Method, system and device for predicting state of health of lithium battery based on relaxation voltage

Abstract

The present disclosure provides a relaxation voltage-based lithium battery state of health prediction method, system and device, wherein the prediction method comprises: obtaining a plurality of groups of sample training data, any sample training data comprising a sample relaxation voltage time series and sample battery state information; extracting first sample statistical features and second sample statistical features based on the sample relaxation voltage time series; splicing and fusing the first sample statistical features and the second sample statistical features to obtain sample fusion features; and training a preset model based on the sample fusion features and the sample battery state information to obtain a battery state of health prediction model. The present disclosure extracts statistical features reflecting global voltage distribution differences and deep features representing different time constant polarization processes at the same time, and fuses them through splicing, thereby balancing the physical interpretability and complex pattern representation capability of the features, and improving the high model training stability and prediction accuracy.

Description

Technical Field

[0001] This disclosure relates to the field of lithium battery health management technology, and in particular to a method, system and device for predicting the health status of lithium batteries based on relaxation voltage. Background Technology

[0002] Lithium iron phosphate batteries have become the mainstream energy storage device for energy storage power stations and electric vehicles due to their advantages such as high safety, low cost and long cycle life. Their state of health (SOH) directly determines the reliability of system operation.

[0003] Battery State of Health (SOH) is an important indicator reflecting the degree of battery performance degradation and is widely used in power battery management, energy storage system maintenance, and battery life management. Existing SOH prediction methods typically rely on charge / discharge capacity testing, internal resistance testing, or data modeling based on complete operating condition curves, which suffers from problems such as high testing costs, insufficient real-time performance, and high requirements for application scenarios.

[0004] Relaxation voltage sequences can reflect the dynamic characteristics of a battery as its internal electrochemical processes gradually reach equilibrium after charging and discharging ceases, containing information related to the battery's aging state. However, modeling using only a single type of feature makes it difficult to simultaneously capture both the overall trend and subtle local variations of the relaxation voltage sequence, thus limiting the model's accuracy and generalization ability. Summary of the Invention

[0005] The technical problem to be solved by this disclosure is to overcome the defects in the prior art and provide a method, system and device for predicting the state of health of lithium batteries based on relaxation voltage. By extracting macroscopic and microscopic features from the relaxation voltage sequence and fusing them by splicing, the model is trained and the state of health is predicted, thereby improving the accuracy of SOH prediction.

[0006] This disclosure solves the above-mentioned technical problems through the following technical solution:

[0007] According to a first aspect of this disclosure, a method for training a battery health state prediction model is provided, the training method comprising:

[0008] Acquire several sets of sample training data, wherein any of the sample training data includes sample relaxation voltage time series and sample battery state information;

[0009] Based on the sample relaxation voltage time series, first sample statistical features and second sample statistical features are extracted; the first sample statistical features are features that reflect the statistical differences in the global voltage distribution in the battery; the second sample statistical features are features of the different time constant planning processes inside the battery.

[0010] The statistical features of the first sample and the statistical features of the second sample are concatenated and fused to obtain the sample fusion features;

[0011] The preset model is trained based on the sample fusion features and the sample battery state information to obtain the battery health state prediction model for predicting the actual battery health state.

[0012] Optionally, the sample relaxation voltage time series includes sample capacity data and sample voltage data at any time point;

[0013] After the step of acquiring several sets of sample training data, and before the step of extracting the first sample statistical features and the second sample statistical features based on the sample relaxation voltage time series, the training method further includes:

[0014] Anomaly identification and processing are performed on the sample relaxation voltage time series;

[0015] The anomaly identification and processing includes: identifying anomalies in the sample size data based on a first preset algorithm, and repairing the identified anomalies in the sample size data using a median algorithm of adjacent normal cycles.

[0016] And / or,

[0017] The anomaly identification and processing includes: identifying abnormal relaxation voltage data based on a preset voltage threshold range and a voltage change threshold between adjacent sampling times, and repairing the abnormal relaxation voltage data by using the average relaxation voltage of the corresponding sampling times of adjacent normal cycles before and after the abnormal cycle.

[0018] Optionally, the first sample statistical characteristics include: sample standard deviation characteristics and sample kurtosis characteristics;

[0019] The step of extracting statistical features of the first sample based on the sample relaxation voltage time series includes:

[0020] Based on the second preset algorithm, the standard deviation feature and the kurtosis feature of the sample relaxation voltage time series are extracted;

[0021] And / or,

[0022] The second sample statistical features include: the sample voltage attenuation difference characteristics within different time intervals;

[0023] The step of extracting statistical features of the second sample based on the sample relaxation voltage time series includes:

[0024] The sample relaxation voltage time series is input into a multi-scale feature extraction network to extract the sample voltage decay difference features.

[0025] Optionally, the step of concatenating and fusing the first sample statistical features and the second sample statistical features to obtain sample fusion features includes:

[0026] The sample standard deviation feature and the sample kurtosis feature are combined to form the first sample statistical feature vector;

[0027] The sample voltage attenuation difference features are used to form a second sample statistical feature vector;

[0028] The first sample statistical feature vector and the second sample statistical feature vector are concatenated along the feature dimension to obtain the sample fusion feature.

[0029] Optionally, after obtaining the battery health state prediction model, the method further includes validating the battery health state prediction model, the validation step including:

[0030] The sample fusion features and the sample battery state information are divided into a training set and a validation set, or a training set and a test set.

[0031] The sample fusion features from the validation set or the test set are input into the trained battery health status prediction model to obtain the predicted battery health status.

[0032] The performance of the battery health state prediction model is evaluated based on the error between the predicted battery health state and the actual battery health state information.

[0033] The performance evaluation metrics include at least one or more of the following: mean square error, root mean square error, and mean absolute error.

[0034] And / or,

[0035] The preset model includes the XGBoost model; the XGBoost model is globally searched and tuned using a Bayesian optimization method.

[0036] According to a second aspect of this disclosure, a method for predicting battery health status is provided, the method comprising:

[0037] Obtain the actual relaxation voltage time series of the battery under test;

[0038] Based on the actual relaxation voltage time series, a first actual statistical feature and a second actual feature are extracted; wherein, the first actual statistical feature is a statistical feature reflecting the difference in global voltage distribution in the battery, and the second actual feature is a feature characterizing the polarization process of different time constants inside the battery.

[0039] The first actual statistical feature and the second actual feature are spliced ​​and fused together to obtain the actual fused feature;

[0040] The actual fused features are input into the battery health status prediction model trained by the training method described in any one of claims 1 to 5, and the actual health status prediction result of the battery under test is output.

[0041] According to a third aspect of this disclosure, a training system for a battery health state prediction model is provided, the training system comprising:

[0042] The sample data acquisition module is used to acquire several sets of sample training data, wherein any sample training data includes sample relaxation voltage time series and sample battery state information.

