Lead-acid battery life estimation method and system based on multi-parameter data model

Abstract

This invention relates to the field of lead-acid battery technology, and particularly to a method and system for estimating the lifespan of lead-acid batteries based on a multi-parameter data model. The method loads battery operation monitoring data, extracts characteristic frequency ranges, and obtains a list of key impedance frequency points. Based on this list, it performs charge-discharge characteristic correlation analysis to filter out a list of high-frequency correlated charge-discharge parameters. Combining physical parameters such as electrode materials and electrolyte characteristics with the aforementioned feature list, it collects first operating data tagged with cycle life to train a basic lifespan estimation model. For the target battery model, it collects corresponding second operating data and tags, performs transfer learning on the basic model, constructs a highly adaptable target lifespan estimation model, and executes the estimation task. This method integrates electrochemical impedance spectroscopy and dynamic operating condition characteristics to improve aging characterization capabilities. Through transfer learning, it enhances the model's generalization ability and is suitable for high-precision lifespan prediction of different battery models under small sample conditions.

Description

Technical Field

[0001] This invention relates to the field of lead-acid battery technology, and in particular to a method and system for estimating the lifespan of lead-acid batteries based on a multi-parameter data model. Background Technology

[0002] Lead-acid batteries, as one of the most widely used electrochemical energy storage devices, are extensively applied in electric vehicles, uninterruptible power supplies, communication base stations, and renewable energy systems. As their service life extends, battery performance gradually degrades, eventually leading to capacity decay, increased internal resistance, and even failure, severely impacting system safety and reliability. Therefore, accurately estimating the remaining lifespan of lead-acid batteries is crucial for enabling predictive maintenance and improving system operating efficiency.

[0003] Traditional battery life estimation methods mainly rely on capacity decay curve fitting or empirical models, which typically require long-term observation of discharge capacity changes, making it difficult to accurately capture early degradation trends. In recent years, electrochemical impedance spectroscopy (EIS) has become an important tool for battery state assessment due to its ability to non-destructively reflect internal charge transfer, diffusion processes, and interfacial reactions. The characteristic frequency response changes contained in EIS data are closely related to battery aging mechanisms, especially in reflecting typical failure modes such as negative electrode sulfation and positive electrode softening and shedding. However, single EIS features have poor adaptability to complex operating conditions and are easily affected by test conditions, making it difficult to independently support high-precision life prediction. Meanwhile, operating parameters such as voltage, current, temperature, and internal resistance change rate during dynamic charge and discharge processes can reflect the external behavior of the battery in real time, but lack the ability to deeply analyze the internal aging mechanisms.

[0004] The above content is only used to help understand the technical solution of the present invention and does not represent an admission that the above content is prior art. Summary of the Invention

[0005] The main objective of this invention is to provide a method and system for estimating the lifespan of lead-acid batteries based on a multi-parameter data model, aiming to solve the technical problems of insufficient integration of electrochemical impedance spectroscopy and dynamic charge-discharge characteristics, incomplete characterization of aging mechanisms, and poor cross-model adaptability in existing lead-acid battery lifespan estimation methods.

[0006] To achieve the above objectives, the present invention provides a method for estimating the lifespan of lead-acid batteries based on a multi-parameter data model, the method comprising:

[0007] Load the operational monitoring data of the target lead-acid battery, perform characteristic frequency range extraction, and obtain a list of key impedance frequency points;

[0008] Based on the operational monitoring data, the list of key impedance frequency points is traversed, and a charge-discharge characteristic correlation analysis is performed to obtain a list of high-frequency correlated charge-discharge parameters.

[0009] Using battery electrode material parameters, electrolyte characteristic parameters, the list of key impedance frequency points, and the list of high-frequency associated charge and discharge parameters as constraints, collect several one-to-one corresponding first battery operating condition time series data and first identifier cycle life labels to train the life estimation basic model.

[0010] Using the target lead-acid battery model, the list of key impedance frequency points, and the list of high-frequency associated charge and discharge parameters, collect several corresponding second battery operating condition time-series data and second identifiers of cycle life. Perform transfer learning on the life estimation basic model to obtain the target lead-acid battery life estimation model and execute the lead-acid battery life estimation task.

[0011] Optionally, the operational monitoring data of the loaded target lead-acid battery is used to extract characteristic frequency ranges to obtain a list of key impedance frequency points, including:

[0012] Constrained by battery electrode material parameters and electrolyte characteristic parameters, impedance spectrum test data are collected, impedance modulus-frequency curve characteristic point analysis is performed, and characteristic impedance frequency interval set is obtained.

[0013] From the operational monitoring data, extract the battery electrochemical impedance spectroscopy data, and statistically analyze the frequency points in each interval of the characteristic impedance frequency interval set whose impedance modulus change rate is greater than the change rate threshold and the impedance characteristic trigger frequency set.

[0014] The frequency points whose trigger frequencies are greater than or equal to the frequency threshold are added to the list of key impedance frequency points.

[0015] Optionally, the step of extracting battery electrochemical impedance spectroscopy data from the operational monitoring data and statistically analyzing the set of frequency points in each interval of the characteristic impedance frequency interval set where the rate of change of impedance modulus is greater than the rate of change threshold and the set of impedance characteristic trigger frequencies includes:

[0016] From the battery electrochemical impedance spectroscopy data, extract the first set of impedance spectroscopy data up to the Mth set of impedance spectroscopy data, where M represents the number of impedance spectroscopy data sets.

[0017] The first group of impedance spectrum data is divided into frequency intervals up to the Mth group of impedance spectrum data to obtain the first frequency interval impedance dataset up to the Kth frequency interval impedance dataset, where K represents the number of frequency intervals.

[0018] Delete the interval datasets with impedance data sample size less than the sample size threshold to obtain several retained frequency interval impedance datasets.

[0019] Traverse the several retained frequency interval impedance datasets, and based on the characteristic impedance frequency interval set, count the frequency points in each interval where the rate of change of impedance magnitude is greater than the rate of change threshold and the impedance characteristic trigger frequency set.

[0020] Optionally, the step of traversing the plurality of retained frequency interval impedance datasets, and based on the characteristic impedance frequency interval set, statistically analyzing the set of frequency points in each interval where the rate of change of impedance magnitude is greater than a rate of change threshold and the set of impedance characteristic trigger frequencies, includes:

[0021] From the set of characteristic impedance frequency intervals, extract the first characteristic frequency interval up to the Pth characteristic frequency interval, where P represents the number of characteristic frequency intervals;

[0022] Extract the impedance data of the first retention interval from the plurality of retention frequency interval impedance datasets;

[0023] Calculate the rate of change of impedance modulus within the first characteristic frequency interval, filter out frequency points with a rate of change greater than the rate of change threshold, and count their trigger frequencies, which are set as the trigger frequencies of the first interval.

[0024] Until the rate of change of impedance modulus within the Pth characteristic frequency interval is calculated, frequency points with a rate of change greater than the rate of change threshold are selected, and their trigger frequencies are statistically analyzed and set as the trigger frequencies of the Pth interval.

[0025] The frequency points corresponding to the first interval trigger frequency up to the Pth interval trigger frequency are added to the frequency point set, and the trigger frequency value is added to the impedance characteristic trigger frequency set.

[0026] Optionally, based on the operational monitoring data, the step of traversing the list of key impedance frequency points and performing charge-discharge characteristic correlation analysis to obtain a list of high-frequency correlated charge-discharge parameters includes:

[0027] Extract the first key frequency point from the list of key impedance frequency points;

[0028] Based on the first key frequency point, extract the first set of impedance-charge-discharge parameter correlation data from the operation monitoring data;

[0029] Based on the Pearson correlation coefficient threshold, feature filtering is performed on the first set of associated data to obtain a set of charging and discharging parameters that are strongly correlated with the impedance modulus.

[0030] The frequency of occurrence of the charging and discharging parameter set in multiple sets of associated data is statistically analyzed, and the charging and discharging parameters with a frequency greater than the frequency threshold are set as the first high-frequency associated charging and discharging parameters.

[0031] The process continues until all key frequency points have undergone correlation analysis, high-frequency correlated charge and discharge parameters are summarized, and a list of high-frequency correlated charge and discharge parameters is output.

[0032] Optionally, the step of collecting several one-to-one corresponding first battery operating condition time-series data and first cycle life labels, constrained by battery electrode material parameters, electrolyte characteristic parameters, the list of key impedance frequency points, and the list of high-frequency correlated charge and discharge parameters, and training a basic life estimation model, includes:

[0033] Based on the battery electrode material parameters and electrolyte characteristic parameters, a first-level constraint is constructed;

[0034] Based on the list of key impedance frequency points and the list of high-frequency associated charge and discharge parameters, a secondary constraint is constructed.

[0035] Load the battery sample set to be analyzed, wherein the battery sample set to be analyzed has the same electrode material to be analyzed, electrolyte characteristics to be analyzed, list of key frequency points to be analyzed, related charge and discharge parameters to be analyzed, battery operating impedance time series data, and charge and discharge dynamic monitoring time series information;

[0036] When the compositional similarity between the electrode material to be analyzed and the battery electrode material is greater than or equal to the compositional similarity threshold, and the conductivity deviation between the electrolyte characteristics to be analyzed and the electrolyte characteristic parameters is less than the deviation threshold, it is considered to satisfy the first-level constraint.

[0037] When the first-level constraint is met, and the frequency overlap between the list of key impedance frequency points and the list of key frequency points to be analyzed is greater than or equal to the overlap threshold, and the parameter matching degree between the list of high-frequency associated charge and discharge parameters and the associated charge and discharge parameters to be analyzed is greater than or equal to the matching threshold, it is considered that the second-level constraint is met.

[0038] When the secondary constraint is met, the battery operating impedance timing data and the charge / discharge dynamic monitoring timing information are set as the first battery operating condition timing data. The cycle life termination time is statistically analyzed for the battery sample set to be analyzed to obtain the first label identifying the cycle life.

