Method for predicting state of health of lithium ion battery based on time-frequency mixed characteristics

Abstract

The invention belongs to the field of energy storage battery health state estimation, and particularly relates to a method for estimating the health state of an energy storage battery by using a machine learning algorithm and a data model based on extracting and processing time-frequency mixed features of the battery. According to the method, the key information of the battery health state can be comprehensively extracted by mixing the time-frequency characteristics, and the prediction accuracy and stability are improved. Compared with a single time domain or frequency domain characteristic, the method combines an incremental capacity curve (time domain characteristic) and an EIS impedance parameter (frequency domain characteristic), can more accurately characterize the degradation process of the battery, and can interpretively identify key factors influencing the health state of the battery. In addition, the method is suitable for different temperature complex working conditions and has high generalization ability. A transparent and credible prediction result is provided for the BMS, and the use and maintenance strategy of the battery is optimized.

Description

Technical Field

[0001] The present invention belongs to the field of health status estimation of energy storage batteries. More specifically, it relates to a method for estimating the health status of energy storage batteries based on extracting and processing time-frequency mixed features of the battery and using machine learning algorithms and data models. Background Art

[0002] As lithium-ion batteries (LIBs) become increasingly commonplace in modern society, the safety risks and range anxiety they raise are becoming increasingly prominent. The performance degradation and aging of LIBs can easily lead to system failure, posing potential risks of property damage and personal injury. Range anxiety often stems from a mismatch between the mileage displayed on the instrument panel and the battery's actual remaining capacity. Consequently, battery state-of-health (SOH) prediction has garnered significant attention. Currently, mainstream SOH scalar characterization methods are primarily based on historical cycle discharge capacity or internal resistance parameters. Specifically, in energy-focused applications such as electric vehicles, capacity parameters are often used as SOH indicators; while internal resistance parameters are primarily used in power-focused applications such as hybrid vehicles. To achieve these goals, existing approaches have proposed three types of solutions: model-driven, data-driven, and hybrid approaches. Data-driven, model-free approaches demonstrate greater applicability. However, existing data-driven approaches generally suffer from limitations in single features, poor model adaptability, and insufficient interpretability. Many existing methods rely solely on time- or frequency-domain features, which struggle to fully characterize the battery degradation process, resulting in limited prediction accuracy. At the same time, under complex operating conditions such as varying temperatures and rates, the model's generalization capabilities are weak, making it difficult to meet practical application requirements. Furthermore, while some data-driven methods can provide highly accurate predictions, they lack interpretability, making it difficult for battery management systems (BMS) to optimize charging and discharging strategies based on these predictions. Summary of the Invention

[0003] In response to the above technical bottlenecks, the present invention discloses a hybrid feature extraction method based on mixed time domain and frequency domain data, which extracts features associated with battery degradation, fully improves interpretability, and constructs a machine learning model to improve prediction accuracy while enhancing the model's adaptability to complex working conditions, providing BMS with transparent and reliable SOH prediction results, and helping to optimize battery management and maintenance strategies.

[0004] The method disclosed in the present invention for predicting the health status of a lithium-ion battery based on time-frequency hybrid characteristics includes:

[0005] Step 1: Find a battery dataset with high data quality, sufficient samples, and strong representativeness;

[0006] Step 2: Preprocess the data set, including target value extraction, outlier and missing value processing;

[0007] Step 3: Extract two key frequencies in the mid-frequency region of the impedance data as frequency domain features, and simultaneously process them to obtain the battery incremental capacity curve, and obtain the peak value, peak area, and corresponding voltage as time domain features;

[0008] Step 4: Input the time domain features and frequency domain features as health factors into the machine learning model, estimate the battery health status, and use machine learning evaluation indicators.

[0009] In the above technical solution, step 1 specifically includes:

[0010] The dataset is derived from experimental data from the Cavendish Laboratory at the University of Cambridge, UK. It contains data from 12 commercial lithium-ion button cells with a rated capacity of 45 mAh. The cells were continuously charged and discharged in climate chambers at 25 °C, 35 °C, and 45 °C. During the entire cycle, they were charged at a rate of 1 C to a maximum voltage of 4.2 V and discharged at a rate of 2 C until the voltage dropped to 3 V.