[0043] The sample feature extraction module is used to extract a first sample statistical feature and a second sample statistical feature based on the sample relaxation voltage time series; the first sample statistical feature is a feature reflecting the statistical difference in the global voltage distribution in the battery; the second sample statistical feature is a feature of the different time constant planning processes inside the battery;

[0044] The sample feature fusion module is used to concatenate and fuse the first sample statistical features and the second sample statistical features to obtain sample fusion features;

[0045] The model training module is used to train a preset model based on the sample fusion features and the sample battery state information to obtain the battery health state prediction model for predicting the actual battery health state.

[0046] Optionally, the sample relaxation voltage time series includes sample capacity data and sample voltage data at any time point;

[0047] The training system further includes a preprocessing module, which is used to perform anomaly identification and processing on the sample relaxation voltage time series after acquiring several sets of sample training data and before extracting the first sample statistical features and the second sample statistical features based on the sample relaxation voltage time series.

[0048] The anomaly identification and processing includes: identifying anomalies in the sample size data based on a first preset algorithm, and repairing the identified anomalies in the sample size data using a median algorithm of adjacent normal cycles.

[0049] And / or,

[0050] The anomaly identification and processing includes: identifying abnormal relaxation voltage data based on a preset voltage threshold range and a voltage change threshold between adjacent sampling times, and repairing the abnormal relaxation voltage data by using the average relaxation voltage of the corresponding sampling times of adjacent normal cycles before and after the abnormal cycle.

[0051] Optionally, the first sample statistical characteristics include: sample standard deviation characteristics and sample kurtosis characteristics;

[0052] The sample feature extraction module is used to extract the sample standard deviation feature and the sample kurtosis feature from the sample relaxation voltage time series based on a second preset algorithm.

[0053] And / or,

[0054] The second sample statistical features include: the sample voltage attenuation difference characteristics within different time intervals;

[0055] The sample feature extraction module is used to input the sample relaxation voltage time series into a multi-scale feature extraction network to extract the sample voltage decay difference features.

[0056] Optionally, the sample feature fusion module is used to combine the sample standard deviation feature and the sample kurtosis feature to form a first sample statistical feature vector;

[0057] The sample voltage attenuation difference features are used to form a second sample statistical feature vector;

[0058] The first sample statistical feature vector and the second sample statistical feature vector are concatenated along the feature dimension to obtain the sample fusion feature.

[0059] Optionally, the training system further includes a model validation module, which, after obtaining the battery health state prediction model, further validates the battery health state prediction model.

[0060] The model validation module is used to divide the sample fusion features and the sample battery state information into a training set and a validation set, or a training set and a test set.

[0061] The sample fusion features from the validation set or the test set are input into the trained battery health status prediction model to obtain the predicted battery health status.

[0062] The performance of the battery health state prediction model is evaluated based on the error between the predicted battery health state and the actual battery health state information.

[0063] The performance evaluation metrics include at least one or more of the following: mean square error, root mean square error, and mean absolute error.

[0064] And / or,

[0065] The preset model includes the XGBoost model; the XGBoost model is globally searched and tuned using a Bayesian optimization method.

[0066] According to a fourth aspect of this disclosure, a battery health state prediction system is provided, characterized in that the prediction system comprises:

[0067] The actual data acquisition module is used to acquire the actual relaxation voltage time series of the battery under test;

[0068] The actual feature extraction module is used to extract a first actual statistical feature and a second actual feature based on the actual relaxation voltage time series; wherein, the first actual statistical feature is a statistical feature reflecting the difference in global voltage distribution in the battery, and the second actual feature is a feature characterizing the polarization process of different time constants inside the battery;

[0069] The actual feature fusion module is used to splice and fuse the first actual statistical feature and the second actual feature to obtain the actual fused feature;

[0070] The health prediction module is used to input the actual fused features into the battery health state prediction model of the first aspect of this disclosure, and output the actual health state prediction result of the battery under test.

[0071] According to a fifth aspect of this disclosure, an electronic device is provided, including a memory, a processor, and a computer program stored in the memory and for running on the processor, characterized in that, when the processor executes the computer program, it implements a training method for a battery health state prediction model according to a first aspect of this disclosure, and / or a battery health state prediction method according to a second aspect of this disclosure.

[0072] According to a sixth aspect of this disclosure, a computer-readable storage medium is provided having a computer program stored thereon, characterized in that, when the computer program is executed by a processor, it implements a training method for a battery health state prediction model according to a first aspect of this disclosure, and / or a battery health state prediction method according to a second aspect of this disclosure.

[0073] According to a seventh aspect of this disclosure, a computer program product is provided, comprising a computer program, characterized in that, when the computer program is executed by a processor, it implements a training method for a battery health state prediction model according to a first aspect of this disclosure, and / or a battery health state prediction method according to a second aspect of this disclosure.

[0074] Based on common knowledge in the field, the above-mentioned preferred conditions can be combined arbitrarily to obtain various preferred embodiments of this disclosure.

[0075] The positive and progressive effects of this disclosure are as follows:

[0076] This disclosure extracts features based on the relaxation voltage time series after battery charging, enabling battery health status prediction without relying on complex disassembly or additional invasive detection. The data is easy to acquire and has strong engineering applicability. Furthermore, this disclosure simultaneously extracts statistical features reflecting global voltage distribution differences and deep features characterizing polarization processes at different time constants, and fuses them through a splicing method, thereby balancing the physical interpretability of the features and the ability to represent complex patterns, thus improving the completeness of feature expression.

[0077] Furthermore, this disclosure performs anomaly identification and processing on the sample training data before model training, which can reduce the adverse effects of capacity jumps, sensor noise, and abnormal relaxation voltage data on model training, and improve the stability and prediction accuracy of model training. This disclosure uses the XGBoost model to learn the fused features, which can make full use of the nonlinear mapping relationship between the fused features and the battery health status, thereby improving the health status prediction performance. Attached Figure Description

[0078] Figure 1 This is a schematic diagram of the training method for the battery health status prediction model provided in Embodiment 1 of this disclosure;

[0079] Figure 2 This is a schematic diagram of the sample fusion process provided in Embodiment 1 of this disclosure;

[0080] Figure 3 These are four sets of battery discharge capacity curves provided in Embodiment 1 of this disclosure;

[0081] Figure 4 This is a graph showing the relaxation voltage change of battery 3 after constant current and constant voltage charging, as provided in Embodiment 1 of this disclosure.

[0082] Figure 5 This is a feature extraction heatmap of battery 3 provided in Embodiment 1 of this disclosure;

[0083] Figure 6 This is a graph showing the change in capacity of the battery 3 provided in Embodiment 1 of this disclosure as a function of the number of cycles.

