with high precision, high efficiency, and high robustness, solving the technical problems of insufficient precision, computational complexity, and neglect of early changes in existing battery health state prediction methods. It can achieve real-time, high-precision, and high-efficiency prediction of battery health state. Attached Figure Description

[0049] Figure 1 A flowchart illustrating a battery health status prediction method provided in an embodiment of this application;

[0050] Figure 2 A schematic diagram illustrating the implementation process of a battery health state prediction method provided in this application embodiment;

[0051] Figure 3 This is a schematic diagram of the full-cycle SOH degradation curve of the battery in the embodiments of this application;

[0052] Figure 4 This is a schematic diagram showing the distribution of Pearson coefficients and Spearman coefficients in different voltage ranges in the embodiments of this application;

[0053] Figure 5 A schematic diagram comparing the SOH prediction performance and error of different models provided in the embodiments of this application;

[0054] Figure 6 This is a schematic diagram of the battery health state prediction device provided in an embodiment of this application. Detailed Implementation

[0055] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. It should be understood that the specific embodiments described herein are merely illustrative and are not intended to limit the scope of this application.

[0056] The following is a brief introduction to the terminology and algorithm implementation principles involved in this application:

[0057] 1. Battery Health Status: Battery health status is called SOH (State of Health). This indicator reflects the battery's health status and performance degradation. Currently, the definition of battery SOH is reflected in several aspects, including capacity, charge, internal resistance, number of cycles, and peak power. It is usually defined by capacity decay, which is the ratio of the maximum capacity that the battery can currently release to the battery's rated capacity, as shown in the following formula:

[0058]

[0059] In the formula C aged C represents the current capacity of the battery. rated This refers to the battery's rated capacity.

[0060] 2. Gray Wolf Optimization Algorithm

[0061] Grey Wolf optimization (GWO) is a swarm intelligence optimization algorithm that simulates the social hierarchy and hunting behavior of grey wolves. It was proposed by Mirjalili et al. in 2014. It seeks the optimal solution to a problem by simulating the leadership hierarchy and cooperative hunting mechanisms within a wolf pack.

[0062] Its core idea lies in dividing wolf packs into four levels and optimizing them through hunting behavior. The four levels are:

[0063] α (alpha wolf): the optimal solution.

[0064] β (suboptimal wolf): suboptimal solution.

[0065] δ(Ordinary Wolf): Third best solution.

[0066] ω(bottom wolf): other candidate solutions.

[0067] The hunting behavior (optimization process) includes:

[0068] Encirclement: The wolf pack gradually surrounds the prey (potential optimal solution).

[0069] Hunting: The α, β, and δ wolves guide the entire wolf pack (population) to move toward the prey's location (global optimal area).

[0070] Attack: Ultimately capture the prey (converge to the optimal solution).

[0071] 3. LSSVM Algorithm

[0072] LSSVM (Least Squares Support Vector Machine) is an important variant of the standard Support Vector Machine (SVM), proposed by Suykens and Vandewalle in 1999. It significantly reduces computational complexity by modifying the SVM's optimization objective function, transforming the inequality constraint problem into solving a system of linear equations.

[0073] Compared to SVM, LSSVM uses the Sum Squared Error Loss function, transforming the optimization problem into solving a system of linear equations, which greatly improves computational efficiency. At the same time, LSSVM changes the constraints to equality constraints, further transforming the optimization problem into a linear one.

[0074] 4. AdaBoost Algorithm

[0075] AdaBoost, short for Adaptive Boosting, is a powerful and efficient ensemble learning algorithm proposed by Yoav Freund and Robert Schapire in 1995. Its core idea is to combine multiple weak, simple learners to form a strong, accurate learner.

[0076] The core idea and working principle of the AdaBoost algorithm are as follows:

[0077] Sequential training: AdaBoost does not train all weak learners simultaneously, but trains them sequentially.

[0078] Focus on errors: In each round of training, it pays more attention (gives higher weight) to samples that were misclassified or mispredicted by the previous weak learner.