[0039] Optionally, the step involves collecting a number of corresponding second battery operating condition time-series data and second cycle life labels using the target lead-acid battery model, the list of key impedance frequency points, and the list of high-frequency associated charge and discharge parameters. Transfer learning is then performed on the life estimation basic model to obtain the target lead-acid battery life estimation model. This life estimation basic model is a weighted integration of the outputs of multiple sub-models with different feature dimensions, including:

[0040] Obtain multiple life estimation sub-models of the aforementioned life estimation basic model;

[0041] An attention mechanism fusion layer is constructed to map the output features of multiple lifetime estimation sub-models to the input of the fusion layer;

[0042] Using the second battery operating condition time series data as input to multiple life estimation basic sub-models, and using the second label identifying cycle life as output of the attention mechanism fusion layer, the life estimation basic model is transferred to the life estimation basic model. By dynamically adjusting the sub-model weights to optimize the fusion strategy, the target lead-acid battery life estimation model is obtained.

[0043] Furthermore, to achieve the above objectives, the present invention also provides a lead-acid battery life estimation system based on a multi-parameter data model, the system comprising:

[0044] The feature extraction module is used to load the operation monitoring data of the target lead-acid battery, perform feature frequency range extraction, and obtain a list of key impedance frequency points.

[0045] The correlation analysis module is used to traverse the list of key impedance frequency points based on the operation monitoring data, perform charge and discharge characteristic correlation analysis, and obtain a list of high-frequency correlated charge and discharge parameters.

[0046] The model training module is used to collect several one-to-one corresponding first battery operating condition time series data and first identifier cycle life labels, with battery electrode material parameters, electrolyte characteristic parameters, the list of key impedance frequency points and the list of high-frequency related charge and discharge parameters as constraints, and to train the life estimation basic model.

[0047] The transfer modeling module is used to collect several corresponding second battery operating condition time-series data and second identifier cycle life labels based on the target lead-acid battery model, the list of key impedance frequency points, and the list of high-frequency associated charge and discharge parameters. It then performs transfer learning on the life estimation basic model to obtain the target lead-acid battery life estimation model and executes the lead-acid battery life estimation task.

[0048] Furthermore, to achieve the above objectives, the present invention also provides a lead-acid battery life estimation device based on a multi-parameter data model. The device includes: a memory, a processor, and a lead-acid battery life estimation program based on a multi-parameter data model stored in the memory and executable on the processor. The lead-acid battery life estimation program based on a multi-parameter data model is configured to implement the steps of the lead-acid battery life estimation method based on a multi-parameter data model as described above.

[0049] In addition, to achieve the above objectives, the present invention also provides a medium storing a lead-acid battery life estimation program based on a multi-parameter data model, wherein when the lead-acid battery life estimation program based on the multi-parameter data model is executed by a processor, the program implements the steps of the lead-acid battery life estimation method based on the multi-parameter data model as described above.

[0050] This invention provides a method for estimating the lifespan of lead-acid batteries based on a multi-parameter data model. The method integrates the intrinsic aging characteristics of electrochemical impedance spectroscopy with the external operational characteristics of dynamic charge-discharge, constructing a multi-dimensional, highly correlated health status characterization system that can more comprehensively and accurately reflect the battery's degradation mechanism under actual operating conditions. By extracting characteristic frequency ranges based on electrode materials and electrolyte properties, the physical interpretability and aging sensitivity of key impedance frequency points are enhanced. Correlation analysis is used to screen high-frequency charge-discharge parameters strongly correlated with key frequency points, achieving effective coupling of internal and external features. A two-stage modeling strategy combining a basic model and transfer learning is employed. Based on training a general model using common aging data, the model is fine-tuned using a small amount of measured data from the target battery model, significantly improving the model's generalization ability and cross-model adaptability, and solving the problem of low lifespan estimation accuracy under small sample conditions. The final target lifespan estimation model exhibits high accuracy, strong robustness, and good engineering applicability, providing reliable technical support for intelligent operation and maintenance and lifespan management of lead-acid batteries. Attached Figure Description

[0051] Figure 1 This is a flowchart illustrating an embodiment of the lead-acid battery life estimation method based on a multi-parameter data model according to the present invention.

[0052] Figure 2 This is a structural block diagram of an embodiment of the lead-acid battery life estimation system based on a multi-parameter data model according to the present invention.

[0053] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0054] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0055] Reference Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the lead-acid battery life estimation method based on a multi-parameter data model according to the present invention. An embodiment of the lead-acid battery life estimation method based on a multi-parameter data model according to the present invention is presented.

[0056] In one embodiment, the lead-acid battery life estimation method based on a multi-parameter data model includes:

[0057] Step S100: Load the operation monitoring data of the target lead-acid battery, perform characteristic frequency range extraction, and obtain a list of key impedance frequency points.

[0058] The target lead-acid battery can be a specific model of lead-acid battery for which life estimation is to be performed. It can be used as the evaluation object for the life estimation task, and its operational data and model information are used for model adaptation and prediction. Operational monitoring data can be time-series data reflecting the external behavior of the battery collected during dynamic charging and discharging. This data can be used to provide the external operational characteristics of the battery under actual operating conditions for correlation analysis and model training. In an exemplary embodiment, operational monitoring data can be obtained in real time through sensors by acquiring parameters such as voltage, current, temperature, and internal resistance change rate. Exemplarily, operational monitoring data can include, but is not limited to, constant current discharge voltage curves, pulse charging current responses, and temperature rise rate sequences. Battery electrode material parameters can be physicochemical parameters describing the composition, structure, and electrochemical characteristics of the positive and negative electrode active materials of the lead-acid battery. These can be used as constraints for extracting characteristic frequency ranges, ensuring that the selected frequency points match the material aging mechanism. Electrolyte characteristic parameters can be parameters characterizing the physicochemical properties of the electrolyte, such as concentration, density, ionic conductivity, and sulfate solubility. These can be used to define characteristic frequency ranges and influence the frequency response corresponding to ion migration and interfacial reaction kinetics. For example, electrolyte characteristic parameters may include, but are not limited to, sulfuric acid concentration gradient, electrolyte viscosity, ion transport number, etc.

[0059] The characteristic frequency range can be an electrochemical impedance spectroscopy frequency range that is sensitive to aging mechanisms, determined based on the characteristics of electrode materials and electrolytes. It can be used to define the search space for key impedance frequency points, improving physical interpretability and aging sensitivity. In one specific embodiment, the characteristic frequency range can be obtained by back-calculating the frequency range based on the time constants of charge transfer or diffusion processes corresponding to failure mechanisms such as negative electrode sulfation and positive electrode softening and shedding. For example, the characteristic frequency range may include, but is not limited to, the charge transfer-dominated frequency band, the double-layer charge-discharge frequency band, and the ion diffusion-restricted frequency band. The list of key impedance frequency points can be a set of discrete frequency points extracted from the characteristic frequency range that are highly sensitive to the battery's aging state. It can be used to characterize the aging characteristics of the internal electrochemical processes of the battery, serving as a core indicator of the intrinsic state of EIS. Furthermore, the list of key impedance frequency points can be obtained by selecting frequency points within the characteristic frequency range that have significant impedance magnitude or phase angle change rate. In an exemplary embodiment, the list of key impedance frequency points can be coupled with a list of high-frequency correlated charge-discharge parameters through correlation analysis, and used as one of the input features of the basic and target models for lifetime estimation. For example, the list of key impedance frequency points may include, but is not limited to, low-frequency diffusion characteristic points, mid-frequency interface response characteristic points, and high-frequency ohmic impedance characteristic points.

[0060] Loading the operational monitoring data of the target lead-acid battery can be done by reading historical or current operational monitoring data from a data storage unit or real-time acquisition system. Furthermore, this operation can provide raw input for subsequent feature extraction and correlation analysis. Performing characteristic frequency range extraction to obtain a list of key impedance frequency points can be done by determining the aging-sensitive EIS frequency range based on battery electrode material parameters and electrolyte characteristic parameters, and extracting representative frequency points from it. Further, this operation can be achieved by back-calculating the frequency corresponding to the time constant based on a physical model (such as an equivalent circuit) and selecting the peak response point, or by calculating the impedance change rate within a preset characteristic frequency range using a sliding window and selecting local extrema as key frequency points. This allows the key impedance frequency points to have clear physical meaning and high sensitivity to typical aging mechanisms.

[0061] Step S200: Based on the operation monitoring data, traverse the list of key impedance frequency points, perform charge and discharge characteristic correlation analysis, and obtain a list of high-frequency correlated charge and discharge parameters.

[0062] The correlation analysis of charge-discharge characteristics can be a statistical or machine learning method used to quantify the correlation strength between key impedance frequency points and various parameters in operational monitoring data. It can be used to screen external operational parameters strongly correlated with key impedance frequency points, achieving effective coupling of internal and external features. The high-frequency correlated charge-discharge parameter list can be a subset of dynamic charge-discharge operational parameters with strong statistical correlation to key impedance frequency points, selected through correlation analysis. It can be used to construct a mapping bridge between external behavior and internal aging state, enhancing the dimensionality and robustness of health status characterization. In a specific embodiment, the high-frequency correlated charge-discharge parameter list can extract candidate parameters from operational monitoring data, calculate their Pearson correlation coefficient, mutual information, or regression weight with key impedance frequency points, and retain parameters above a threshold. Furthermore, the high-frequency correlated charge-discharge parameter list can, together with the key impedance frequency point list, constitute multidimensional health status features and serve as joint input to the basic and target models for lifetime estimation. For example, the high-frequency correlated charge-discharge parameter list may include, but is not limited to, the voltage decay rate at the end of discharge, the charge acceptance index, and the slope of the instantaneous change in internal resistance.

[0063] By traversing the list of key impedance frequency points and performing charge-discharge characteristic correlation analysis, a list of high-frequency correlated charge-discharge parameters is obtained. This can be achieved by calculating the statistical correlation between each key impedance frequency point and various parameters in the operational monitoring data, and then filtering out strongly correlated parameters. Furthermore, this operation can be implemented by using linear correlation coefficients (such as Pearson's) for univariate screening, or by using LASSO regression or random forest feature importance ranking for multivariate joint screening. This allows for deep coupling between the internal electrochemical state and external dynamic behavior, constructing a highly correlated health state characterization system.

[0064] Step S300: Using battery electrode material parameters, electrolyte characteristic parameters, a list of key impedance frequency points, and a list of high-frequency related charge and discharge parameters as constraints, collect several one-to-one corresponding first battery operating condition time-series data and first label indicating cycle life, and train the life estimation basic model.