[0084] Figure 7 This is a graph showing the predicted SOH of battery 3 provided in Embodiment 1 of this disclosure;

[0085] Figure 8 This is a flowchart illustrating the battery health status prediction method provided in Embodiment 2 of this disclosure;

[0086] Figure 9 This is a schematic diagram of the structure of the training system for the battery health state prediction model provided in Embodiment 3 of this disclosure;

[0087] Figure 10 This is a schematic diagram of the battery health status prediction system provided in Embodiment 4 of this disclosure;

[0088] Figure 11 This is a schematic diagram of the structure of the electronic device provided in Embodiment 5 of this disclosure. Detailed Implementation

[0089] The present disclosure is further illustrated below by way of embodiments, but the present disclosure is not limited to the scope of the embodiments described herein.

[0090] The prefixes such as "first" and "second" used in this disclosure are merely for distinguishing different descriptive objects and do not limit the position, order, priority, quantity, or content of the described objects. The use of ordinal numbers and other prefixes used to distinguish descriptive objects in this disclosure does not constitute a limitation on the described objects. The description of the described objects is given in the claims or the context of the embodiments, and should not be construed as an unnecessary limitation. Furthermore, in the description of this embodiment, unless otherwise stated, "multiple" means two or more.

[0091] Battery State of Health (SOH) is an important indicator reflecting the degree of battery performance degradation and is widely used in power battery management, energy storage system maintenance, and battery life management. Existing SOH prediction methods typically rely on charge / discharge capacity testing, internal resistance testing, or data modeling based on complete operating condition curves, which suffers from problems such as high testing costs, insufficient real-time performance, and high requirements for application scenarios.

[0092] Relaxation voltage sequences can reflect the dynamic characteristics of a battery as its internal electrochemical processes gradually reach equilibrium after charging and discharging ceases, containing information related to the battery's aging state. However, modeling using only a single type of feature makes it difficult to simultaneously capture both the overall trend and subtle local variations of the relaxation voltage sequence, thus limiting the model's accuracy and generalization ability.

[0093] Therefore, it is necessary to provide a method, system, and device for predicting the health status of lithium batteries based on relaxation voltage, in order to improve prediction accuracy and robustness.

[0094] Example 1

[0095] like Figure 1 As shown, this embodiment provides a training method for a battery health status prediction model. The training method includes:

[0096] S11: Obtain several sets of sample training data. Any sample training data includes the sample relaxation voltage time series and sample battery state information.

[0097] S12: Extract the first sample statistical features and the second sample statistical features based on the sample relaxation voltage time series; the first sample statistical features are the statistical features that reflect the differences in the global voltage distribution in the battery; the second sample statistical features are the features of the different time constant planning processes inside the battery;

[0098] S13: The statistical features of the first sample and the statistical features of the second sample are concatenated and fused to obtain the sample fusion features;

[0099] S14: Train the preset model based on sample fusion features and sample battery state information to obtain a battery health state prediction model for predicting the actual battery health state.

[0100] Among them, the relaxation voltage after the charging process of the sample relaxation voltage time series serves as an independent and pure source of aging characteristics. During the relaxation phase, the external power failure causes the input current to return to zero, completely eliminating the strong coupling interference between dynamic load and instantaneous ohmic polarization. This allows the spontaneous depolarization process of the terminal voltage to accurately map the real thermodynamic equilibrium evolution of internal charge transfer and ion diffusion. This mechanism eliminates the dependence on continuous and harsh charging and discharging conditions, and only requires the use of fragmented rest periods of the equipment (such as 30 to 1800 seconds) to extract effective features, thus broadening the adaptability of the health state estimation model to real and complex operating conditions such as electric vehicles or energy storage power stations.

[0101] This disclosure extracts features based on the relaxation voltage time series after battery charging, enabling battery health status prediction without relying on complex disassembly or additional invasive detection. The data is easy to acquire and has strong engineering applicability. Furthermore, this disclosure simultaneously extracts statistical features reflecting global voltage distribution differences and deep features characterizing polarization processes at different time constants, and fuses them through a splicing method, thereby balancing the physical interpretability of the features and the ability to represent complex patterns, thus improving the completeness of feature expression.

[0102] The sample relaxation voltage time series in this embodiment includes sample capacity data and sample voltage data at any time point.

[0103] After acquiring several sets of sample training data and before extracting the first and second sample statistical features based on the sample relaxation voltage time series, the training method further includes:

[0104] Anomaly identification and processing were performed on the sample relaxation voltage time series.

[0105] The anomaly identification and processing includes: identifying anomalies in the sample size data based on a first preset algorithm, and repairing the identified abnormal sample size data using the median algorithm of adjacent normal cycles.

[0106] Anomaly identification and processing also include: identifying abnormal relaxation voltage data based on a preset voltage threshold range and a voltage change threshold between adjacent sampling times, and repairing the abnormal relaxation voltage data by using the average relaxation voltage of the corresponding sampling times of adjacent normal cycles before and after the abnormal cycle.

[0107] In one implementation, the full life cycle charge-discharge sequence of different individual batteries is collected. For example, based on the thermodynamic equilibrium characteristics during the aging process of lithium batteries, the relaxation voltage sequence after constant current and constant voltage charging is extracted (e.g., the extraction range is 30-1800 seconds, the sampling interval is 30 seconds, and a 60-dimensional relaxation voltage time series is constructed).

[0108] Given the non-stationarity and local regeneration effect of battery degradation sequences, traditional mean-sensitive anomaly detection is abandoned. Instead, a robust Z-score algorithm based on the absolute deviation of adjacent cyclic medians (MAD) is adopted (a commonly used statistical technique for identifying outliers in a dataset). The calculation formula is as follows:

[0109]

[0110] in, For the current capacity, is the median of the capacity of 5 consecutive cycles, and 1.4826 is the normal distribution fit constant. When When a point is identified as an abnormal point in the aging trajectory, adaptive smoothing repair is performed using the median value of adjacent healthy trajectory sequences to obtain sample size data.

[0111] Furthermore, to address the relaxation start drift caused by differences in the initial internal resistance of different batteries, a personalized envelope threshold (e.g., 3.35-3.48V for batteries 1, 2, and 4) is adopted, combined with polarization voltage abrupt change detection (adjacent gradient difference). The process involves removing sensor noise and internal short-circuit transient anomalies, and ensuring the continuity of the voltage spatial manifold through sequential interpolation; ultimately, sample voltage data is generated. Through the above anomaly handling and repair process, highly robust cleaning and feature alignment of non-stationary degradation sequences of lithium-ion batteries are achieved.