[0011] In the above technical solution, step 2 specifically includes:

[0012] Step 2.1: Extract the battery health status data, which can be expressed as:

[0013]

[0014] Among them, Q n is the capacity of the battery when it is currently fully charged (state of charge SOC = 100), and Q0 is the rated capacity of the battery;

[0015] Step 2.2: Use the Z-score method to filter out outliers, and replace the data exceeding the threshold with the average value before and after;

[0016] Step 2.3: For batteries with missing voltage data, obtain the current by differentiating the capacity-time curve, identify the constant current charging interval, and reconstruct the voltage-time curve within the 3.8 V to 4.2 V voltage window.

[0017] In the above technical solution, step 3 specifically includes:

[0018] Step 3.1: The frequency domain features are obtained by extracting data at two key frequencies, 2.16 Hz and 17.8 Hz, from the impedance data. The slope and curvature of the Nyquist plot in the above frequency data are used as frequency domain features F1 - F4, and the first and second derivatives of the imaginary part with respect to frequency are used as frequency domain features F5 - F8.

[0019] Step 3.2: Using the voltage-time curve extracted above, perform incremental capacity curve analysis on each battery, extracting three key data: curve peak, corresponding peak area, and corresponding voltage, and use these data as time domain features.

[0020] In the above technical solution, step 4 specifically includes:

[0021] Step 4.1: Input the extracted features into the machine learning model. The selected machine learning models include: LightGBM model, XGBoost model, SVR model, and LightGBM model optimized by genetic algorithm;

[0022] Step 4.2: Genetic algorithm will be used to adjust parameters and select features within the LightGBM model;

[0023] Step 4.3. The selected machine learning evaluation metrics include: mean square error (MSE), root mean square error (RMSE), and coefficient of determination (R²).

[0024] Step 4.4: The parameters of the lightGBM model optimized using GA include: learning_rate, num_leaves, max_depth, n_estimators, min_child_samples, reg_alpha, reg_lambda, bagging_fraction, and bagging_freq.

[0025] Because the present invention adopts the above technical means, it has the following beneficial effects:

[0026] This method, through hybrid time-frequency features, comprehensively extracts key information about the battery's state of health, improving prediction accuracy and stability. Compared to single time-domain or frequency-domain features, this method combines incremental capacity curves (time-domain features) with EIS impedance parameters (frequency-domain features), enabling more precise characterization of battery degradation and interpretable identification of key factors influencing battery state of health. Furthermore, this method is applicable to complex operating conditions at varying temperatures and exhibits strong generalization capabilities. This method provides transparent and reliable prediction results for the BMS, helping to optimize battery usage and maintenance strategies. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 A flowchart of the implementation steps of the method for predicting the health status of a lithium-ion battery based on time-frequency hybrid features provided by an embodiment of the present invention;

[0028] Figure 2 The capacity decay curve of the battery provided in the embodiment of the present invention;

[0029] Figure 3Iterative flow chart of the genetic algorithm for LightGBM model tuning and feature screening provided by the embodiment of the present invention;

[0030] Figure 4 A comparison chart of the actual SOH value of a 25C05 battery and the predicted SOH value of the battery at 25 °C using the optimized GA-LightGBM model provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0031] To make the objectives, technical solutions, and advantages of the embodiments of the present invention more clear, the technical solutions of the embodiments of the present invention will be clearly and completely described below in conjunction with the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Generally, the components of the embodiments of the present invention described and shown in the drawings herein can be arranged and designed in various different configurations.

[0032] Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the invention as claimed, but rather merely represents selected embodiments of the present invention. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without creative effort are intended to fall within the scope of protection of the present invention.

[0033] It should be noted that similar reference numerals and letters represent similar items in the following figures. Therefore, once an item is defined in one figure, it does not need to be further defined or explained in subsequent figures. It should be noted that the features of the embodiments of the present invention can be combined with each other without conflict.

[0034] The present invention discloses a method for predicting the health status of a lithium-ion battery based on time-frequency hybrid characteristics (hereinafter referred to as the prediction method). The prediction method includes the following steps:

[0035] Step 1: Find a battery dataset with high data quality, sufficient samples, and strong representativeness.

[0036] Step 2: Preprocess the data set, including target value extraction, outlier and missing value processing.

[0037] Step 3: Extract two key frequencies in the mid-frequency region of the impedance data as frequency domain features, and simultaneously process them to obtain the battery incremental capacity curve, and obtain the peak value, peak area, and corresponding voltage as time domain features.