[0079] Adaptive adjustment: Through this "error-focused" mechanism, the algorithm adaptively adjusts the distribution of training data, enabling subsequent weak learners to focus on learning the most difficult samples to classify.

[0080] Weighted voting: After training, all weak learners are combined into the final strong learner through weighted voting (for classification problems) or weighted averaging (for regression problems). The weak learner with higher accuracy has greater say (weight) in the voting.

[0081] like Figure 1 As shown, this embodiment provides a battery health status prediction method, which can be used to predict the health status of various batteries such as lithium-ion batteries, and includes the following steps:

[0082] Step S1: Obtain the charging voltage data of the battery under test.

[0083] Step S2: Extract health indicators to indicate the battery health status from the charging voltage data located in the target voltage range; the health indicators include at least two of the following: equal-voltage charging time, equal-voltage charging energy, average voltage, power spectral density value, Euclidean distance, and Manhattan distance.

[0084] Step S3: The health indicators are fused using a weighted fusion method to obtain fused health indicators; wherein, the weight of each health indicator is determined based on its correlation with the battery health status.

[0085] Step S4: Input the fused health indicators into the trained prediction model to obtain the battery health status prediction result corresponding to the battery under test; the prediction model is constructed by weighted combination of multiple support vector machine models optimized by swarm intelligence optimization algorithm through an ensemble learning framework.

[0086] Step S5: Obtain the health status of the nickel-cadmium battery under test in the next charge-discharge cycle based on the predicted battery capacity value.

[0087] The battery under test can be a lithium-ion battery, etc., without specific limitations. The charging voltage data of the battery under test can be understood as the charging voltage data of the most recent charging cycle or the current charging cycle. Specifically, it can be time-series voltage data sampled at equal time intervals within a specific voltage range during the constant current charging phase, and this specific voltage range covers the target voltage range. The target voltage range refers to an optimal sub-range determined from the battery charging voltage curve. Within this optimal sub-range, the equal-voltage differential charging time shows the strongest and most stable correlation with the degradation of battery health, while also enabling rapid data acquisition. Therefore, it becomes the optimal voltage window for extracting health indicators for battery health prediction. Thus, extracting health indicators only from the charging voltage data within the target voltage range, and then predicting battery health based on these extracted health indicators, can maximize the linear and monotonic correlation between the extracted health indicators and battery aging, ensuring the sensitivity of the health indicators from the outset. Charging voltage data within the target voltage range refers to charging voltage data that falls within the target voltage range. For example, if the target voltage range is [3.86V, 4.08V], and the voltage range corresponding to the charging voltage data of the battery under test is [3.06V, 4.20V], then charging voltage data within the target voltage range refers to charging voltage data of the battery under test that falls within [3.86V, 4.08V].

[0088] In this process, health indicators (also known as health features) indicative of battery health status can be extracted from charging voltage data within the target voltage range using one or more methods, such as time-domain analysis, frequency-domain analysis, or similarity comparison. These extracted health indicators effectively reflect subtle degradation trends in battery health, particularly the impact of capacity changes in the early stages of battery life. After determining the health indicators for the charging voltage data of the battery under test, a weighted fusion method can be used to fuse these indicators, resulting in a fused health indicator. The weight of each health indicator is determined based on its correlation with the battery health status, specifically determined when training the prediction model using the training sample set. By comprehensively capturing battery aging information—that is, extracting multiple health indicators and then fusing them based on their weights—the resulting fused health indicator is more representative and stable than any single health indicator, effectively improving the quality of the model's input data. Fusion methods used for fusing health indicators include, but are not limited to, weighted average, neural networks, and principal component analysis. The fused health indicator is a scalar value representing the battery health status obtained by fusing health indicators.