[0112] To address the issue of battery non-stationary degradation sequences being highly susceptible to local noise contamination, an adaptive joint data cleaning mechanism based on local manifold continuity was constructed. Traditional mean filtering, which easily distorts effective aging details, was abandoned. Instead, a robust Z-score algorithm based on the absolute deviation of adjacent cyclic medians (MAD) was employed for capacity anomaly identification, supplemented by a customized relaxation polarization gradient threshold. This mechanism effectively eliminates local capacity regeneration effects, internal short-circuit transient anomalies, and sensor flying-point noise. Furthermore, through adaptive smooth interpolation of preceding and following time sequences, it preserves the true physical details of the original micro-polarization fluctuations, improving the signal-to-noise ratio and effective data utilization of the front-end feature input matrix.

[0113] Relaxation voltage refers to the transient depolarization response sequence of a battery's terminal voltage as it spontaneously recovers to its open-circuit voltage (OCV) over time after constant current-constant voltage (CC-CV) charging is completed and the external current is cut off, allowing it to enter a resting state. This process is unaffected by external current interference and can accurately map the thermodynamic equilibrium evolution process across multiple spatial scales within the battery.

[0114] As batteries age, the internal solid electrolyte interphase (SEI) thickens, active materials are lost, and ion transport channels become blocked, directly altering the shape of the relaxation voltage curve. The evolution of the relaxation voltage is not a simple change in slope, but rather a superposition of multiple physical processes with different time constants:

[0115] Short-term relaxation (tens of seconds): mainly dominated by charge transfer polarization, reflecting the decay of reaction kinetics at the electrode / electrolyte interface;

[0116] Long relaxation time (hundreds to thousands of seconds): mainly dominated by concentration polarization, reflecting the slow solid-phase diffusion of lithium ions inside the positive and negative electrode active particles that is hindered.

[0117] Conventional single voltage change rate methods cannot isolate the complex aging mechanisms described above. Therefore, this disclosure specifically designs a multi-scale relaxation voltage feature extraction process to address the unique properties of relaxation voltage with multiple time constants and multiple polarizations, thereby achieving accurate decoupling and mapping of aging mechanisms at different depths.

[0118] Specifically, the first sample statistical features in this embodiment include: sample standard deviation features and sample kurtosis features;

[0119] The steps for extracting the statistical features of the first sample based on the sample relaxation voltage time series include:

[0120] Based on the second preset algorithm, the standard deviation feature and kurtosis feature of the sample relaxation voltage time series are extracted.

[0121] In one implementation, statistics reflecting the differences in global voltage distribution are extracted, such as standard deviation (std) and kurtosis. The standard deviation This reflects the increased volatility caused by increased internal impedance; kurtosis The steepness of the rate of change of relaxation voltage.

[0122]

[0123]

[0124] Where v represents voltage data.

[0125] The second sample statistical characteristics include: the sample voltage attenuation difference characteristics in different time intervals;

[0126] The steps for extracting statistical features of the second sample based on the sample relaxation voltage time series include:

[0127] The sample relaxation voltage time series is input into a multi-scale feature extraction network to extract the sample voltage decay difference features.

[0128] In one implementation, statistical features of the second sample are extracted through piecewise relaxation polarization gradients. Specifically, for different time constant polarization processes within the battery, the relaxation voltage decay at specific time points is extracted nonlinearly to characterize aging at different depths.

[0129]

[0130]

[0131]

[0132] Subsequently, by utilizing the coupling strength between the Pearson correlation metric and the target SOH, redundant interference was eliminated, and 3-4 samples with high information entropy voltage attenuation difference characteristics were retained:

[0133]

[0134] In this embodiment, as Figure 2 As shown, the steps for concatenating and fusing the statistical features of the first sample and the statistical features of the second sample to obtain the sample fused features include:

[0135] S21: Combine the sample standard deviation feature and the sample kurtosis feature to form the first sample statistical feature vector;

[0136] S22: Combine the sample voltage attenuation difference features into a second sample statistical feature vector;

[0137] S23: Concatenate the statistical feature vectors of the first sample and the statistical feature vectors of the second sample along the feature dimension to obtain the sample fusion feature.

[0138] In one implementation, to overcome the limitation that a single receptive field cannot capture the voltage relaxation dynamics across the entire frequency band, a multi-scale polarization attention network (MSPAN) is constructed.

[0139] First, normalization is used to eliminate the absolute value offset of the initial terminal voltage:

[0140] The model simultaneously captures the fast transient response (small scale) in the early relaxation phase and the slow diffusion response (large scale) in the later relaxation phase using a parallel multi-scale one-dimensional convolutional array (receptive field scale configuration of 3, 5, and 7). The multi-scale feature map mapping process is as follows:

[0141]

[0142] in These are convolution kernels of different scales.

[0143] Following multi-scale parallel extraction, a novel self-attention mechanism is introduced. Since the key relaxation range of a battery shifts at different aging stages, this mechanism maps features to... The matrix dynamically calculates aging sensitivity weights for different time steps and feature channels through scaling the dot product:

[0144]

[0145] Finally, through global flat pooling and fully connected dimensionality reduction, a highly condensed 64-dimensional microscopic aging characterization vector is output. .

[0146] By concatenating macroscopic physical statistical features with deep microscopic implicit features along the channel dimension, a heterogeneous fusion feature matrix containing multi-scale temporal dynamics and global statistical priors is constructed.

[0147] The Multi-Scale Polarization Attention Network (MSPAN) addresses the challenge of decoupling complex battery polarization mechanisms from single voltage features. Based on the extraction of global degradation statistical prior features, this network uses parallel one-dimensional convolutional branches with multiple receptive fields (scales of 3, 5, and 7) to precisely align and extract physical processes with different time constants within the relaxation voltage. Furthermore, by incorporating a dynamic self-attention mechanism, the model adaptively assigns higher weights to different receptive fields and key relaxation frequency bands based on the battery's current aging state, achieving deep physical feature mining of high-dimensional nonlinear aging trajectories.

[0148] The preset model in this embodiment includes the XGBoost model; the XGBoost model performs global search and tuning through Bayesian optimization to accurately map the high-dimensional heterogeneous feature set composed of extracted features to the health status of the battery.

[0149] The aging process of lithium-ion batteries is highly nonlinear, time-varying, and involves multiple coupled factors, making it difficult for traditional linear regression models to capture deep degradation patterns. This disclosure employs the XGBoost algorithm as the regression prediction engine for SOH (Solar Oxygen Optimization). XGBoost is a highly efficient ensemble learning algorithm based on the gradient boosting framework. Its core idea is to iteratively add new classification regression trees (CART) to fit the residuals of the previous prediction, and finally sum the prediction results of multiple trees as the final SOH output.

[0150] In the feature mapping process disclosed herein, the objective function of XGBoost is defined as a combination of the training loss and the structural risk regularization term, as shown in the following formula:

[0151]

[0152] In the formula, The total number of input feature samples; To measure the true SOH value Compared with model predictions The loss function for the error between them; This represents the total number of decision trees; Indicates the first The internal structure of the tree; This is a regularization penalty term.