[0065] The first battery operating condition time-series data can be a general operating data sequence covering multiple battery models, used to train the life estimation base model. It can provide learning samples of common aging patterns to support the construction of a general model. In an exemplary embodiment, the first battery operating condition time-series data can be used to conduct long-term cycle tests on multiple models of lead-acid batteries under standard or diversified operating conditions and record the operational monitoring data. The first label identifying cycle life can be a supervision signal that corresponds one-to-one with the first battery operating condition time-series data and represents the remaining number of cycles or the proportion of capacity decay. It can be used as the target output for training the life estimation base model to achieve end-to-end mapping learning.

[0066] The basic lifespan estimation model can be an initial lifespan prediction model trained on common data from multiple models, possessing the ability to identify general aging patterns. It can be used to capture common degradation patterns across models, providing a prior knowledge base for subsequent transfer learning. In one specific embodiment, the basic lifespan estimation model can be trained using a list of key impedance frequency points and a list of high-frequency correlated charge-discharge parameters as input, with a label indicating the first cycle life as a supervisory signal, employing a deep neural network or ensemble learning method. Furthermore, the basic lifespan estimation model can serve as a starting point model for transfer learning, fine-tuned with a small amount of measured data from the target lead-acid battery.

[0067] Constrained by battery electrode material parameters, electrolyte characteristic parameters, a list of key impedance frequency points, and a list of high-frequency correlated charge-discharge parameters, several one-to-one corresponding time-series data of the first battery's operating conditions and a first label identifying cycle life are collected. This can be achieved by collecting operating data and corresponding life labels for multiple battery models under different operating conditions, while satisfying the aforementioned parameter constraints. Furthermore, this operation ensures that the training data covers common aging mechanisms, supporting the effective learning of the general model. The basic model for life estimation can be trained by inputting the time-series data of the first battery's operating conditions (including key impedance frequency points and high-frequency correlated parameters) into the model, using the first label identifying cycle life as a supervisory signal for parameter optimization. Furthermore, this operation can obtain an initial model with the ability to identify common aging across different battery models.

[0068] Step S400: Using the target lead-acid battery model, a list of key impedance frequency points, and a list of high-frequency associated charge and discharge parameters, collect several corresponding second battery operating condition time-series data and second identifier of cycle life labels, perform transfer learning on the life estimation basic model, obtain the target lead-acid battery life estimation model, and execute the lead-acid battery life estimation task.

[0069] The target lead-acid battery model can be the specific lead-acid battery product model identifier of the life estimation model to be deployed. It can be used to distinguish the differences in structure, materials, and processes of different batteries, guiding the data selection and model adaptation for transfer learning. The second battery operating condition time series data can be a small sequence of actual operating data collected for the target lead-acid battery model. It can be used to provide aging behavior samples unique to the target model for model fine-tuning. In a specific embodiment, the second battery operating condition time series data can be obtained by performing a limited number of cycle tests on the target model battery in the target application scenario or accelerated aging experiment and recording the operational monitoring data. The second label identifying cycle life can be a supervision signal that corresponds one-to-one with the second battery operating condition time series data and represents the remaining life of the target battery. It can be used as a fine-tuning target in the transfer learning stage to guide the base model to adapt to the characteristics of the target model.

[0070] The target lead-acid battery life estimation model can be a final life prediction model specifically designed for the target lead-acid battery model, fine-tuned through transfer learning. It can be used to perform specific life estimation tasks and output high-precision, robust remaining life prediction results. In an exemplary embodiment, the target lead-acid battery life estimation model can be fine-tuned based on the life estimation foundation model, utilizing second battery operating condition time-series data and second labeling of cycle life. Furthermore, the target lead-acid battery life estimation model can receive operational monitoring data of the target lead-acid battery as input and output life prediction values.

[0071] Using the target lead-acid battery model, a list of key impedance frequency points, and a list of high-frequency associated charge and discharge parameters, collect several corresponding second battery operating condition time-series data and second-identifier cycle life labels. This can be done by collecting a small amount of measured operating data and corresponding life labels within the same feature framework for the target model. Furthermore, this operation can provide suitable fine-tuning samples for the target model, ensuring the effectiveness of transfer learning. Transfer learning is then performed on the life estimation base model to obtain the target lead-acid battery life estimation model. This can be done by freezing some layer parameters of the base model, fine-tuning only the output layer or shallow network, and performing a small number of iterative training sessions using the second battery operating condition time-series data and labels. Further, this operation can be achieved by using a fine-tuning strategy to adjust all or part of the network weights, or by using a feature extraction + new classification head strategy to train only the newly added top-level prediction module. This can improve the model's adaptability to the target model under small sample conditions and avoid overfitting. The lead-acid battery life estimation task can be performed by inputting the real-time operating monitoring data of the target lead-acid battery into the target lead-acid battery life estimation model and outputting the predicted remaining life value. Furthermore, this operation enables high-precision and robust engineered lifetime prediction.

[0072] Taking the intelligent operation and maintenance of backup power supplies for communication base stations as an example, the lead-acid battery life estimation method based on a multi-parameter data model in this embodiment can be used for valve-regulated lead-acid battery packs deployed in communication base stations in remote areas. Regular EIS testing is performed, and voltage, current, and temperature data during charging and discharging are recorded simultaneously. The system first determines the characteristic frequency range (e.g., 10mHz–1Hz) based on the positive electrode PbO2 porosity and electrolyte density of the battery model, and extracts key impedance frequency points (e.g., the phase angle at 0.1Hz). Subsequent analysis reveals a high correlation between the voltage decay rate at the end of discharge and the 0.1Hz impedance, incorporating it into the high-frequency correlated charging and discharging parameter list. A life estimation model trained using historical data from multiple battery models is combined with a small amount of measured data from the base station's batteries over the past three months for transfer learning, generating a dedicated life estimation model. Maintenance personnel can then use this model to provide early warnings of battery replacement cycles, avoiding the risk of sudden power outages.

[0073] In one embodiment, operational monitoring data of the target lead-acid battery is loaded, characteristic frequency range extraction is performed, and a list of key impedance frequency points is obtained, including:

[0074] Constrained by battery electrode material parameters and electrolyte characteristic parameters, impedance spectrum test data are collected, impedance modulus-frequency curve characteristic point analysis is performed, and characteristic impedance frequency interval set is obtained.

[0075] From the operational monitoring data, extract the battery electrochemical impedance spectroscopy data, and statistically analyze the frequency points in each characteristic impedance frequency range where the rate of change of impedance modulus is greater than the threshold and the impedance characteristic trigger frequency set.

[0076] Add the frequency points whose impedance characteristic trigger frequencies are greater than or equal to the frequency threshold to the list of key impedance frequency points.

[0077] Impedance spectroscopy (ISS) data can be raw data on the change of complex impedance with frequency obtained by applying a small-signal AC excitation to a lead-acid battery under controlled conditions. This data can serve as the basis for constructing an impedance modulus-frequency curve to extract characteristic frequency ranges related to aging mechanisms. In this embodiment, ISS data can be obtained by applying a wideband sinusoidal perturbation and measuring the voltage and current response under static or specific charging states of the battery using an electrochemical workstation or an embedded EIS module. The impedance modulus-frequency curve can be a function curve characterizing the battery's frequency response, plotted with frequency on the x-axis and impedance modulus on the y-axis. This curve can be used to visualize the frequency response regions corresponding to different electrochemical processes, supporting feature point analysis. Furthermore, the impedance modulus-frequency curve can be obtained by taking the modulus of the impedance spectroscopy data and plotting it on a logarithmic frequency scale.

[0078] The characteristic impedance frequency range set can be a set of aging-sensitive frequency bands matched with the characteristics of electrode materials and electrolytes, obtained based on the characteristic point analysis of the impedance modulus-frequency curve. This set can be used to limit the effective range of subsequent dynamic EIS analysis, ensuring that the selected frequency bands correspond to the actual aging mechanism (such as low-frequency diffusion, mid-frequency interface reaction). In a specific embodiment, the characteristic impedance frequency range set can be obtained by identifying inflection points, extreme points, or abrupt slope changes on the impedance modulus-frequency curve, and defining physically meaningful frequency bands in conjunction with material parameters. For example, the characteristic impedance frequency range set may include, but is not limited to, one or more of the following: low-frequency Warburg diffusion range, mid-frequency charge transfer range, and high-frequency ohmic resistance range. Furthermore, the characteristic impedance frequency range set can serve as a range constraint for the statistical rate of change of impedance modulus, guiding the generation of the frequency point set and the impedance characteristic trigger frequency set.

[0079] Constrained by battery electrode material parameters and electrolyte characteristics, impedance spectroscopy (EIS) data is collected. This can be achieved by setting the frequency range, excitation amplitude, and state of charge conditions for EIS testing based on the battery's positive and negative electrode material structures and the electrolyte's physicochemical properties, and then performing impedance measurements. Furthermore, this operation can be further refined by setting a low-frequency lower limit based on the porosity of the positive electrode PbO2 to ensure coverage of the diffusion process; or by adjusting the excitation current amplitude based on the electrolyte's ionic conductivity to avoid polarization distortion. This allows the collected impedance spectroscopy data to focus on physical processes related to the material aging mechanism, improving the relevance of subsequent characteristic analyses.

[0080] The impedance modulus-frequency curve feature point analysis is performed to obtain a set of characteristic impedance frequency intervals. This can be achieved by differentiating or analyzing the impedance modulus-frequency curve to identify slope abrupt changes, inflection points, or plateau boundaries, and then dividing the frequency bands in conjunction with material parameters. Further, this operation can be achieved by using the second derivative zero-point detection method to locate characteristic corner frequencies; or by back-calculating the corresponding frequency bands for each process based on the fitting results of the equivalent circuit model, thereby generating a set of frequency intervals with clear electrochemical physical significance, providing a mechanism-guided search space for dynamic EIS analysis. Battery electrochemical impedance spectroscopy (EIS) data can be extracted from operational monitoring data, reflecting the dynamic EIS response of the target battery under actual operating conditions. It can be used to provide the impedance dynamics of individual batteries under real-world usage conditions, and to identify actual aging-sensitive points. In this embodiment, battery EIS data can be obtained by embedding small-amplitude high-frequency perturbation signals during charging and discharging, simultaneously acquiring voltage and current responses, and performing FFT or lock-in amplification processing. For example, battery EIS data can include, but is not limited to, one or more of the following: online EIS data, offline periodic EIS data, and transient impedance data under pulse perturbation.