[0089] In this process, after determining the fused health index of the charging voltage data of the battery under test, the fused health index is input into the trained prediction model, which then outputs the predicted battery health status of the battery under test. Because swarm intelligence optimization algorithms optimize the parameters of the support vector machine model, they avoid the blindness of human selection and enable the model to achieve optimal performance. Swarm intelligence optimization algorithms include, but are not limited to, GWO, PSO (Particle Swarm Optimization), and GA (Genetic Algorithm). Ensemble learning frameworks can be understood as ensemble learning algorithms, including but not limited to AdaBoost, Gradient Boosting, Random Forest, and Stacking. Taking AdaBoost as an example, the AdaBoost algorithm integrates multiple weak predictors (also called weak learners) to form a strong predictor, further enhancing the model's generalization ability and prediction accuracy. The battery health status prediction result can include information such as the battery health status value. Of course, the battery health status prediction result can also only include information such as the predicted current battery capacity of the battery under test. Furthermore,

[0090] In summary, the battery health status prediction method provided in the above embodiments organically combines the selection of charging voltage data within the optimal voltage range (i.e., the target voltage range), the extraction and fusion of multiple health indicators, and a high-performance hybrid prediction model. This enables collaborative prediction of battery health status with high accuracy, high efficiency, and high robustness, solving the technical problems of insufficient accuracy, computational complexity, and neglect of early changes in existing battery health status prediction methods. It can achieve real-time, high-accuracy, and high-efficiency prediction of battery health status.

[0091] In one embodiment, before step S1, the following steps are included:

[0092] Obtain a training sample set containing multiple training samples, each training sample including a charging voltage data sample and a corresponding battery health status sample;

[0093] The charging voltage data samples in all training samples are analyzed based on the bicorrelation analysis method to determine the target voltage range;

[0094] For each training sample, health indicators are extracted from the charging voltage data samples located in the target voltage range.

[0095] A method combining entropy weight and correlation factor is used to fuse the health indicators corresponding to each training sample to obtain the fused health indicator for each training sample.

[0096] The prediction model is trained based on the fused health indicators of all training samples and the corresponding battery health status samples.

[0097] The training samples can be generated based on historical data of the battery under test, or based on historical data of other batteries of the same model as the battery under test. For each training sample (i.e., each battery cycle), a corresponding fused health metric can be calculated. Furthermore, after obtaining the training sample set, preprocessing operations such as denoising, interpolation, or normalization can be performed on the training samples.

[0098] In one embodiment, the historical charging voltage data in all training samples are analyzed based on a dual correlation analysis method to determine the target voltage range, including:

[0099] Within a preset charging voltage range, multiple candidate voltage ranges are constructed using different combinations of start and stop voltages;

[0100] For each candidate voltage range, the corresponding equal-voltage charging time sequence is extracted from all training samples, and the Pearson correlation coefficient and Spearman correlation coefficient between the equal-voltage charging time sequence and the historical battery health state sequence corresponding to all training samples are calculated.

[0101] The correlation coefficient and data sampling duration of each candidate voltage range are comprehensively evaluated, and the candidate voltage range that meets the preset performance conditions is selected as the target voltage range. The data sampling duration of each candidate voltage range is determined based on the equal voltage difference charging duration sequence corresponding to that candidate voltage range.

[0102] The preset charging voltage range can be set according to actual needs, such as [3.80V, 4.20V]. Optionally, the starting voltage range of the candidate voltage interval can be set to [3.80V, 4.00V]. When constructing multiple candidate voltage intervals with different combinations of starting and cutting-off voltages, the interval can be set to 0.01V. For example, assuming the current candidate voltage interval is [3.80V, 4.20V], the next candidate voltage interval could be [3.81V, 4.20V], etc. Optionally, the cutting-off voltage of the candidate voltage intervals can be uniformly set to 4.20V.

[0103] It is understandable that for each charge-discharge cycle, the voltage-time data, i.e., the charging voltage data, of the constant current charging phase can be extracted. Then, on the voltage-time curve, the time t1(k) at which the voltage reaches the starting voltage Vstart of the voltage range and the time t2(k) at which the voltage reaches the cutoff voltage Vend of the voltage range are found. The equal-voltage differential charging duration of this charging voltage data is the difference between t2(k) and t1(k). Based on the above operations, for each candidate voltage range, the corresponding equal-voltage differential charging duration can be extracted from each training sample, thereby obtaining an equal-voltage differential charging duration sequence composed of the equal-voltage differential charging durations of all training samples. The specific calculation process of Pearson correlation coefficient and Spearman correlation coefficient can be found in existing technologies and will not be elaborated here.