[0153] By explicitly introducing a regularization term into the objective function, XGBoost effectively controls model complexity. For small sample sizes and localized noise data that are prone to occur in battery aging tests, this mechanism significantly suppresses overfitting of the model to local anomalies, ensuring strong robustness and generalization ability of the SOH mapping.

[0154] The XGBoost model contains a large number of hyperparameters that significantly affect prediction accuracy (such as learning rate, maximum tree depth, subsample ratio, minimum weight of leaf nodes, etc.). Traditional grid search or random search suffers from the curse of dimensionality, characterized by high blindness and extremely high computational cost, when dealing with high-dimensional parameter spaces. To address this, this disclosure creatively introduces a Bayesian optimization algorithm to perform efficient global optimization in the surrogate space.

[0155] Bayesian optimization is a heuristic strategy for global optimization of computationally expensive black-box functions. It mainly consists of two core modules: the Gaussian Process Surrogate Model (GSM): based on existing hyperparameter evaluation history, it constructs a probability distribution surrogate model of the objective function to fit the mapping topology between the high-dimensional hyperparameter space and the SOH prediction error.

[0156] Acquisition Function: This disclosure uses the expected improvement or confidence upper limit as the acquisition function. This function intelligently guides the sampling direction of the next optimal hyperparameter combination by continuing a fine search near the known low-error region.

[0157] Unlike blind search without prior guidance, Bayesian optimization can fully utilize historical evaluation information from each iteration to update the posterior distribution of the Gaussian process. When faced with high-dimensional inputs that incorporate multi-scale relaxation features, this mechanism can quickly overcome local minima traps within a very small number of iterations (e.g., 50), accurately locating the global subspace of XGBoost's optimal hyperparameters that fits the current battery degradation characteristics, thus achieving an optimal balance between model training efficiency and SOH prediction accuracy.

[0158] An Extreme Gradient Boosting (XGBoost) prediction engine with joint Bayesian surrogate space global optimization was constructed. Addressing the high-dimensional spatial characteristics of multi-source heterogeneous fusion features, the engine utilizes intelligent iteration of a Bayesian Gaussian surrogate model and acquisition function to efficiently overcome local optima traps, achieving accurate global localization of high-dimensional hyperparameters of the XGBoost model with extremely low computational cost. Combined with XGBoost's built-in structural risk regularization penalty mechanism, this architecture effectively suppresses overfitting on small sample and locally oscillating datasets, endowing the SOH mapping model with predictive accuracy and generalization robustness even when facing individual batteries with different initial internal resistances.

[0159] In this embodiment, after obtaining the battery health state prediction model, the method further includes validating the battery health state prediction model. The validation steps include:

[0160] The sample fusion features and sample battery state information are divided into training set and validation set, or training set and test set.

[0161] Input the fused features of samples from the validation set or test set into the trained battery health status prediction model to obtain the predicted battery health status.

[0162] The performance of the battery health state prediction model is evaluated based on the error between the predicted battery health state and the actual battery health state information.

[0163] The performance evaluation metrics include at least one or more of the following: mean square error, root mean square error, and mean absolute error.

[0164] The fused features of the input validation set are fed into the optimized XGBoost mapping model, and the output SOH prediction sequence is calculated using the coefficient of determination (COP). The root mean square error (RMSE) and mean absolute error (MAE) are used to fully quantify the prediction fidelity of the battery health status prediction model in capturing the early plateau region and the later capacity drop region of the battery.

[0165] The following examples illustrate the implementation principle of the training method for the battery health state prediction model disclosed herein:

[0166] I. Specific Implementation Methods for Data Acquisition and Cleaning:

[0167] Data source: such as Figure 3 As shown, charge-discharge cycle data of four 110Ah ternary lithium batteries were collected. The charge-discharge regime was as follows: constant current charging (1C, i.e., 110A) to 3.65V, then constant voltage charging to current ≤0.05C, resting for 30 minutes (collecting relaxation voltage, 30-1800 seconds, 30-second intervals), and then constant current discharging (1C) to 2.5V. The discharge capacity and relaxation voltage data of each cycle were recorded. The number of cycles for each battery was 1500-2000.

[0168] II. Abnormal Capacity Cleaning:

[0169] Calculate the Z-score for the capacity of each loop: For loop i, take the capacity of the two loops before and after it (a total of 4 neighbors), and calculate the median M and the absolute deviation of the median MAD of the neighbors;

[0170] When Z-score > 20, it is considered abnormal, and the capacity i is repaired using the median M of the neighbors;

[0171] Abnormal relaxation voltage cleaning:

[0172] Voltage thresholds are set for each battery: the normal range for batteries 1, 2, and 4 is 3.35-3.48V, and for battery 3 it is 3.34-3.55V (determined based on aging tests of this battery model; test results are as follows). Figure 4 (as shown)

[0173] If the relaxation voltage of a certain cycle exceeds the threshold, or the voltage difference between adjacent time points is >0.05V, it is judged as abnormal and is repaired by the average of the relaxation voltages of the two most recent normal cycles before and after the cycle.

[0174] Detailed implementation of feature extraction, fusion, and SOH prediction:

[0175] III. Statistical Feature Extraction and Screening:

[0176] Calculate the statistical characteristics of relaxation voltage: standard deviation (std), kurtosis, decay_600s (voltage difference between 1230 and 1800 seconds), decay_900s (voltage difference between 330 and 1230 seconds), and decay_1620s (voltage difference between 210 and 1800 seconds).

[0177] like Figure 5 As shown, the Pearson correlation coefficients between each feature and SOH were calculated, and highly correlated features were selected by battery number: the selection threshold for battery 1 was 0.88 (retaining std, decay_600s, and decay_1620s), and the selection threshold for battery 3 was 0.9 (retaining std, kurtosis, and decay_900s); where std, decay_600s, and decay_1620s are the extracted statistical features.

[0178] IV. Deep Feature Extraction (MSFFCM Training):

[0179] Model parameters: Input dimension 60 (relaxation voltage time point), parallel convolution kernel size 3, 5, 7, number of filters 32, number of attention heads 2, fully connected layer output dimension 64;

[0180] Training environment: PyTorch 2.0, GPU is NVIDIA RTX 3070, training epochs 50, batch size=32, optimizer Adam (learning rate 0.001).

[0181] Training results: The training loss (MSE) of the MSFFCM model decreased to below 0.0001, and the validation loss stabilized at around 0.0002, indicating that the model converged;

[0182] V. Feature Fusion and Data Partitioning:

[0183] Align statistical features and deep features according to the number of iterations, and construct a fused feature matrix (e.g., after fusion, battery 1 has 1780 samples and 67 features (3 statistical features + 64 deep features)).

[0184] The training and test sets are divided in an 8:2 ratio to ensure that the training set contains 80% of the cyclic data for each battery and the test set contains 20%.