[0081] The rate of change of impedance modulus can be the rate at which the battery impedance modulus changes with the number of cycles or time at a specific frequency point. It can be used to quantify the sensitivity of each frequency point to aging and serve as a criterion for screening key frequencies. In a specific embodiment, the rate of change of impedance modulus can be obtained by differentiating or fitting the slope of the impedance modulus at multiple time points at the same frequency point. The rate of change threshold can be a preset critical value for the rate of change of impedance modulus, used to determine whether a frequency point has sufficient aging sensitivity. It can be used to filter out frequency points with gradual changes and insensitivity to aging, while retaining high-response regions. The frequency point set can be the set of all frequency points within the characteristic impedance frequency range that satisfy the condition that the rate of change of impedance modulus is greater than the rate of change threshold. It can be used to initially screen candidate frequency points that have a significant response to aging.

[0082] The impedance characteristic trigger frequency set can be a set of frequency points within the characteristic impedance frequency range that generate a significant impedance response due to activation by the aging mechanism. Its triggering is determined by a rate-of-change threshold and can be used to identify frequency regions where aging-related electrochemical behavior actually occurs, serving as candidate sources for key frequency points. In an exemplary embodiment, the impedance characteristic trigger frequency set can be obtained by including frequency points in the frequency set whose corresponding impedance modulus change rate exceeds the rate-of-change threshold. Furthermore, the impedance characteristic trigger frequency set can be used as input to add its elements to the list of key impedance frequency points after being filtered by a frequency threshold. Exemplarily, the impedance characteristic trigger frequency set may include, but is not limited to, one or more of the following: sulfation trigger frequency set, positive electrode softening trigger frequency set, and electrolyte drying trigger frequency set.

[0083] Extracting battery electrochemical impedance spectroscopy (EIS) data from operational monitoring data can be achieved by simultaneously injecting small-signal perturbations during charge-discharge monitoring, separating the AC response components using signal processing techniques, and reconstructing the EIS data. Further, this operation can be accomplished by inserting a multi-frequency sinusoidal perturbation sequence into the constant-current discharge gap and extracting the impedance at each frequency point using lock-in amplification; or by using broadband pseudo-random signal excitation and reconstructing the frequency domain response through system identification methods. This allows for the acquisition of the battery's dynamic impedance response under real-world operating conditions, compensating for the disconnect between offline EIS and actual aging conditions. The statistical characteristic impedance frequency interval set includes the set of frequency points within each interval whose impedance magnitude change rate exceeds a threshold and the impedance characteristic trigger frequency set. This can be achieved by calculating the impedance magnitude change rate over multiple cycles for each frequency point within a characteristic interval, and selecting points exceeding the change rate threshold to form two sets. Furthermore, this operation can be achieved by fitting a linear degradation trend to each frequency point, using the absolute value of the slope as the rate of change; or by using a sliding window to calculate the impedance difference ratio between adjacent cycles as the rate of change, thereby identifying dynamic response frequencies that are sensitive to aging in actual operation and forming a candidate set of aging indicator frequencies with a high signal-to-noise ratio.

[0084] A frequency threshold can serve as the lowest frequency limit for further filtering of highly sensitive frequency points within the impedance characteristic trigger frequency set. It can also be used to exclude frequency points in excessively low frequency bands that are susceptible to noise or non-aging interference, improving the stability and measurability of critical frequency points. Frequency points with trigger frequencies greater than or equal to the frequency threshold within the impedance characteristic trigger frequency set are added to the critical impedance frequency point list. This can be achieved by traversing the impedance characteristic trigger frequency set, retaining elements with frequency values ​​not lower than the frequency threshold, and merging them into the critical impedance frequency point list. Furthermore, this operation can be further optimized by setting the frequency threshold to 10 MHz to exclude points in the ultra-low frequency band susceptible to temperature drift; or by dynamically setting the frequency threshold based on the EIS hardware measurement accuracy to ensure repeatability. This allows for the elimination of low-frequency points that are difficult to measure stably or are easily interfered with, while maintaining aging sensitivity, thereby improving the engineering practicality and robustness of critical frequency points.

[0085] Taking the health assessment of lead-acid batteries in a renewable energy storage system as an example, the lead-acid battery life estimation method based on a multi-parameter data model in this embodiment can be as follows: In a photovoltaic-lead-acid energy storage system, an online EIS test is automatically performed after each day's charging. The system first sets the EIS test frequency range to 1 mHz–1 kHz based on the specific surface area of ​​the negative electrode sponge lead and the electrolyte density of the battery model, and collects impedance spectrum data. By performing curvature analysis on the impedance modulus-frequency curve, two characteristic impedance frequency ranges are identified: 0.01–0.1 Hz (corresponding to the sulfation diffusion process) and 1–10 Hz (corresponding to the positive electrode interface reaction). Subsequently, daily EIS results are extracted from the operational monitoring data of the past 30 days, and the rate of change of impedance modulus at each frequency point is calculated. It is found that the rate of change at 0.03 Hz and 5 Hz consistently exceeds 5% / cycle, forming a set of impedance characteristic trigger frequencies. The frequency threshold is then set to 0.01 Hz, and both 0.03 Hz and 5 Hz are included in the list of key impedance frequency points. This list is then used for correlation analysis and life modeling, significantly improving the early warning capability for early capacity degradation.

[0086] In one embodiment, battery electrochemical impedance spectroscopy data is extracted from operational monitoring data, and the set of frequency points in each interval of the characteristic impedance frequency interval set whose impedance modulus change rate is greater than the change rate threshold and the set of impedance characteristic trigger frequencies are statistically analyzed, including:

[0087] From the battery electrochemical impedance spectroscopy data, extract the first set of impedance spectroscopy data up to the Mth set of impedance spectroscopy data, where M represents the number of impedance spectroscopy data sets.

[0088] The impedance spectrum data is divided into the first group up to the Mth group according to the frequency interval, and the impedance data datasets of the first frequency interval up to the Kth frequency interval are obtained, where K represents the number of frequency intervals.

[0089] Delete the interval datasets with impedance data sample size less than the sample size threshold to obtain several retained frequency interval impedance datasets.

[0090] Traverse several retained frequency range impedance datasets, and based on the characteristic impedance frequency range set, count the frequency points in each range where the rate of change of impedance magnitude is greater than the rate of change threshold and the impedance characteristic trigger frequency set.

[0091] The first set of impedance spectroscopy data can be the first batch of EIS measurement results extracted from the battery electrochemical impedance spectroscopy data in chronological order. This can be used as the starting data point for multi-cycle aging evolution analysis, participating in frequency band division and rate of change statistics. The Mth set of impedance spectroscopy data can be the Mth batch of EIS measurement results extracted from the battery electrochemical impedance spectroscopy data, where M is the total number of sets. This can be used to represent the impedance state of the battery in the later stages of service or at a specific moment, for constructing a time-series evolution sequence. For example, the Mth set of impedance spectroscopy data can be one or more of the following: end-cycle EIS data, accelerated aging endpoint spectrum, or randomly sampled EIS snapshots. M (number of impedance spectroscopy data sets) can be a preset total number of EIS data batches extracted from operational monitoring data for time-series analysis, used to determine the time resolution and coverage of aging trend statistics. Extracting the first set of impedance spectroscopy data up to the Mth set from the battery electrochemical impedance spectroscopy data can be done by grouping and slicing the battery electrochemical impedance spectroscopy data according to timestamps or cycle numbers, forming M ordered EIS snapshots. Furthermore, this operation can be achieved by extracting a set of EIS data at fixed cycle intervals (e.g., every 10 cycles), or by triggering EIS acquisition and grouping based on running events (e.g., after deep discharge), thereby constructing a time-series EIS sequence that reflects the battery aging evolution and supports dynamic rate of change analysis.

[0092] The first frequency interval impedance dataset can be a subset of the impedance samples belonging to the first characteristic frequency interval from all M groups of impedance spectrum data. It can be used to support independent statistical analysis of frequency bands related to a specific aging mechanism. In an exemplary embodiment, the first frequency interval impedance dataset can be one or more of a low-frequency diffusion interval dataset, a mid-frequency interface reaction interval dataset, or a high-frequency ohmic interval dataset. The Kth frequency interval impedance dataset can be a subset of the impedance samples belonging to the Kth characteristic frequency interval from all M groups of impedance spectrum data. It can be used to achieve parallel processing and independent evaluation of multiple aging-sensitive frequency bands. K (number of frequency intervals) can be the total number of independent frequency bands contained in the characteristic impedance frequency interval set, and can be used to determine the granularity of frequency band division and the ability to decouple multiple mechanisms. Dividing the first group of impedance spectrum data up to the Mth group of impedance spectrum data according to frequency intervals to obtain the first frequency interval impedance dataset up to the Kth frequency interval impedance dataset can be done by assigning each group of EIS data to the corresponding interval according to frequency based on the K frequency band boundaries of the characteristic impedance frequency interval set, and aggregating them to form K time-spanning frequency band datasets. Furthermore, this operation can be achieved by performing frequency band masking on each set of EIS data and assigning it to the corresponding interval set, or by establishing a frequency-interval mapping table to redistribute all impedance points in batches. This enables frequency band decoupling guided by aging mechanism, allowing the evolution of different electrochemical processes to be tracked independently.

[0093] The sample size threshold can be a preset minimum effective impedance sample count limit, used to determine whether a frequency range dataset has statistical reliability. It can be used to filter out sparse data ranges caused by missing tests, signal loss, or abnormal operating conditions, improving the robustness of subsequent analysis. In a specific embodiment, the sample size threshold can be one or more of the following: a fixed minimum sample size (e.g., 10 times), a dynamic threshold calculated based on confidence intervals, or a relative threshold set according to the proportion of total data. The retained frequency range impedance dataset can be a subset of impedance data from each frequency range that has sufficient statistical significance after removing ranges with insufficient sample size. It can be used to ensure that subsequent rate of change calculations are based on reliable data, avoiding noise or missing values ​​interfering with aging trend identification. In this embodiment, the retained frequency range impedance dataset can be one or more of the following: a high-completeness low-frequency dataset, a stable mid-frequency response dataset, or a continuous high-frequency observation dataset. Furthermore, the retained frequency range impedance dataset can be obtained by counting the number of impedance samples in each frequency range impedance dataset and retaining only the set ≥ the sample size threshold. For example, the retained frequency range impedance dataset can be used as a traversal object to statistically analyze the rate of change of impedance magnitude and generate a set of frequency points and a set of impedance characteristic trigger frequencies.