[0104] In this embodiment, for each candidate voltage range, after obtaining the equal-voltage differential charging time corresponding to all training samples in that candidate voltage range, the average of the equal-voltage differential charging times corresponding to all training samples in that candidate voltage range can be determined as the data sampling time for that candidate voltage range. It can be understood that the wider the voltage range, the higher the correlation between the equal-voltage differential charging time and the state of equilibrium (SOH) may be; however, the wider the voltage range, the longer the data sampling time corresponding to that voltage range will be. In this embodiment, a candidate voltage range whose overall performance meets the preset conditions refers to a voltage range with sufficiently high correlation coefficients (i.e., a large absolute value or average absolute value of the Pearson correlation coefficient and the Spearman correlation coefficient) and sufficiently short data sampling time. For example, suppose the data sampling time for candidate voltage range A [3.80V, 4.20V] is 4 minutes, and the average absolute value of the correlation coefficient is 0.95, while the data sampling time for candidate voltage range A [3.86V, 4.08V] is 2 minutes, and the average absolute value of the correlation coefficient is 0.93. Although the correlation of range A (0.95) is slightly higher than that of range B (0.93), the data sampling time of range B (2 minutes) is much shorter than that of range A (4 minutes). That is, range B achieves a doubling of sampling efficiency with a negligible loss of accuracy (2%). Therefore, range B is selected as the candidate voltage range whose overall performance meets the preset conditions.

[0105] Under constant current charging conditions, the charging time under constant voltage difference directly reflects the battery's charging rate. A very early and persistent sign of battery aging is the increase in internal impedance and the slowdown in lithium-ion diffusion rate. These subtle changes may not yet be apparent in capacity, but they immediately and sensitively affect charging behavior, leading to longer charging times within the same voltage range. Therefore, by selecting candidate voltage ranges that meet preset performance conditions as target voltage ranges, and then extracting health indicators within these ranges, the extracted health indicators can effectively reflect the subtle degradation trends of the battery state. Simultaneously, the target voltage range determined through dual correlation analysis (Pearson + Spearman) maximizes the linear and monotonic correlation between the extracted health indicators and battery aging, ensuring the sensitivity of the features from the outset, making the selected voltage range more robust to different aging modes.

[0106] Thus, by accurately obtaining the target voltage range, the accuracy of battery health status prediction can be further improved.

[0107] In one embodiment, a method combining entropy weights and correlation factors is used to fuse the health indicators corresponding to each training sample, resulting in a fused health indicator for each training sample, including:

[0108] For each training sample, calculate the health indicator set X = [X1, X2, ... X]. i …,X m The i-th health indicator X in ] i The Pearson correlation coefficient γ between the battery health status sample and the corresponding battery health status sample i The correlation coefficient ρ with Spearman i ; m represents the number of health indicators;

[0109] According to the formula Calculate health index X respectively i Information entropy E i ; where X' ij ε is the normalized value of the i-th health indicator in the j-th training sample, where n is the number of training samples and ε approaches zero.

[0110] According to the formula and Determine health indicator X i weight w i ;

[0111] According to the formula Calculate the Fusion Health Index (IHF) for each training sample.

[0112] The weights of each health indicator are obtained through the aforementioned method. By weighted summation of the health indicators for each training sample, a fused health indicator for each training sample can be obtained. In this way, by fusing multiple health indicators into a single, highly representative indicator, the dimensionality of the model input is greatly simplified, the model's computation speed is accelerated, and the efficiency of battery health status prediction is further improved.

[0113] In one embodiment, the constructed prediction model is trained based on the fused health indicators of all training samples and the corresponding battery health status samples, including:

[0114] Initialize the population for swarm intelligence optimization algorithms;

[0115] With the goal of minimizing the battery health state prediction error, the parameters of the support vector machine model are iteratively optimized based on the fused health indicators of all training samples and the corresponding battery health state samples.