[0185] VI. Bayesian Optimization and XGBoost Prediction:

[0186] Hyperparameter search range: max_depth: 3-10, learning_rate: 0.01-0.3, n_estimators: 100-500, subsample: 0.7-1.0;

[0187] Optimization tool: BayesianOptimization library (a Python global optimization library based on Gaussian processes and Bayesian inference), 50 iterations, optimal hyperparameters: max_depth=6, learning_rate=0.08, n_estimators=350, subsample=0.9;

[0188] Prediction results: such as Figures 6 to 7 And Table 1 (test set metrics for 4 batteries). Figure 6 This is a comparison chart of the actual capacity and predicted capacity of battery 3. Figure 7 A comparison chart of predicted SOH and actual SOH for battery 3.

[0189] Table 1

[0190]

[0191] Results verification: The RMSE of each battery was less than 0.8%, and the R² was greater than 0.95, meeting the industrial-grade SOH prediction accuracy requirements (RMSE < 1.0%).

[0192] This disclosure constructs a multi-scale polarization feature fusion architecture driven by both macro and micro dimensions, successfully decoupling the physical evolution of different time constants in the relaxation voltage. Compared to the single-dimensional working voltage change rate in existing technologies, which is highly susceptible to external current coupling interference, this architecture can deeply capture the thermodynamic equilibrium evolution law reflecting the true aging state of the battery. Experiments show that this mechanism accurately captures the nonlinear capacity drop trend in the later stages of the battery's life cycle. In the verification of four 110Ah real-vehicle-specification lithium iron phosphate batteries, the root mean square error (RMSE) of SOH prediction was strictly controlled below 0.72%, and the coefficient of determination (R²) was consistently above 0.95. Its mapping fidelity is superior to existing single-feature schemes.

[0193] This paper discloses a robust cleaning mechanism based on a joint approach using median absolute deviation (MAD) and dynamic polarization threshold to address frequent charge / discharge interruptions, local regeneration effects, and high-frequency flying-spot noise from sensors under real-world dynamic operating conditions. This mechanism adheres to the principle of local manifold continuity in battery degradation trajectories, achieving adaptive smoothing repair of anomalous sequences without disrupting realistic microscopic polarization details. In a stringent noise injection experiment with 5%-10% high-intensity anomalous cycles introduced into the validation set, the model's prediction accuracy degradation consistently remained within 0.1%. This improves the robustness of traditional models, which are prone to collapse under low-quality operating data.

[0194] This disclosure focuses on the spontaneous depolarization response (30-1800 seconds) of a battery after power loss, removing from the physical level the strong dependence on specific ambient temperatures, specific charge / discharge rates, and complete continuous cycling procedures. Because it extracts relative relaxation parameters characterizing the increase in internal resistance and the degradation of diffusion capability, rather than absolute electrical parameters, this approach exhibits excellent cross-platform scalability. When migrating to lithium-ion batteries with other rated capacities or different systems, only the envelope voltage threshold and Z-score boundary constant at the front end need to be adaptively adjusted, without redesigning feature engineering or reconstructing the deep network architecture, thus reducing the marginal cost of large-scale engineering deployment.

[0195] Example 2

[0196] like Figure 8 As shown, this embodiment provides a method for predicting battery health status. The prediction method includes:

[0197] S31: Obtain the actual relaxation voltage time series of the battery under test;

[0198] S32: Extract the first actual statistical feature and the second actual feature based on the actual relaxation voltage time series; wherein, the first actual statistical feature is a statistical feature reflecting the difference in global voltage distribution in the battery, and the second actual feature is a feature characterizing the polarization process of different time constants inside the battery;

[0199] S33: The first actual statistical feature and the second actual feature are spliced ​​and fused to obtain the actual fused feature;

[0200] S34: Input the actual fused features into the battery health status prediction model and output the actual health status prediction result of the battery under test.

[0201] The battery health state prediction method provided in this disclosure obtains the actual relaxation voltage time series of the battery under test and extracts a first actual statistical feature reflecting the difference in global voltage distribution and a second actual feature characterizing the polarization process of different time constants. This allows for a comprehensive and multi-dimensional characterization of the battery's internal state. By splicing and fusing the two types of features and inputting them into the battery health state prediction model, the dynamic information related to aging contained in the relaxation voltage can be effectively utilized, avoiding the problems of insufficient information from a single feature or the original voltage sequence and large noise interference. Furthermore, this disclosure does not require a complete charge-discharge test or disassembly analysis of the battery. It can achieve health state prediction based solely on a portion of the voltage response during the relaxation stage, offering advantages such as speed, non-destructiveness, and ease of online implementation. It also improves the accuracy and robustness of the prediction results.

[0202] Example 3

[0203] like Figure 9As shown, this embodiment provides a training system for a battery health state prediction model. The training system includes:

[0204] The sample data acquisition module 101 is used to acquire several sets of sample training data. Any sample training data includes sample relaxation voltage time series and sample battery state information.

[0205] The sample feature extraction module 102 is used to extract first sample statistical features and second sample statistical features based on the sample relaxation voltage time series; the first sample statistical features are features of statistics reflecting the differences in global voltage distribution in the battery; the second sample statistical features are features of different time constant planning processes inside the battery.

[0206] The sample feature fusion module 103 is used to concatenate and fuse the statistical features of the first sample and the statistical features of the second sample to obtain the sample fusion features;

[0207] The model training module 104 is used to train a preset model based on sample fusion features and sample battery state information to obtain a battery health state prediction model for predicting the actual battery health state.

[0208] The sample relaxation voltage time series in this embodiment includes sample capacity data and sample voltage data at any time point.

[0209] The training system also includes a preprocessing module 105, which is used to identify and process anomalies in the sample relaxation voltage time series after acquiring several sets of sample training data and before extracting the first sample statistical features and the second sample statistical features based on the sample relaxation voltage time series.

[0210] The anomaly identification and processing includes: identifying anomalies in the sample size data based on a first preset algorithm, and repairing the identified abnormal sample size data using the median algorithm of adjacent normal cycles.

[0211] Anomaly identification and processing include: identifying abnormal relaxation voltage data based on a preset voltage threshold range and a voltage change threshold between adjacent sampling times, and repairing the abnormal relaxation voltage data by using the average relaxation voltage of the corresponding sampling times of adjacent normal cycles before and after the abnormal cycle.

[0212] Optionally, the first sample statistical characteristics include: sample standard deviation characteristics and sample kurtosis characteristics;

[0213] The sample feature extraction module 102 is used to extract the sample standard deviation feature and sample kurtosis feature from the sample relaxation voltage time series based on the second preset algorithm;

[0214] The second sample statistical characteristics include: the sample voltage attenuation difference characteristics in different time intervals;

[0215] The sample feature extraction module 102 is used to input the sample relaxation voltage time series into the multi-scale feature extraction network to extract the sample voltage decay difference features.