[0094] Deleting impedance data sets with sample sizes less than a threshold results in several retained frequency range impedance datasets. This can be achieved by counting the number of valid samples in each frequency range impedance dataset and removing sets below the threshold. Further, this operation can be implemented by setting the sample size threshold to 50% of the total number of groups M to ensure valid observations in more than half of the periods, or by using a sliding window validation method to retain only ranges with data in N consecutive cycles. This eliminates sparse or unreliable frequency bands, improving the statistical stability and reliability of subsequent rate of change estimation. Iterating through several retained frequency range impedance datasets, based on the characteristic impedance frequency range set, the set of frequency points within each range whose impedance magnitude change rate exceeds the rate of change threshold and the set of impedance characteristic trigger frequencies are statistically analyzed. This can be done by calculating the impedance magnitude change rate at M time points for all frequency points within each retained range, and selecting points exceeding the rate of change threshold to form two sets. Furthermore, this operation can be achieved by fitting a linear regression slope to each frequency point as the rate of change and taking the absolute value to compare with the threshold, or by calculating the moving average rate of change of the impedance difference between adjacent groups to identify a continuous upward / downward trend. This allows for the accurate identification of aging-sensitive dynamic response frequencies from high-quality time-series EIS data, generating a high signal-to-noise ratio key feature candidate set.

[0095] For example, in the scenario of health management of lead-acid batteries in electric vehicle start-stop systems, the lead-acid battery life estimation method based on a multi-parameter data model in this embodiment can be as follows: A 12V start-stop battery automatically performs an online EIS test once a week during vehicle operation, accumulating M=24 sets of impedance spectrum data. Based on the previously determined characteristic impedance frequency interval set (including K=3 intervals: 0.01–0.1Hz, 0.5–2Hz, 50–100Hz), the system divides each set of EIS data into three frequency band subsets. Upon inspection, the 0.01–0.1Hz interval was found to have only 18 valid samples due to low temperature causing some test failures, which is lower than the set sample size threshold (20), and was therefore removed; the other two intervals were retained to form a retained frequency interval impedance dataset. Subsequently, the system calculated the monthly change rate of impedance modulus at each frequency point in the 0.5–2Hz interval, and found that the change rate at 1.2Hz reached 8% / month, exceeding the 5% change rate threshold, so it was added to the frequency point set and marked as the impedance characteristic trigger frequency related to sulfation. This frequency point is ultimately included in the list of critical impedance frequency points for use as input in the lifetime model.

[0096] In one embodiment, several retained frequency interval impedance datasets are traversed. Based on the characteristic impedance frequency interval set, the set of frequency points in each interval where the rate of change of impedance magnitude is greater than the rate of change threshold and the set of impedance characteristic trigger frequencies are statistically analyzed, including:

[0097] From the set of characteristic impedance frequency intervals, extract the first characteristic frequency interval up to the Pth characteristic frequency interval, where P represents the number of characteristic frequency intervals;

[0098] The first characteristic frequency interval can be the first aging mechanism-related frequency band extracted from the characteristic impedance frequency interval set. It can be used as the starting unit for interval-by-interval analysis to calculate the rate of change of impedance modulus. In this embodiment, the first characteristic frequency interval can include, but is not limited to, one or more of the following: the negative electrode sulfation-dominant frequency band, the positive electrode interface reaction frequency band, and the electrolyte diffusion-limiting frequency band. The Pth characteristic frequency interval can be the Pth aging-sensitive frequency band extracted from the characteristic impedance frequency interval set, where P is the total number of intervals. It can be used to cover the aging response region of a specific electrochemical process and support parallel analysis of multiple mechanisms. P (the number of characteristic frequency intervals) can be the total number of independent frequency bands contained in the characteristic impedance frequency interval set, which can be used to determine the granularity and coverage of the decoupled analysis of aging mechanisms. Extracting the first characteristic frequency interval up to the Pth characteristic frequency interval from the characteristic impedance frequency interval set can be done by sequentially traversing all P frequency bands in the characteristic impedance frequency interval set and extracting them one by one for subsequent analysis. Furthermore, this operation can be implemented through structured traversal, thereby enabling structured and ordered processing of frequency bands corresponding to different aging mechanisms.

[0099] Extract the impedance data for the first retention interval from several retention frequency interval impedance datasets;

[0100] The first retained frequency range impedance data can be a high-quality subset of impedance time-series data corresponding to the first characteristic frequency range extracted from several retained frequency range impedance datasets. This subset can be used as the basic input for calculating the rate of change and triggering frequency statistics within this frequency band. Extracting the first retained frequency range impedance data from several retained frequency range impedance datasets can be done by matching and extracting the corresponding impedance time-series subset from the retained datasets based on the currently processed characteristic frequency range (such as the first range). Furthermore, this operation can be implemented through a frequency band alignment mechanism, thereby ensuring strict alignment between the analysis data and the physical mechanism frequency band, avoiding cross-range contamination.

[0101] Calculate the rate of change of impedance modulus within the first characteristic frequency interval, filter out frequency points with a rate of change greater than the rate of change threshold, and count their trigger frequencies, which are set as the trigger frequencies of the first interval.

[0102] The frequency point can be a discrete test point with a specific frequency value within a certain characteristic frequency range. It can be used as the basic unit for calculating the rate of change of impedance modulus and as a candidate key frequency point. The trigger frequency can be the number of times a frequency point meets the condition of "rate of change of impedance modulus greater than the rate of change threshold" in M ​​sets of EIS data. It can be used to quantify the sensitivity and stability of the frequency point to aging. A high trigger frequency indicates that its response has reproducibility and reliability. In an exemplary embodiment, the trigger frequency can be obtained by traversing M sets of data and counting the number of cycles in which the rate of change of each frequency point exceeds the threshold. For example, the trigger frequency can be one or more of the following, including but not limited to low-frequency sulfation trigger frequency, medium-frequency positive electrode shedding trigger frequency, and high-frequency contact resistance degradation trigger frequency. The first interval trigger frequency can be the trigger frequency value corresponding to the high rate of change frequency point obtained after screening within the first characteristic frequency range. It can be used to characterize the stable response intensity of the most representative aging-sensitive point in this range.

[0103] Calculate the rate of change of impedance magnitude within the first characteristic frequency range, filter frequency points whose rate of change exceeds a threshold, and statistically analyze their trigger frequencies, setting them as the first interval trigger frequencies. This can be achieved by calculating the rate of change of impedance magnitude for each frequency point in the first retained impedance data range across M sets of data, retaining points exceeding the threshold, and counting the number of times each point is triggered as the first interval trigger frequency. Further, this operation can be performed by fitting a linear degradation slope to each frequency point, taking the absolute value as the rate of change, and using the trigger frequency as the number of cycles in which the slope exceeds the threshold; or by using a sliding window to calculate the rate of change between adjacent cycles, counting N consecutive times exceeding the threshold as one valid trigger. This allows for the identification of representative frequency points and their response strengths that are both sensitive and stable within a specific aging mechanism frequency band.

[0104] Until the rate of change of impedance modulus within the Pth characteristic frequency interval is calculated, frequency points with a rate of change greater than the rate of change threshold are selected, and their trigger frequencies are statistically analyzed and set as the trigger frequencies of the Pth interval.

[0105] The trigger frequency of the Pth interval can be the trigger frequency value of the corresponding high rate of change frequency point obtained after screening within the Pth characteristic frequency interval. It can be used to reflect the dynamic activation degree of the Pth aging mechanism in actual operation. The rate of change of impedance modulus within the Pth characteristic frequency interval is calculated, frequency points with a rate of change greater than the rate of change threshold are screened, and their trigger frequencies are statistically analyzed and set as the trigger frequency of the Pth interval. This can be done in the same way as above, but applied to the Pth characteristic frequency interval and its corresponding retained data. Furthermore, this operation can be implemented through a mechanism consistent with the above, thereby enabling parallel sensitive point mining and stability quantification for all aging-related frequency bands.

[0106] Add the frequency points corresponding to the trigger frequencies from the first interval up to the Pth interval into the frequency point set, and add the trigger frequency value into the impedance characteristic trigger frequency set.

[0107] Adding the frequency points corresponding to the trigger frequencies of the first interval up to the Pth interval into the frequency point set can be achieved by collecting all high-rate-of-change frequency points selected from all P intervals and merging them into a unified frequency point set. Furthermore, this operation can be implemented through set merging, thereby constructing a candidate set of highly sensitive frequencies covering multiple aging mechanisms. Adding the trigger frequency values ​​of the first interval up to the Pth interval into the impedance characteristic trigger frequency set can be achieved by storing the corresponding trigger frequency values ​​of each interval sequentially into the impedance characteristic trigger frequency set, corresponding one-to-one with the frequency points. Furthermore, this operation can be implemented through a mapping storage mechanism, thereby establishing a quantitative correlation between frequency points and their aging response stability, supporting subsequent feature weighting or selection.

[0108] For example, in the scenario of health assessment of backup lead-acid batteries for communication base stations, the lead-acid battery life estimation method based on a multi-parameter data model in this embodiment can be as follows: The system has determined that the characteristic impedance frequency interval set includes P = 2 intervals: 0.02–0.08Hz (negative electrode sulfation) and 1–5Hz (positive electrode softening). Impedance time series data corresponding to the two intervals are extracted from the retained dataset. In the 0.02–0.08Hz interval, the 0.05Hz point has an impedance modulus monthly change rate exceeding 6% 18 times in 24 EIS sets, with a trigger frequency of 18; in the 1–5Hz interval, the 3Hz point has a change rate exceeding the threshold 20 times, with a trigger frequency of 20. The system adds 0.05Hz and 3Hz to the frequency point set, and stores 18 and 20 as their trigger frequency values ​​in the impedance characteristic trigger frequency set. In subsequent modeling, 3Hz is given greater weight due to its higher trigger frequency, reflecting its more stable positive electrode aging indication role in this type of battery.