[0116] The optimized support vector machine model parameters are output to form a weak predictor;

[0117] Using an ensemble learning algorithm, each weak predictor is weighted and combined based on its prediction performance on the training sample set to obtain the final strong predictor.

[0118] The following explanation uses the Grey Wolf Optimization Algorithm as the swarm intelligence optimization algorithm, the Least Squares Support Vector Machine model as the support vector machine model, the AdaBoost algorithm as the ensemble learning algorithm, the learner as the predictor, and health features as the health metric. Figure 2 As shown. The population for the swarm intelligence optimization algorithm is initialized, with a population size of 10 and a maximum number of iterations of 10, etc. Simultaneously, the regularization parameter γ∈[10,100] and kernel parameter σ∈[100,500] are set for the least squares support vector machine model. Furthermore, for the AdaBoost algorithm, the number of weak predictors can be set to 10, and the error threshold ξ=0.01.

[0119] The fused health index is presented as an input feature in the training data. Based on the fused health index of all training samples and the corresponding battery health state samples, a fused health index sequence (labeled X_train) and a battery health state sequence (labeled y_train) are obtained. X_train and y_train are then input into a prediction model built on the GWO-LSSVM-AdaBoost algorithm. The Grey Wolf Optimization (GWO) algorithm attempts different parameter combinations, and its fitness is evaluated by using the LSSVM model with the current parameters to predict the SOH based on X_train and calculating the error (e.g., mean squared error, MSE) between the predicted value and y_train. The smaller the error, the higher the fitness. For each set of parameters proposed by GWO, the core task of the LSSVM model is to learn the mathematical function f between X_train and y_train, such that f(X_train) ≈ y_train. The AdaBoost algorithm trains multiple (e.g., 10) LSSVM models (i.e., weak predictors, also known as weak learners) optimized by GWO. It assigns a weight to each weak predictor based on its performance on the training dataset (i.e., its prediction error on X_train), and then performs a weighted combination to obtain the final strong predictor, i.e., the trained prediction model. The more accurate the LSSVM model's predictions, the higher its weight.

[0120] In the GWO-LSSVM-AdaBoost prediction model, the GWO algorithm automatically finds the optimal combination of parameters (e.g., regularization parameters and kernel function parameters) for the LSSVM model, thus avoiding the blindness of manual parameter tuning and significantly improving the accuracy and reliability of the prediction model. LSSVM acts as a "weak learner" or "base model," learning the nonlinear mapping from "fusion health indicators" to "battery SOH." Each LSSVM model optimized by GWO is considered a weak predictor. The AdaBoost algorithm integrates multiple different LSSVM models (each optimized by GWO at different stages) and combines them through weighted summation to form a more powerful and stable "strong predictor," ultimately achieving high-precision and robust battery health state prediction.

[0121] In one embodiment, the formulas for calculating the equal-pressure differential charging duration (HF1) and the equal-pressure differential charging energy (HF2) are as follows:

[0122] HF1=T CDVC (k) = t2(k) - t1(k)

[0123]

[0124] Where k is the number of cycles, t1(k) and t2(k) correspond to the start and end times of the selected voltage segment in each cycle, respectively, and U is the voltage.

[0125] The formulas for calculating the voltage mean characteristic (HF3) and power spectral density characteristic (HF4) are as follows:

[0126]

[0127] Where w is the number of voltage sampling points collected in each cycle within the voltage range, and V p (k) represents the voltage mean characteristic of the k-th cycle, V i S is the voltage value at the i-th sampling point. V (f) is the power spectral density at frequency f;

[0128] The formulas for calculating Euclidean distance (HF5) and Manhattan distance (HF6) are as follows:

[0129]

[0130] HF6=D M (k)=|a1-b1|+|a2-b2|+|a i -b i |

[0131] Among them, a i and b i These represent the voltage data points of the reference voltage curve and the voltage curve of the kth cycle, respectively.