[0216] In this embodiment, the sample feature fusion module 103 is used to combine the sample standard deviation feature and the sample kurtosis feature to form a first sample statistical feature vector;

[0217] The sample voltage attenuation difference features are used to form the second sample statistical feature vector;

[0218] The statistical feature vectors of the first and second samples are concatenated along the feature dimension to obtain the sample fusion feature.

[0219] The training system in this embodiment also includes a model validation module 106, which is used to validate the battery health state prediction model after obtaining it.

[0220] Model validation module 106 is used to divide the sample fusion features and sample battery state information into training set and validation set, or training set and test set.

[0221] Input the fused features of samples from the validation set or test set into the trained battery health status prediction model to obtain the predicted battery health status.

[0222] The performance of the battery health state prediction model is evaluated based on the error between the predicted battery health state and the actual battery health state information.

[0223] Among them, the performance evaluation indicators include at least one or more of the following: mean square error, root mean square error, and mean absolute error;

[0224] And / or,

[0225] The preset models include the XGBoost model; the XGBoost model uses a Bayesian optimization method for global search and tuning.

[0226] For the system embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The system embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separate, and the components as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs.

[0227] This disclosure extracts features based on the relaxation voltage time series after battery charging, enabling battery health status prediction without relying on complex disassembly or additional invasive detection. The data is easy to acquire and has strong engineering applicability. Furthermore, this disclosure simultaneously extracts statistical features reflecting global voltage distribution differences and deep features characterizing polarization processes at different time constants, and fuses them through a splicing method, thereby balancing the physical interpretability of the features and the ability to represent complex patterns, thus improving the completeness of feature expression.

[0228] Furthermore, this disclosure performs anomaly identification and processing on the sample training data before model training, which can reduce the adverse effects of capacity jumps, sensor noise, and abnormal relaxation voltage data on model training, and improve the stability and prediction accuracy of model training. This disclosure uses the XGBoost model to learn the fused features, which can make full use of the nonlinear mapping relationship between the fused features and the battery health status, thereby improving the health status prediction performance.

[0229] Example 4

[0230] like Figure 10 As shown, this embodiment provides a battery health status prediction system, which includes:

[0231] Actual data acquisition module 201 is used to acquire the actual relaxation voltage time series of the battery under test;

[0232] The actual feature extraction module 202 is used to extract a first actual statistical feature and a second actual feature based on the actual relaxation voltage time series; wherein, the first actual statistical feature is a statistical feature reflecting the difference in global voltage distribution in the battery, and the second actual feature is a feature characterizing the polarization process of different time constants inside the battery;

[0233] The actual feature fusion module 203 is used to splice and fuse the first actual statistical feature and the second actual feature to obtain the actual fused feature;

[0234] The health prediction module 204 is used to input the actual fused features into the battery health state prediction model of the first aspect of this disclosure, and output the actual health state prediction result of the battery under test.

[0235] For the system embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The system embodiments described above are merely illustrative, wherein the units described as separate components may or may not be physically separate, and the components as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of this disclosure according to actual needs.

[0236] The battery health status prediction system provided in this disclosure acquires the actual relaxation voltage time series of the battery under test and extracts a first actual statistical feature reflecting the differences in global voltage distribution and a second actual feature characterizing the polarization process of different time constants. This allows for a comprehensive and multi-dimensional characterization of the battery's internal state. By splicing and fusing the two types of features and inputting them into the battery health status prediction model, the dynamic information related to aging contained in the relaxation voltage can be effectively utilized, avoiding the problems of insufficient information from a single feature or the original voltage sequence and high noise interference. Furthermore, this disclosure does not require a complete charge-discharge test or disassembly analysis of the battery; it can achieve health status prediction based solely on a portion of the voltage response during the relaxation stage. This has the advantages of being fast, non-destructive, and easy to implement online, while also improving the accuracy and robustness of the prediction results.

[0237] Example 5

[0238] Figure 11 This is a schematic diagram of the structure of an electronic device shown in an example embodiment of the present disclosure. The electronic device includes a memory, a processor, and a computer program stored in the memory and used to run on the processor. When the processor executes the computer program, it implements the training method of the battery health state prediction model described in Embodiment 1, and / or the battery health state prediction method provided in Embodiment 2. Figure 11 The electronic device 110 shown is merely an example and should not impose any limitation on the functionality and scope of use of the embodiments disclosed herein.

[0239] like Figure 11 As shown, the electronic device 110 can be manifested as a general-purpose computing device, such as a server device. The components of the electronic device 110 may include, but are not limited to: at least one processor 111, at least one memory 112, and a bus 113 connecting different system components (including memory 112 and processor 111).

[0240] Bus 113 includes a data bus, an address bus, and a control bus.

[0241] The memory 112 may include volatile memory, such as random access memory (RAM) 1121 and / or cache memory 1122, and may further include read-only memory (ROM) 1123.

[0242] The memory 112 may also include a program tool 1125 (or utility) having a set (at least one) program module 1124, such program module 1124 including but not limited to: an operating system, one or more application programs, other program modules and program data, each or some combination of these examples may include an implementation of a network environment.

[0243] The processor 111 executes various functional applications and data processing by running computer programs stored in the memory 112, such as the training method of the battery health state prediction model described in any of the above embodiments 1, and / or the battery health state prediction method provided in embodiment 2.

[0244] Electronic device 110 can also communicate with one or more external devices 114 (e.g., keyboard, pointing device, etc.). This communication can be performed via input / output (I / O) interface 115. Furthermore, electronic device 110 can also communicate with one or more networks (e.g., local area network (LAN), wide area network (WAN), and / or public network, such as the Internet) via network adapter 116. As shown, network adapter 116 communicates with other modules of electronic device 110 via bus 113. It should be understood that, although not shown in the figures, other hardware and / or software modules can be used in conjunction with electronic device 110, including but not limited to: microcode, device drivers, redundant processors, external disk drive arrays, RAID (disk array) systems, tape drives, and data backup storage systems.

[0245] It should be noted that although several units / modules or sub-units / modules of the electronic device have been mentioned in the detailed description above, this division is merely exemplary and not mandatory. In fact, according to embodiments of this disclosure, the features and functions of two or more units / modules described above can be embodied in one unit / module. Conversely, the features and functions of one unit / module described above can be further divided and embodied by multiple units / modules.

[0246] Example 6

[0247] This disclosure also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the training method for the battery health state prediction model described in Embodiment 1 above, and / or the battery health state prediction method provided in Embodiment 2.

[0248] The readable storage medium may be more specifically adopted, including but not limited to: portable disk, hard disk, random access memory, read-only memory, erasable programmable read-only memory, optical storage device, magnetic storage device, or any suitable combination thereof.