[0038] Step 4: Input the time domain features and frequency domain features as health factors into the machine learning model, estimate the battery health status, and use machine learning evaluation indicators.

[0039] Example 1

[0040] The method for predicting the health status of a lithium-ion battery based on time-frequency mixed characteristics includes the following steps:

[0041] Step 1: Find a battery dataset with high data quality, sufficient samples, and strong representativeness.

[0042] In the above technical solution, step 1 specifically includes:

[0043] The dataset used is a public dataset from the Cavendish Laboratory at the University of Cambridge, UK. This dataset contains over 20,000 electrochemical impedance spectroscopy (EIS) test records for 12 commercial Eunicell LR2032 lithium-ion button cell batteries with a rated capacity of 45 mAh and a lithium cobalt oxide / graphite system, covering various states of charge and temperature conditions. Each battery group underwent continuous charge-discharge cycling in an incubator at 25°C (numbered 25C01-25C08), 35°C (35C01 and 35C02), and 45°C (45C01 and 45C02). The charge phase was performed at a constant current of 1C (45 mA) to a cutoff voltage of 4.2 V, and the discharge phase was performed at a constant current of 2C (90 mA) to a cutoff voltage of 3 V. After every even-numbered cycle, the batteries were left to rest for 15 minutes at both 0% and 100% state of charge (SOC) for open-circuit EIS testing. The test frequency range was 0.02 Hz to 20 kHz, and the excitation current was 5 mA. Nine differentiated charge-discharge strategies (Stages I-IX) were used for each battery group. During each charge-discharge cycle, EIS measurements were performed at 60 fixed frequency points within the 0.02 Hz to 20 kHz frequency range, with both real and imaginary impedance data recorded. Capacity decay was calculated from the discharge capacity after an odd-numbered cycle. The capacity of the eight commercial lithium-ion batteries used in this study showed a significant capacity decay with increasing cycle number. Of the nine different stages, Stage V (15 minutes of rest after full charge) was selected for this study. After a period of rest, the chemical reactions within the battery cease, leaving the battery in a relatively stable state, simplifying subsequent battery health prediction.

[0044] Step 2: Preprocess the dataset, including target value extraction, outlier and missing value processing.

[0045] In the above technical solution, step 2 specifically includes:

[0046] Step 2.1, wherein the target value of the present invention is the health status data of the battery, which can be expressed as:

[0047]

[0048] Q nis the capacity of the battery when it is currently fully charged (state of charge SOC = 100), and Q0 is the rated capacity of the battery;

[0049] Step 2.2: The Z-score method is mainly used to filter out outliers, and the data exceeding the threshold are replaced by the average value before and after.

[0050] Step 2.3. Since some batteries in the data set lack voltage data, the present invention uses the following method to process and supplement the missing voltage data: although there is no voltage data, the capacity data of a certain battery exists. First, the capacity data is used to derive the time to obtain the current data of the battery; the charging process can be divided into constant current charging and constant voltage charging through the current data; after identifying the constant current charging interval, the time data in the constant current charging interval is obtained; based on the battery with complete data, the starting voltage (3.8 V) and the ending voltage (4.2 V) are obtained, and a linear voltage is set within the voltage window of 3.8 V to 4.2 V to reconstruct the voltage-time curve.

[0051] Step 3: Extract two key frequencies in the mid-frequency region of the impedance data as frequency domain features, and simultaneously process them to obtain the battery incremental capacity curve, and obtain the peak value, peak area, and corresponding voltage as time domain features.

[0052] In the above technical solution, step 3 specifically includes:

[0053] Step 3.1: The frequency domain features are obtained by extracting data at two key frequencies, 2.16 Hz and 17.8 Hz, from the impedance data. The slope and curvature of the Nyquist plot in the above frequency data are used as frequency domain features F1 - F4, and the first and second derivatives of the imaginary part with respect to frequency are used as frequency domain features F5 - F8.

[0054] Step 3.2: Extract the voltage-time curve using the method described in step 2, and perform incremental capacity curve analysis on each battery. The incremental capacity analysis is as follows: The incremental capacity analysis method converts the voltage platform of traditional constant current charging into easily observable peaks and valleys on the incremental capacity curve to correspond to the aging state and aging mechanism within the battery. The specific calculation method is as follows: The incremental capacity is equal to the ratio of the difference between the charge level at the two moments before and after to the voltage difference:

[0055]

[0056] Among them, Q k The amount of electricity charged at the kth moment, Q k-1 The amount of electricity charged at the k-1th moment. V k is the voltage value at the kth moment, V k-1is the voltage value at the k-1th moment; for the multi-segment constant current charging data used in this embodiment, three features based on the IC curve are selected, namely, peak value, corresponding peak area, and peak voltage; these three are used as time domain features F9-F11 to construct a hybrid feature set.