[0132] It is understandable that for each charge-discharge cycle, the voltage-time data, i.e., the charging voltage data, of the constant current charging phase can be extracted. Then, on the voltage-time curve, the time t1(k) at which the voltage reaches the starting voltage Vstart of the voltage range and the time t2(k) at which the voltage reaches the cutoff voltage Vend of the voltage range are found. The equal-voltage differential charging duration is then the difference between t2(k) and t1(k). It should be noted that the specific calculation process for the equal-voltage differential charging duration, equal-voltage differential charging energy, average voltage, power spectral density value, Euclidean distance, and Manhattan distance in this embodiment can refer to existing technologies and will not be repeated here.

[0133] The battery health status prediction method provided in the above embodiments is verified through a specific example below, using health indicators as health features. Under experimental conditions, cycle aging test data were collected for four lithium-ion batteries (labeled CS2_35, CS2_36, CS2_37, and CS2_38) at room temperature, charged at a 0.5C rate using constant current-constant voltage (CC-CV) charging (cutoff voltage of 4.2V), and then discharged. By calculating the ratio of the actual discharge capacity to the rated capacity after each cycle, i.e., SOH, the following results were obtained: Figure 3 The full-cycle SOH degradation curves of the four batteries shown in the figure demonstrate that even under the same operating conditions, the degradation paths of different batteries differ, indicating the uncertainty and complexity of the battery aging process. Next, the charging voltage data and corresponding SOH obtained through the experiments were used as the training sample set to train a prediction model based on GWO-LSSVM-AdaBoost. Please refer to [further details omitted]. Figure 2 First, preprocessing is performed on the training sample set: outliers and noisy data are removed, missing data is filled in by linear interpolation, and the data is normalized to the range [0,1]. Next, a target voltage range is determined using a dual correlation analysis method, including: setting the starting voltage range of the voltage range to [3.80V, 4.00V], with an interval of 0.01V; setting the cutoff voltage to a uniform 4.2V; and calculating the Pearson correlation coefficient (γ) and Spearman correlation coefficient (ρ) between the equal-voltage differential charging time and SOH within different voltage ranges, such as... Figure 4 As shown, considering both correlation coefficient and sampling time, the target voltage range was ultimately determined to be [3.86V, 4.08V], with corresponding γ = 0.9386, ρ = 0.8976, and an average sampling time of 3737.99s per cycle. Next, within the determined target voltage range [3.86V, 4.08V], six health features across three categories were extracted: one category consisted of directly measured features, including equal-voltage charging time (HF1) and equal-voltage charging energy (HF2); another category consisted of indirectly calculated features, including voltage mean (HF3) and power spectral density (HF4); and the third category consisted of similarity features, including Euclidean distance (HF5) and Manhattan distance (HF6). Then, a feature fusion method based on entropy weighting and correlation factors was used to weight and fuse the six health features, resulting in the fused health features for each training sample. Next, the training samples from the early stage data (e.g., the first 257 cycles) are used as the training set to train the model, while the training samples from the remaining cycles (i.e., the remaining cycles excluding the first 257 cycles) are used as the test set to input into the trained model to obtain the output SOH prediction results.

[0134] To verify the accuracy and advantages of the model, Figure 5This diagram illustrates the performance comparison between the GWO-LSSVM-AdaBoost model (hereinafter referred to as the proposed model) provided in this embodiment and LSSVM (labeled as model 1), GWO-BiGRU (labeled as model 2), GWO-BiLSTM (labeled as model 3), GWO-LSSVM (labeled as model 4), and LSSVM-AdaBoost (labeled as model 5). The GWO-LSSVM-AdaBoost model provided in this embodiment has the lowest MAE, MAPE, and RMSE, indicating higher prediction accuracy.