[0109] In one embodiment, based on operational monitoring data, a list of key impedance frequency points is traversed, and a charge-discharge characteristic correlation analysis is performed to obtain a list of high-frequency correlated charge-discharge parameters, including:

[0110] Extract the first critical frequency point from the list of critical impedance frequency points;

[0111] The first key frequency point can be a single frequency point selected sequentially from the list of key impedance frequency points, used for point-by-point correlation analysis. In this embodiment, the first key frequency point can serve as the reference frequency for a single correlation analysis, ensuring that the aging information of each key frequency point is independently and fully extracted. For example, the first key frequency point can be one or more of the following: low-frequency diffusion dominance point, mid-frequency interface reaction point, and high-frequency ohmic response point. Extracting the first key frequency point from the list of key impedance frequency points can be achieved by selecting a frequency point from the list in sequence or by priority as the current analysis object. Furthermore, this operation can be implemented through sequential traversal, thereby enabling point-by-point refined processing of key frequency points and avoiding the overall averaging from masking local sensitive features.

[0112] Based on the first key frequency point, the first set of impedance-charge and discharge parameter correlation data is extracted from the operation monitoring data;

[0113] The first set of impedance-charge-discharge parameter correlation data can be a multi-dimensional sample set composed of impedance magnitude values ​​and various charge-discharge parameters in the operational monitoring data, synchronously collected at the EIS test time or aging stage corresponding to the first key frequency point. In an exemplary embodiment, the first set of impedance-charge-discharge parameter correlation data can align the impedance magnitude values ​​within the same time window or the same cycle with dynamic parameters such as voltage, current, and temperature to form observation samples, providing structured input for statistical correlation analysis and ensuring consistency of internal and external characteristics over time or operating conditions. The impedance magnitude value can be the amplitude of the complex impedance measured in the electrochemical impedance spectrum at a specific frequency. Furthermore, the impedance magnitude value can serve as a quantitative indicator of the intrinsic aging state of the EIS, used to establish a correlation with external operating parameters. Exemplarily, the impedance magnitude value can include, but is not limited to, one or more of low-frequency impedance magnitude values, mid-frequency impedance magnitude values, and high-frequency impedance magnitude values.

[0114] Based on the first key frequency point, the first set of impedance-charge / discharge parameter correlation data is extracted from the operational monitoring data. This can be achieved by aligning the impedance magnitude corresponding to the first key frequency point with the operational monitoring data within the same aging stage or test cycle to construct a multi-dimensional sample. Furthermore, this operation can be achieved by extracting the average charge / discharge parameters within a fixed time window before and after the EIS test time, or by treating each charge / discharge cycle as a sample unit and matching it with the impedance magnitude obtained from the EIS test within that cycle. This ensures that internal and external features are aligned in the physical process, improving the reliability of the correlation analysis.

[0115] Based on the Pearson correlation coefficient threshold, feature filtering is performed on the first set of associated data to obtain a set of charging and discharging parameters that are strongly correlated with the impedance modulus.

[0116] The charging / discharging parameter set can be a subset of charging / discharging operating parameters that are strongly linearly correlated with the current impedance modulus after being filtered by a Pearson correlation coefficient threshold. In one embodiment, the charging / discharging parameter set can be obtained by calculating the Pearson correlation coefficient between each charging / discharging parameter and the impedance modulus, and retaining parameters whose absolute values ​​are greater than a preset threshold. This is used to initially filter weakly correlated or noisy parameters and focus on external behavioral features with statistical significance. The Pearson correlation coefficient threshold can be a preset statistical threshold used to determine whether the charging / discharging parameters and the impedance modulus have a strong linear correlation. Furthermore, the Pearson correlation coefficient threshold can control the strictness of feature screening and balance the number of features and the correlation strength. For example, the Pearson correlation coefficient threshold can be a high correlation threshold (e.g., 0.7), a medium correlation threshold (e.g., 0.5), an adaptive correlation threshold, etc. Based on the Pearson correlation coefficient threshold, feature screening is performed on the first set of associated data to obtain a charging / discharging parameter set that is strongly correlated with the impedance modulus. This can be achieved by calculating the Pearson correlation coefficient between each charging / discharging parameter and the impedance modulus, and retaining parameters whose absolute values ​​are greater than the threshold. Furthermore, this operation can be achieved by setting fixed or dynamic thresholds, thereby eliminating weakly correlated or noisy parameters and focusing on statistically significant external behavioral indicators.

[0117] The frequency of occurrence of the charge and discharge parameter set in multiple sets of associated data is statistically analyzed, and the charge and discharge parameters with a frequency greater than the frequency threshold are set as the first high-frequency associated charge and discharge parameters.

[0118] The frequency threshold can be the minimum number of occurrences or a proportion threshold used to determine whether a certain charge / discharge parameter is stably present in multiple sets of correlated data. In an exemplary embodiment, the frequency threshold can ensure that the selected parameter has robustness across frequency points and operating conditions, avoiding overfitting caused by accidental strong correlations. For example, the frequency threshold may include, but is not limited to, absolute frequency thresholds, relative proportion thresholds, dynamic sliding window frequency thresholds, etc. The first high-frequency correlated charge / discharge parameter can be a stable, strongly correlated charge / discharge parameter whose frequency exceeds the frequency threshold in the correlation analysis of multiple key frequency points. Further, the first high-frequency correlated charge / discharge parameter can serve as a component of the final high-frequency correlated charge / discharge parameter list, representing external characteristics with universality and mechanistic consistency. For example, the first high-frequency correlated charge / discharge parameter may include, but is not limited to, one or more of the following: discharge plateau voltage slope, charging end current decay rate, peak temperature rise rate, etc.

[0119] Statistical analysis of the frequency of occurrence of charge / discharge parameter sets across multiple sets of correlated data can be achieved by iterating through all key frequency points and accumulating the number of times each parameter is selected. Furthermore, this operation can be implemented using hash counting or a frequency mapping table, thereby identifying external characteristics that remain stable under various aging mechanisms and enhancing the generalization ability of the parameters. Setting charge / discharge parameters with frequencies exceeding a frequency threshold as the first high-frequency correlated charge / discharge parameters can be achieved by comparing the cumulative frequency of each parameter with a preset frequency threshold and retaining the qualified parameters. Furthermore, this operation can be implemented using threshold comparison logic, ensuring that the final selected parameters possess both strong single-point correlation and high multi-point stability.

[0120] The process continues until all key frequency points have undergone correlation analysis, high-frequency correlated charge and discharge parameters are summarized, and a list of high-frequency correlated charge and discharge parameters is output.

[0121] Correlation analysis of all key frequency points can be performed by iteratively executing the extraction, correlation, filtering, and statistical steps described above until all frequency points in the key impedance frequency point list have been processed. Furthermore, this operation can be implemented using an iterative control structure, thereby achieving comprehensive coverage of the external feature coupling of all intrinsic aging-sensitive frequencies. Summarizing the high-frequency correlated charge-discharge parameters and outputting a high-frequency correlated charge-discharge parameter list can be achieved by deduplicating and merging all first-level high-frequency correlated charge-discharge parameters to form the final parameter list. Furthermore, this operation can be implemented using set merging and deduplication algorithms, thereby generating structured, high-confidence, multi-dimensional health status characterization inputs to support subsequent modeling.

[0122] Taking the battery health management of a renewable energy storage system as an example, the lead-acid battery life estimation method based on a multi-parameter data model in this embodiment can be implemented in a photovoltaic energy storage station where lead-acid batteries undergo daily charge-discharge cycles and periodically undergo EIS testing. The system sequentially extracts the first key frequency point (e.g., 0.1Hz) from a list of key impedance frequency points (e.g., 0.01Hz, 0.1Hz, 10Hz) and matches it with data such as the slope of the voltage curve, charging current, and temperature rise rate recorded during the daily charge-discharge cycle to form the first set of impedance-charge-discharge parameter correlation data. The Pearson correlation coefficient between each parameter and the 0.1Hz impedance modulus is calculated, and parameters with an absolute value of correlation coefficient > 0.6 (e.g., voltage decay rate at the end of discharge) are retained. After repeating this process for all frequency points, it is found that the voltage decay rate at the end of discharge is selected in all three frequency points, with a frequency of 3, exceeding the frequency threshold of 2, and is therefore set as the first high-frequency correlated charge-discharge parameter. The final result is a list of high-frequency correlated charge and discharge parameters containing three parameters, which is used to fine-tune the lifetime estimation model and significantly improve the ability to identify the early stage of sulfation.

[0123] In one embodiment, constrained by battery electrode material parameters, electrolyte characteristic parameters, a list of key impedance frequency points, and a list of high-frequency associated charge and discharge parameters, several one-to-one corresponding first battery operating condition time-series data and first label indicating cycle life are collected to train a lifespan estimation basic model, including:

[0124] Based on the battery electrode material parameters and electrolyte characteristic parameters, a first-level constraint is constructed;

[0125] The primary constraint can be a physical consistency condition set based on battery electrode material parameters and electrolyte characteristic parameters, used to screen battery samples with similar chemical systems. It ensures the comparability of training data at the level of electrode material composition and electrolyte physicochemical properties, avoiding confusion about aging mechanisms due to differences in material systems. In this embodiment, the primary constraint's operating principle can be explained in context, i.e., by setting component similarity thresholds and deviation thresholds to form screening rules at the material and electrolyte levels. For example, the primary constraint can include, but is not limited to, one or more of the following: positive electrode active material component matching constraint, negative electrode alloy structure similarity constraint, and electrolyte ionic conductivity consistency constraint. Battery electrode material parameters can be benchmark data used to characterize the composition and structural features of the electrode materials used in standard batteries, providing a basis for comparison of the electrode materials to be analyzed. The electrode materials to be analyzed can be the actual composition and structural description of the electrode materials used in the battery sample to be analyzed, which can be used to compare the composition similarity with the battery electrode material parameters to determine whether the primary constraint is met. In one exemplary embodiment, the similarity between the electrode material to be analyzed and the standard parameters can be obtained by calculating the cosine similarity of the material composition vector or by a composition matching rule engine based on expert rules.