[0249] Example 7

[0250] This disclosure also provides a computer program product, including a computer program that, when executed by a processor, implements the training method for the battery health state prediction model described in Embodiment 1 above, and / or the battery health state prediction method provided in Embodiment 2.

[0251] The program code for executing the computer program product of this disclosure can be written in any combination of one or more programming languages, and the program code can be executed entirely on a user device, partially on a user device, as a stand-alone software package, partially on a user device and partially on a remote device, or entirely on a remote device.

[0252] While specific embodiments of this disclosure have been described above, those skilled in the art should understand that these are merely illustrative examples, and the scope of protection of this disclosure is defined by the appended claims. Those skilled in the art can make various changes or modifications to these embodiments without departing from the principles and essence of this disclosure, but all such changes and modifications fall within the scope of protection of this disclosure.

Claims

1. A training method for a battery health status prediction model, characterized in that, The training method includes: Acquire several sets of sample training data, wherein any of the sample training data includes sample relaxation voltage time series and sample battery state information; Based on the sample relaxation voltage time series, first sample statistical features and second sample statistical features are extracted; the first sample statistical features are features that reflect the statistical differences in the global voltage distribution in the battery; the second sample statistical features are features of the different time constant planning processes inside the battery. The statistical features of the first sample and the statistical features of the second sample are concatenated and fused to obtain the sample fusion features; The preset model is trained based on the sample fusion features and the sample battery state information to obtain the battery health state prediction model for predicting the actual battery health state.

2. The training method for the battery health state prediction model according to claim 1, characterized in that, The sample relaxation voltage time series includes sample capacity data and sample voltage data at any time point. After the step of acquiring several sets of sample training data, and before the step of extracting the first sample statistical features and the second sample statistical features based on the sample relaxation voltage time series, the training method further includes: Anomaly identification and processing are performed on the sample relaxation voltage time series; The anomaly identification and processing includes: identifying anomalies in the sample size data based on a first preset algorithm, and repairing the identified anomalies in the sample size data using a median algorithm of adjacent normal cycles; And / or, The anomaly identification and processing includes: identifying abnormal relaxation voltage data based on a preset voltage threshold range and a voltage change threshold between adjacent sampling times, and repairing the abnormal relaxation voltage data by using the average relaxation voltage of the corresponding sampling times of adjacent normal cycles before and after the abnormal cycle.

3. The training method for the battery health state prediction model according to claim 1, characterized in that, The first sample statistical characteristics include: sample standard deviation and sample kurtosis. The step of extracting statistical features of the first sample based on the sample relaxation voltage time series includes: Based on the second preset algorithm, the standard deviation feature and the kurtosis feature of the sample relaxation voltage time series are extracted; And / or, The second sample statistical features include: the sample voltage attenuation difference characteristics within different time intervals; The step of extracting statistical features of the second sample based on the sample relaxation voltage time series includes: The sample relaxation voltage time series is input into a multi-scale feature extraction network to extract the sample voltage decay difference features.

4. The training method for the battery health state prediction model according to claim 3, characterized in that, The step of concatenating and fusing the first sample statistical features and the second sample statistical features to obtain the sample fused features includes: The sample standard deviation feature and the sample kurtosis feature are combined to form the first sample statistical feature vector; The sample voltage attenuation difference features are used to form a second sample statistical feature vector; The first sample statistical feature vector and the second sample statistical feature vector are concatenated along the feature dimension to obtain the sample fusion feature.

5. The training method for the battery health status prediction model according to any one of claims 1-4, characterized in that, After obtaining the battery health state prediction model, the method further includes validating the battery health state prediction model. The validation steps include: The sample fusion features and the sample battery state information are divided into a training set and a validation set, or a training set and a test set. The sample fusion features from the validation set or the test set are input into the trained battery health status prediction model to obtain the predicted battery health status. The performance of the battery health state prediction model is evaluated based on the error between the predicted battery health state and the actual battery health state information. The performance evaluation metrics include at least one or more of the following: mean square error, root mean square error, and mean absolute error. And / or, The preset model includes the XGBoost model; the XGBoost model is globally searched and tuned using a Bayesian optimization method.

6. A method for predicting battery health status, characterized in that, The prediction method includes: Obtain the actual relaxation voltage time series of the battery under test; Based on the actual relaxation voltage time series, a first actual statistical feature and a second actual feature are extracted; wherein, the first actual statistical feature is a statistical feature reflecting the difference in global voltage distribution in the battery, and the second actual feature is a feature characterizing the polarization process of different time constants inside the battery. The first actual statistical feature and the second actual feature are spliced ​​and fused together to obtain the actual fused feature; The actual fused features are input into the battery health status prediction model trained by the training method described in any one of claims 1 to 5, and the actual health status prediction result of the battery under test is output.

7. A training system for a battery health state prediction model, characterized in that, The training system includes: The sample data acquisition module is used to acquire several sets of sample training data, wherein any sample training data includes sample relaxation voltage time series and sample battery state information. The sample feature extraction module is used to extract a first sample statistical feature and a second sample statistical feature based on the sample relaxation voltage time series; the first sample statistical feature is a feature that reflects the statistical difference in the global voltage distribution in the battery; the second sample statistical feature is a feature of the different time constant planning processes inside the battery; The sample feature fusion module is used to concatenate and fuse the first sample statistical features and the second sample statistical features to obtain sample fusion features; The model training module is used to train a preset model based on the sample fusion features and the sample battery state information to obtain the battery health state prediction model for predicting the actual battery health state.

8. A battery health state prediction system, characterized in that, The prediction system includes: The actual data acquisition module is used to acquire the actual relaxation voltage time series of the battery under test; The actual feature extraction module is used to extract a first actual statistical feature and a second actual feature based on the actual relaxation voltage time series; wherein, the first actual statistical feature is a statistical feature reflecting the difference in global voltage distribution in the battery, and the second actual feature is a feature characterizing the polarization process of different time constants inside the battery; The actual feature fusion module is used to splice and fuse the first actual statistical feature and the second actual feature to obtain the actual fused feature; The health prediction module is used to input the actual fused features into the battery health state prediction model trained by the training method of any one of claims 1 to 5, and output the actual health state prediction result of the battery under test.

9. An electronic device comprising a memory, a processor, and a computer program stored in the memory and for running on the processor, characterized in that, When the processor executes the computer program, it implements the training method of the battery health state prediction model according to any one of claims 1 to 5, and / or the battery health state prediction method according to claim 6.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the training method of the battery health state prediction model according to any one of claims 1 to 5, and / or the battery health state prediction method according to claim 6.

11. A computer program product, comprising a computer program, characterized in that, When the computer program is executed by the processor, it implements the training method of the battery health state prediction model according to any one of claims 1 to 5, and / or the battery health state prediction method according to claim 6.

Images