[0135] Based on the same inventive concept as the foregoing embodiments, such as Figure 6 As shown, this embodiment of the invention also provides a battery health status prediction device, comprising:

[0136] The acquisition module is used to acquire the charging voltage data of the battery under test;

[0137] The extraction module is used to extract health indicators that indicate the health status of the battery from the charging voltage data located in the target voltage range; the health indicators include at least two of the following: equal-voltage charging time, equal-voltage charging energy, average voltage, power spectral density value, Euclidean distance, and Manhattan distance;

[0138] The fusion module is used to fuse health indicators using a weighted fusion method to obtain fused health indicators; wherein, the weight of each health indicator is determined based on its correlation with the battery health status;

[0139] The processing module is used to input the fused health indicators into the trained prediction model to obtain the health status prediction result of the battery to be tested; the prediction model is constructed by weighted combination of multiple support vector machine models optimized by swarm intelligence optimization algorithm through an ensemble learning framework.

[0140] Specific limitations regarding the battery health state prediction device can be found in the limitations of the battery health state prediction method described above, and will not be repeated here. Each module in the aforementioned battery health state prediction device can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device in hardware form, or stored in the memory of a computer device in software form, so that the processor can call and execute the corresponding operations of each module.

[0141] In summary, the battery health status prediction device provided in the above embodiments organically combines the selection of charging voltage data within the target voltage range, the extraction and fusion of multiple health indicators, and a high-performance hybrid prediction model. This enables collaborative prediction of battery health status with high accuracy, high efficiency, and high robustness, solving the technical problems of insufficient accuracy, computational complexity, and neglect of early changes in existing battery health status prediction methods. It can achieve real-time, high-accuracy, and high-efficiency prediction of battery health status.

[0142] This application also provides a computing device, including a processor and a memory, wherein the memory stores computer program instructions for execution on the processor, and when the processor executes the computer program instructions, it implements the battery health state prediction method as described above.

[0143] This application also provides a computer-readable storage medium storing computer instructions, which, when executed by a processor, implement the battery health status prediction method described above.

Claims

1. A method for predicting battery health status, characterized in that, The method includes: Step S1: Obtain the charging voltage data of the battery under test; Step S2: Extract health indicators for indicating battery health status from the charging voltage data located in the target voltage range. The health indicators include at least two of the following: equal-voltage charging time, equal-voltage charging energy, average voltage, power spectral density value, Euclidean distance, and Manhattan distance. The target voltage range refers to an optimal sub-range determined from the battery charging voltage curve. Within this optimal sub-range, the equal-voltage charging time shows the strongest and most stable correlation with the degradation of battery health status, and it can also achieve rapid data acquisition, thus becoming the optimal voltage window for extracting health indicators for predicting battery health status. Step S3: The health indicators are fused using a weighted fusion method to obtain fused health indicators; wherein, the weight of each health indicator is determined based on its correlation with the battery health status. Step S4: Input the fused health indicators into the trained prediction model to obtain the battery health status prediction result corresponding to the battery under test; the prediction model is constructed by weighted combination of multiple support vector machine models optimized by swarm intelligence optimization algorithm through an ensemble learning framework.

2. The method as described in claim 1, characterized in that, Before step S1, the following are included: Obtain a training sample set containing multiple training samples, each training sample including a charging voltage data sample and a corresponding battery health status sample; The charging voltage data samples in all training samples are analyzed based on the bicorrelation analysis method to determine the target voltage range; For each training sample, health indicators are extracted from the charging voltage data samples located in the target voltage range. A method combining entropy weight and correlation factor is used to fuse the health indicators corresponding to each training sample to obtain the fused health indicator for each training sample. The prediction model is trained based on the fused health indicators of all training samples and the corresponding battery health status samples.

3. The method as described in claim 2, characterized in that, Historical charging voltage data from all training samples were analyzed using a dual correlation analysis method to determine the target voltage range, including: Within a preset charging voltage range, multiple candidate voltage ranges are constructed using different combinations of start and stop voltages; For each candidate voltage range, the corresponding equal-voltage charging time sequence is extracted from all training samples, and the Pearson correlation coefficient and Spearman correlation coefficient between the equal-voltage charging time sequence and the historical battery health state sequence corresponding to all training samples are calculated. The correlation coefficient and data sampling duration of each candidate voltage range are comprehensively evaluated, and the candidate voltage range that meets the preset performance conditions is selected as the target voltage range. The data sampling duration of each candidate voltage range is determined based on the equal voltage difference charging duration sequence corresponding to that candidate voltage range.