[0126] Electrolyte characteristic parameters can serve as benchmark data characterizing the physicochemical properties of electrolytes used in standard batteries, providing a basis for consistency verification of the characteristics of the electrolyte to be analyzed. The characteristics of the electrolyte to be analyzed can be the physicochemical properties of the actual electrolyte in the battery sample to be analyzed, which can be used to calculate the conductivity deviation with the electrolyte characteristic parameters to determine whether the first-level constraints are met. Furthermore, the conductivity deviation can be the absolute or relative difference in conductivity between the characteristics of the electrolyte to be analyzed and the characteristic parameters of the standard electrolyte, which can be used to quantify the degree of consistency of the physicochemical properties of the electrolyte. The deviation threshold can be the upper limit of the allowable conductivity deviation, used to determine whether the electrolyte meets the first-level constraints, and can be used as another criterion for determining the first-level constraints, limiting the acceptable fluctuation range of electrolyte characteristics. The composition similarity threshold can be the minimum composition similarity limit for determining whether the electrode material to be analyzed and the electrode material parameters of the standard battery belong to the same chemical system, and can be used as one of the criterion for determining the first-level constraints, controlling the comparability range at the material level. Based on the battery electrode material parameters and electrolyte characteristic parameters, a primary constraint is constructed, which can be achieved by setting component similarity thresholds and deviation thresholds, forming screening rules at the material and electrolyte levels. Furthermore, this operation can be implemented by using cosine similarity calculations of material component vectors or by using absolute deviations (such as conductivity differences less than a set threshold), thereby establishing an entry threshold for chemical system consistency and eliminating aging path interference caused by material differences.

[0127] Based on the list of key impedance frequency points and the list of high-frequency associated charge and discharge parameters, a second-level constraint is constructed.

[0128] The secondary constraints can be correlation conditions set based on a list of key impedance frequency points and a list of high-frequency associated charge and discharge parameters. These conditions are used to screen battery samples with highly aligned feature spaces. They ensure a strong coupling relationship between the selected samples' EIS response characteristics and external dynamic behavior, enhancing the causal correlation between input features and aging state. In a specific embodiment, the operating principle of the secondary constraints can be explained in context, i.e., by setting overlap and matching thresholds to form feature space alignment screening rules. For example, the secondary constraints may include, but are not limited to, one or more of the following: impedance frequency overlap constraints, charge and discharge parameter matching constraints, and timing response synchronicity constraints.

[0129] The critical impedance frequency point list can be a set of frequency points reflecting the aging-sensitive frequency bands of a standard battery, serving as a comparison benchmark for the critical frequency point list to be analyzed. The critical frequency point list to be analyzed can be a set of frequency points extracted from the EIS data of the battery sample to be analyzed, reflecting its aging state. It can be used to calculate the frequency overlap with the critical impedance frequency point list to determine whether the second-order constraints are met. The frequency overlap can be the intersection ratio or weighted overlap between the critical impedance frequency point list and the critical frequency point list to be analyzed on the frequency axis, and can be used to measure the consistency between the two sets of frequency points in the aging-sensitive frequency bands. The overlap threshold can be the minimum acceptable standard for determining whether the frequency overlap meets the second-order constraints, and can be used to ensure that the selected samples have sufficient alignment in the EIS feature dimensions.

[0130] The high-frequency correlated charge-discharge parameter list can be a set of charge-discharge characteristic parameters strongly correlated with the high-frequency impedance response, serving as a matching benchmark for the correlated charge-discharge parameters to be analyzed. The correlated charge-discharge parameters to be analyzed can be a set of candidate charge-discharge characteristic parameters extracted from the operational monitoring data of the battery sample to be analyzed. These parameters can be used to evaluate the parameter matching degree with the high-frequency correlated charge-discharge parameter list, determining whether the secondary constraints are met. The parameter matching degree refers to the similarity between the high-frequency correlated charge-discharge parameter list and the correlated charge-discharge parameters to be analyzed in terms of parameter type, quantity, and statistical distribution. It can be used to assess the consistency of external operating characteristics and support the determination of secondary constraints. The matching degree threshold is the minimum acceptable standard for determining whether the parameter matching degree meets the secondary constraints, ensuring the comparability of charge-discharge behavior characteristics.

[0131] Based on the list of key impedance frequency points and the list of high-frequency associated charge and discharge parameters, a secondary constraint is constructed. This constraint can be achieved by setting overlap and matching thresholds, forming a screening rule for feature space alignment. Furthermore, this operation can be implemented by sequentially calculating the frequency overlap and parameter matching degree and comparing them with their respective thresholds. This ensures that the samples are highly coupled in both internal and external feature dimensions, improving the signal-to-noise ratio and causality of the training data.

[0132] Load the battery sample set to be analyzed, wherein the battery sample set to be analyzed has the same electrode material to be analyzed, electrolyte characteristics to be analyzed, list of key frequency points to be analyzed, related charge and discharge parameters to be analyzed, battery operating impedance time series data, and charge and discharge dynamic monitoring time series information;

[0133] The battery sample set to be analyzed can be a collection of multiple battery samples to be evaluated. Each sample includes its electrode materials, electrolyte characteristics, key frequency points, related parameters, and runtime sequence data. This data can be used as a screening object before training the life estimation basic model to extract valid samples that meet the two-level constraints. In an exemplary embodiment, the battery sample set to be analyzed can include, but is not limited to, one or more of the following: laboratory accelerated aging sample set, field-retired battery sample set, and multi-manufacturer battery sample set of the same model. Loading the battery sample set to be analyzed can be done by reading a multi-dimensional battery sample set containing materials, electrolyte, frequency points, parameters, and runtime sequence data from a database or experimental platform. Furthermore, this operation can be implemented through system calls to local storage or remote database interfaces, thereby providing the original data source for the two-level constraint screening.

[0134] When the compositional similarity between the electrode material to be analyzed and the battery electrode material is greater than or equal to the compositional similarity threshold, and the conductivity deviation between the electrolyte properties to be analyzed and the electrolyte property parameters is less than the deviation threshold, it is considered to satisfy the first-level constraint.

[0135] Determining whether the compositional similarity between the electrode material to be analyzed and the battery electrode material parameters is greater than or equal to a compositional similarity threshold can be achieved by calculating the similarity between the two in terms of main elemental composition, additive ratios, etc. (such as cosine similarity, Jaccard index), and comparing it with the threshold. Furthermore, this determination can be achieved by using cosine similarity calculation of material composition vectors or by using a compositional matching rule engine based on expert rules, thus enabling preliminary screening at the material level. Determining whether the conductivity deviation between the electrolyte properties to be analyzed and the electrolyte property parameters is less than a deviation threshold can be achieved by calculating the absolute difference or relative error of the two conductivity values ​​and determining whether it is within the allowable range. Furthermore, this determination can be achieved by using absolute deviation (i.e., the absolute value of the difference between the two conductivity values ​​is less than a set threshold) or relative deviation (i.e., the proportion of the difference to the benchmark value is less than a set threshold), thus completing the first-level constraint electrolyte consistency verification.

[0136] When the first-level constraint is met, and the frequency overlap between the list of key impedance frequency points and the list of key frequency points to be analyzed is greater than or equal to the overlap threshold, and the parameter matching degree between the list of high-frequency associated charge and discharge parameters and the associated charge and discharge parameters to be analyzed is greater than or equal to the matching threshold, it is considered that the second-level constraint is met.

[0137] When the first-level constraint is met, and the frequency overlap between the list of key impedance frequency points and the list of key frequency points to be analyzed is greater than or equal to the overlap threshold, and the parameter matching degree between the list of high-frequency associated charge-discharge parameters and the list of associated charge-discharge parameters to be analyzed is greater than or equal to the matching threshold, the second-level constraint is considered met. This can be achieved by calculating the frequency overlap and parameter matching degree sequentially after passing the first-level constraint, and comparing them with their respective thresholds. If all are satisfied, the second-level constraint is considered passed. Furthermore, this determination process can be executed by automated scripts or rule engines, thereby enabling fine-grained screening at the feature level and ensuring that samples are highly aligned in the aging characterization dimension.

[0138] When the secondary constraint is met, the battery operating impedance timing data and charge / discharge dynamic monitoring timing information are set as the first battery operating condition timing data. The cycle life termination time of the battery sample set to be analyzed is statistically analyzed to obtain the first label identifying the cycle life.

[0139] The battery operating impedance time-series data can be a sequence of electrochemical impedance spectroscopy data collected in chronological order during multiple cycles. This data can be used as a component of the first battery operating condition time-series data, providing dynamic evolution information on the internal aging state. For example, the battery operating impedance time-series data can include, but is not limited to, one or more of Nyquist plot sequences, Bode magnitude sequences, and phase angle time-series curves. The charge-discharge dynamic monitoring time-series information can be a time series of external operating parameters such as voltage, current, and temperature recorded synchronously during cyclic testing. This data can be used as another component of the first battery operating condition time-series data, providing external behavioral characteristics. In a specific embodiment, the charge-discharge dynamic monitoring time-series information can include, but is not limited to, one or more of constant current discharge voltage decay curves, pulse charging voltage drop response sequences, and temperature rise rate time sequences. Setting the battery operating impedance time-series data and the charge-discharge dynamic monitoring time-series information as the first battery operating condition time-series data can involve merging the impedance and charge-discharge time-series data of samples that pass two levels of constraints into a unified format multi-channel input sequence. Furthermore, this operation can be achieved through timestamp alignment and channel concatenation, thereby constructing a high-quality, highly consistent training input.

[0140] The cycle life end time of the battery sample set to be analyzed is statistically analyzed to obtain a first label identifying the cycle life. This label can be generated based on the actual number of failure cycles or the capacity decay endpoint for each sample. Furthermore, this operation can be achieved by detecting whether the capacity has decayed to 80% of the initial capacity, thus providing accurate supervisory signals to support the effective training of the life estimation model. The cycle life end time can be the number of cycles or the running time when the battery reaches the preset failure standard (e.g., capacity decay to 80%) during cycle testing. This can be used as the basis for generating the first label identifying the cycle life, providing a true target value for supervised learning.