[0057] Step 4: Input the time domain features and frequency domain features as health factors into the machine learning model to estimate the battery health status and use machine learning evaluation indicators.

[0058] In the above technical solution, step 4 specifically includes:

[0059] Step 4.1: Input the extracted features into the machine learning model. The selected machine learning models include: LightGBM model, XGBoost model, SVR model, and LightGBM model optimized by genetic algorithm. The model and algorithm principles are as follows:

[0060] Support Vector Machine Regression (SVR):

[0061] SVR is a data-driven method based on structural risk minimization (SRM). ε-SVR is a commonly used SVR model whose goal is to find the optimal estimation function:

[0062]

[0063] Where x represents the input data, w represents the weight, b represents the intercept, and φ(x) represents the feature space. The goal of SVR is to find an equation that makes the maximum deviation of f(x) from the training data less than a predefined value ε; by introducing slack variables 𝜉i and 𝜉i*, the goal is to minimize:

[0064]

[0065] XGBoost (Extreme Gradient Boosting):

[0066] XGBoost is an optimization implementation based on gradient boosted decision trees (GBDTs). It uses the Boosting ensemble learning framework to improve the model's predictive power by combining multiple decision trees in a weighted manner. The basic idea is to train a tree, calculate the residual (the error between the predicted value and the true value), then train the next tree to fit the residual of the previous tree. This process continues iteratively until the desired number of trees is reached or the error converges.

[0067] LightGBM (Light gradient boosting machine):

[0068] LightGBM is also a variant of GBDT, designed specifically to address the limitations of GBDT when processing massive amounts of data, effectively improving its practical application value and computational efficiency. LightGBM's decision tree model adopts a splitting method based on a leaf growth strategy. The computational complexity of this strategy is relatively low, but it is necessary to suppress overfitting by limiting the tree depth and the minimum sample size of leaf nodes. In addition, LightGBM uses a histogram-based decision tree algorithm to discretize the eigenvalues ​​into several "buckets" (buckets), and by searching for split points within these "buckets", it significantly reduces storage requirements and computational consumption. The parallelization framework of this algorithm includes three optimization strategies: feature parallelism when the feature dimension is high, data parallelism when the data scale is large, and voting parallelism when both the feature dimension and the number of votes are high.

[0069] Genetic algorithm:

[0070] Genetic algorithm is a type of random search method developed based on the biological evolution mechanism (natural selection and the principle of survival of the fittest). Through selection, crossover and mutation operations, the initial population continuously generates a group of individuals with higher fitness in iterative evolution, and eventually converges to the optimal solution; in the selection stage, by screening individuals with better fitness in the population, inferior individuals are eliminated to achieve the genetic inheritance of dominant genes. This study adopts the roulette wheel selection method to determine the selection probability based on the proportion of individual fitness to the total fitness of the population; in order to prevent the degeneration of high-quality individuals, an elite retention strategy is introduced to allow the individuals with the highest fitness to be directly copied to the next generation without modification. The crossover operation generates new individuals by performing pairwise gene recombination on the selected individuals, thereby increasing the probability of inheritance of dominant characteristics; assuming that the parent individuals participating in the crossover are X1 and X2, the offspring individuals generated are:

[0071]

[0072]

[0073] In the formula, r∈[1,0], the crossover method ensures that all offspring individuals are in the feasible region, and the mutation operation adopts a small-scale random perturbation method in the neighborhood of the current value. Assuming that the parameter changes within the interval, the mutation operation is as follows:

[0074]

[0075] Step 4.2: Genetic algorithm will be used to adjust parameters and select features within the LightGBM model;

[0076] Step 4.3. The selected machine learning evaluation indicators include mean square error (MSE), root mean square error (RMSE) and coefficient of determination (R²).