4. The method as described in claim 2, characterized in that, A method combining entropy weighting and correlation factors is used to fuse the health indicators corresponding to each training sample, resulting in a fused health indicator for each training sample, including: For each training sample, calculate the set of health indicators separately. X=[X 1 ,X 2 ,…X i …,X m ] The i-th health indicator X i Pearson correlation coefficient between the corresponding battery health status samples γ i Correlation coefficient with Spearman ρ i ; m represents the number of health indicators; According to the formula Calculate health indicators separately X i Information entropy E i ;in, It is the normalized first i The health indicator at the first j The values ​​in the training samples, where n is the number of training samples. Ɛ Approaching zero; According to the formula and Determine health indicators X i weight w i ; According to the formula Calculate the fusion health index for each training sample IHF .

5. The method as described in claim 2, characterized in that, The prediction model is trained based on the fused health indicators and corresponding battery health status samples from all training samples, including: Initialize the population for swarm intelligence optimization algorithms; With the goal of minimizing the battery health state prediction error, the parameters of the support vector machine model are iteratively optimized based on the fused health indicators of all training samples and the corresponding battery health state samples. The optimized support vector machine model parameters are output to form a weak predictor; Using an ensemble learning algorithm, each weak predictor is weighted and combined based on its prediction performance on the training sample set to obtain the final strong predictor.

6. The method according to any one of claims 1 to 5, characterized in that, The swarm intelligence optimization algorithm is the Grey Wolf Optimization Algorithm, the support vector machine model is the Least Squares Support Vector Machine Model, and the ensemble learning algorithm is the AdaBoost Algorithm.

7. The method as described in claim 1, characterized in that, Equal voltage difference charging time ( HF 1 ) and equal pressure difference charging energy ( HF 2 The calculation formula for ) is as follows: in, k It is the number of loops. t 1 (k) , t 2 (k) These correspond to the start and end times of the selected voltage segment in each loop, respectively. U Indicates voltage; Voltage mean characteristics ( HF 3 ) and power spectral density characteristics ( HF 4 The calculation formula for ) is as follows: Where n is the number of voltage sampling points collected in each cycle within the voltage range. This is the characteristic of the average voltage in the k-th cycle. V i It is the first i Voltage values ​​at each sampling point S V ( f ) is frequency f Power spectral density at; Euclidean distance ( HF 5 ) and Manhattan distance ( HF 6 The calculation formula for ) is as follows: in, a i and b i Representing the reference voltage curve and the first k Voltage data points of the voltage curve for each cycle.

8. A battery health status prediction device, characterized in that, include: The acquisition module is used to acquire the charging voltage data of the battery under test; The extraction module is used to extract health indicators that indicate the battery's health status from charging voltage data located within a target voltage range. The health indicators include at least two of the following: isobaric charging time, isobaric charging energy, average voltage, power spectral density, Euclidean distance, and Manhattan distance. The target voltage range refers to an optimal sub-range determined from the battery charging voltage curve. Within this optimal sub-range, the isobaric charging time exhibits the strongest and most stable correlation with the degradation of the battery's health status, while also enabling rapid data acquisition, thus becoming the optimal voltage window for extracting health indicators to predict the battery's health status. The fusion module is used to fuse health indicators using a weighted fusion method to obtain fused health indicators; wherein, the weight of each health indicator is determined based on its correlation with the battery health status; The processing module is used to input the fused health indicators into the trained prediction model to obtain the health status prediction result of the battery to be tested; the prediction model is constructed by weighted combination of multiple support vector machine models optimized by swarm intelligence optimization algorithm through an ensemble learning framework.

9. A computing device, characterized in that, The device includes a processor and a memory, the memory storing computer program instructions for execution on the processor, wherein when the processor executes the computer program instructions, it implements the battery health state prediction method as claimed in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, It stores computer instructions, which, when executed by a processor, implement the battery health status prediction method as described in any one of claims 1 to 7.

Images