[0077] Step 4.3: After determining the features, models, and evaluation metrics, these features were input into different models under different conditions (temperatures). Eight batteries were selected from the dataset for the experiment, numbered 25C01, 25C05, 25C06, 25C07, 35C01, 35C02, 45C01, and 45C02. These batteries underwent stable charge-discharge cycles at three constant temperatures: 25°C, 35°C, and 45°C. A cross-validation strategy was used for each temperature condition. In the 25°C experiment, one battery was selected as the test set, and the remaining three as the training set to construct the SOH estimation model. Similarly, in the 35°C and 45°C experiments, one battery was selected as the test set and the other as the training set, and validation was repeated under the same temperature conditions. This cross-validation framework effectively ensured the model's generalizability to different individual batteries of the same model.

[0078] The present invention discloses a hybrid feature extraction method based on mixed time domain and frequency domain data, which extracts features associated with battery degradation, fully improves interpretability, and constructs a machine learning model to improve prediction accuracy while enhancing the model's adaptability to complex working conditions, providing BMS with transparent and reliable SOH prediction results, helping to optimize battery management and maintenance strategies.

[0079] It will be easily understood by those skilled in the art that the above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A method for predicting the health status of a lithium-ion battery based on time-frequency mixing characteristics, characterized in that: The following steps are involved: Step 1: Find a battery dataset with high data quality, sufficient samples, and strong representativeness; Step 2: Preprocess the data set, including target value extraction, outlier and missing value processing; Step 3: Extract two key frequencies in the mid-frequency region of the impedance data as frequency domain features, and simultaneously process them to obtain the battery incremental capacity curve, and obtain the peak value, peak area, and corresponding voltage as time domain features; Step 4: Input the time domain features and frequency domain features as health factors into the machine learning model, estimate the battery health status, and use machine learning evaluation indicators.

2. The method for optimizing the health status of lithium-ion batteries based on equivalent circuit feature screening according to claim 1, characterized in that: Step 1 specifically includes: The dataset is derived from experimental data from the Cavendish Laboratory at the University of Cambridge, UK. It contains data from 12 commercial lithium-ion button cells with a rated capacity of 45 mAh. The cells were continuously charged and discharged in climate chambers at 25 °C, 35 °C, and 45 °C. During the entire cycle, they were charged at a rate of 1 C to a maximum voltage of 4.2 V and discharged at a rate of 2 C until the voltage dropped to 3 V.

3. The method for predicting the health status of a lithium-ion battery based on time-frequency mixing characteristics according to claim 1, characterized in that: Step 2 specifically includes: Step 2.1: Extract the battery health status data, which can be expressed as:

4. Among them, Q n is the capacity of the battery when it is currently fully charged (state of charge SOC = 100), and Q0 is the rated capacity of the battery; Step 2.2: Use the Z-score method to filter out outliers, and replace the data exceeding the threshold with the average value before and after; Step 2.3: For batteries with partially missing voltage data, obtain the current by differentiating the capacity-time curve. After identifying the constant current charging range, reconstruct the voltage-time curve within the 3.8 V to 4.2 V voltage window.

5. The method for predicting the health status of a lithium-ion battery based on time-frequency mixing characteristics according to claim 1, characterized in that: Step 3 specifically includes: Step 3.

1. Extract the slope and curvature of the Nyquist plot at the two key frequencies of 2.16 Hz and 17.8 Hz from the impedance data as frequency domain features F1-F4. The first and second derivatives of the imaginary part of the frequency data with respect to frequency are used as frequency domain features F5-F8. Step 3.2: Perform incremental capacity curve analysis on each battery using the extracted voltage-time curves. Extract three key data points: curve peak, corresponding peak area, and corresponding voltage. Use these three data points as time-domain features F9-F11 to construct a hybrid feature set.

6. The method for predicting the health status of a lithium-ion battery based on time-frequency mixing characteristics according to claim 1, characterized in that: Step 4 specifically includes: Step 4.1: Input the features extracted in step 3 above into the machine learning model. The selected machine learning models include: LightGBM model, XGBoost model, SVR model, and LightGBM model optimized by genetic algorithm.

7. Step 4.2: Genetic algorithm will be used to adjust parameters and select features within the LightGBM model; Step 4.3: Select machine learning evaluation metrics including mean square error (MSE), root mean square error (RMSE), and coefficient of determination (R²); Step 4.4: The parameters of the lightGBM model optimized using GA include: learning_rate, num_leaves, max_depth, n_estimators, min_child_samples, reg_alpha, reg_lambda, bagging_fraction, and bagging_fre.

Images