[0141] Taking the construction of a universal model for multi-manufacturer lead-acid batteries as an example, the lead-acid battery life estimation method based on a multi-parameter data model in this embodiment can be as follows: An energy storage system integrator needs to build a unified life prediction model for 12V100Ah valve-regulated lead-acid batteries from different suppliers. First, define the first-level constraints: similarity of positive electrode PbO2 content ≥90%, electrolyte conductivity deviation <5%; second-level constraints: overlap of key frequency points (such as 0.05Hz, 1Hz) ≥80%, matching degree of high-frequency correlation parameters (such as voltage slope at the end of discharge, internal resistance change rate) ≥85%. Load 500 sets of battery samples from 3 manufacturers from the historical database, and retain 320 sets of valid samples after screening by the two-level constraints. Integrate the EIS time series and charge / discharge monitoring data of these samples into the first battery operating condition time series data, and use its actual failure cycle number (such as the capacity dropping to 80% after the 480th cycle) as the first identifier cycle life label for training the life estimation basic model. The model can be quickly adapted to any newly arrived battery model and can be deployed with only minor adjustments based on a small amount of actual test data.

[0142] In one embodiment, using the target lead-acid battery model, a list of key impedance frequency points, and a list of high-frequency associated charge and discharge parameters, several corresponding second battery operating condition time-series data and second label indicating cycle life are collected. Transfer learning is then performed on the life estimation basic model to obtain the target lead-acid battery life estimation model. The life estimation basic model is a weighted integration model of the outputs of multiple sub-models with different feature dimensions, including:

[0143] Obtain multiple life estimation sub-models from the life estimation base model;

[0144] An attention mechanism fusion layer is constructed to map the output features of multiple lifetime estimation sub-models to the input of the fusion layer;

[0145] Using the second battery operating condition time series data as input to multiple life estimation basic sub-models, and the label of the second cycle life as the output of the attention mechanism fusion layer, the life estimation basic model is transferred to the life estimation basic model. By dynamically adjusting the sub-model weights to optimize the fusion strategy, the life estimation model of the target lead-acid battery is obtained.

[0146] The lifespan estimation foundational sub-model can be multiple heterogeneous sub-models constituting the lifespan estimation foundational model. Each sub-model focuses on different feature dimensions (such as EIS features, charge / discharge parameters, or combinations thereof) for lifespan prediction. These sub-models can be used to capture battery aging information from multiple perspectives, providing complementary prediction viewpoints and enhancing the overall model's expressive power. In this embodiment, the lifespan estimation foundational sub-model can include, but is not limited to, one or more of the following: EIS-dominated sub-models, dynamic parameter-dominated sub-models, and hybrid feature fusion sub-models. Obtaining multiple lifespan estimation foundational sub-models can be achieved by training multiple independent sub-models using different feature subsets (such as only key impedance frequency points, only high-frequency associated charge / discharge parameters, or a combination of both) during the lifespan estimation foundational model training phase. Furthermore, this operation can be achieved by configuring the input feature space of each sub-model separately, thereby constructing a multi-perspective aging characterization capability and improving the information coverage breadth of the foundational model.

[0147] The output features can be intermediate prediction representations or feature vectors generated by each lifetime estimation sub-model after processing the input data. These can be used as input to the attention mechanism fusion layer, carrying the judgment basis of each sub-model on the lifetime status under the current operating condition. The input to the fusion layer can be a joint input tensor formed by concatenating or stacking the output features of multiple lifetime estimation sub-models. This can be used to provide a unified representation space for multi-source prediction information for the attention mechanism fusion layer. Constructing the attention mechanism fusion layer can involve designing a learnable neural network module that receives the output features of multiple sub-models and outputs a weighted fusion prediction result. Furthermore, the attention mechanism fusion layer can be constructed by using a fully connected layer + softmax to generate attention weights, or by using a query-key-value (QKV) structure to implement a multi-head attention mechanism, thereby introducing a dynamic weight allocation mechanism and making the fusion process adaptive to the input. In an exemplary embodiment, the attention mechanism fusion layer can include, but is not limited to, additive attention fusion layers, multiplicative attention fusion layers, and multi-head attention fusion layers.

[0148] Mapping the output features of multiple lifetime estimation sub-models to the input of the fusion layer can be achieved by concatenating the output vectors of each sub-model along the channel or time dimension to form a joint input tensor for the fusion layer. Furthermore, this operation can be implemented through tensor concatenation or stacking operations, thus providing structured multi-source information input for the attention mechanism. Using the second battery operating condition time-series data as input to multiple lifetime estimation sub-models can be achieved by simultaneously inputting the second battery operating condition time-series data of the target model into all lifetime estimation sub-models. Furthermore, this operation can be implemented through a parallel forward propagation mechanism, thereby activating the parallel evaluation of the target battery's current state by each sub-model.

[0149] The attention mechanism fusion layer can be a learnable weighted fusion module built on the attention mechanism. It dynamically allocates the contribution weights of each sub-model's output, enabling adaptive weighted fusion of prediction results from multiple sub-models, strengthening highly discriminative feature channels, and suppressing noise interference. In this embodiment, the attention mechanism fusion layer can calculate the relevance score between each sub-model's output and the current input context using trainable parameters, and generate a weight vector after softmax normalization. For example, the attention mechanism fusion layer works collaboratively with the lifetime estimation base sub-model to form an end-to-end trainable ensemble prediction architecture. Using the label of the second identifier cycle lifetime as the output of the attention mechanism fusion layer, the final output of the fusion layer can be compared with the label of the second identifier cycle lifetime during transfer learning to calculate the loss. Furthermore, this operation can be implemented using supervised learning loss functions such as mean squared error or cross-entropy, thereby guiding the attention mechanism and sub-models to optimize together, making the fusion result approximate the true lifetime label.

[0150] Sub-model weights can be dynamically adjustable weight coefficients assigned by the attention mechanism fusion layer to each lifespan estimation sub-model. These weights reflect the relative importance of each sub-model under the current input conditions, supporting adaptive fusion decisions. The fusion strategy can be a rule or mechanism that integrates the outputs of multiple sub-models into a final prediction result. It determines how multi-source information works synergistically to generate robust predictions, and in this scheme, it is dynamically optimized by the attention mechanism. Optimizing the fusion strategy by dynamically adjusting sub-model weights can be achieved during transfer learning training by updating the attention mechanism parameters through backpropagation, allowing the sub-model weights to adapt adaptively to the input conditions. Furthermore, this operation can be implemented through end-to-end joint fine-tuning of the sub-models and attention layer, or by freezing the sub-model parameters and fine-tuning only the attention mechanism fusion layer. This enables adaptive responses to the aging characteristics of the target model, improving prediction accuracy and robustness under small sample sizes.

[0151] Taking the health management of lead-acid batteries in a renewable energy storage system as an example, the lead-acid battery life estimation method based on a multi-parameter data model in this embodiment can be as follows: A photovoltaic energy storage station uses deep-cycle lead-acid batteries, and its life estimation basic model includes three sub-models: Sub-model A focuses on the impedance phase angle of 0.01–1Hz, Sub-model B analyzes the voltage recovery rate at the end of charging, and Sub-model C integrates both. After deploying a new batch of batteries of the same model, the system collects the operating time-series data of the second battery in the first 20 cycles and the corresponding capacity decay label. During the transfer learning stage, the three sub-models process the input data in parallel, and the attention mechanism fusion layer dynamically assigns higher weights to sub-model B according to the current cycle's charge and discharge mode (such as shallow charge and shallow discharge vs. deep cycle) (because the EIS change is not significant under shallow charge conditions). After fine-tuning, the target lead-acid battery life estimation model can more accurately predict the remaining life of this specific batch of batteries under actual light fluctuations.

[0152] In addition, refer to Figure 2 To achieve the above objectives, the present invention also provides a lead-acid battery life estimation system based on a multi-parameter data model, the system comprising:

[0153] The feature extraction module 10 is used to load the operation monitoring data of the target lead-acid battery, perform feature frequency range extraction, and obtain a list of key impedance frequency points.

[0154] The correlation analysis module 20 is used to traverse the list of key impedance frequency points based on the operation monitoring data, perform charge and discharge characteristic correlation analysis, and obtain a list of high-frequency correlated charge and discharge parameters.

[0155] The model training module 30 is used to collect several one-to-one corresponding first battery operating condition time series data and first identifier cycle life labels, with battery electrode material parameters, electrolyte characteristic parameters, the list of key impedance frequency points and the list of high-frequency related charge and discharge parameters as constraints, and to train the life estimation basic model.

[0156] The transfer modeling module 40 is used to collect several corresponding second battery operating condition time series data and second identifier cycle life labels based on the target lead-acid battery model, the list of key impedance frequency points and the list of high-frequency associated charge and discharge parameters, and to perform transfer learning on the life estimation basic model to obtain the target lead-acid battery life estimation model and execute the lead-acid battery life estimation task.

[0157] Other embodiments or specific implementations of the lead-acid battery life estimation system based on a multi-parameter data model described in this invention can be referred to the above-described method embodiments, and will not be repeated here.

[0158] Furthermore, to achieve the above objectives, the present invention also provides a lead-acid battery life estimation device based on a multi-parameter data model. The device includes: a memory, a processor, and a lead-acid battery life estimation program based on a multi-parameter data model stored in the memory and executable on the processor. The lead-acid battery life estimation program based on a multi-parameter data model is configured to implement the steps of the lead-acid battery life estimation method based on a multi-parameter data model as described above.

[0159] In addition, to achieve the above objectives, the present invention also provides a medium storing a lead-acid battery life estimation program based on a multi-parameter data model, wherein when the lead-acid battery life estimation program based on the multi-parameter data model is executed by a processor, the program implements the steps of the lead-acid battery life estimation method based on the multi-parameter data model as described above.

[0160] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for estimating the lifespan of lead-acid batteries based on a multi-parameter data model, characterized in that, The method includes: Load the operational monitoring data of the target lead-acid battery, perform characteristic frequency range extraction, and obtain a list of key impedance frequency points; Based on the operational monitoring data, the list of key impedance frequency points is traversed, and a charge-discharge characteristic correlation analysis is performed to obtain a list of high-frequency correlated charge-discharge parameters. Using battery electrode material parameters, electrolyte characteristic parameters, the list of key impedance frequency points, and the list of high-frequency associated charge and discharge parameters as constraints, collect several one-to-one corresponding first battery operating condition time series data and first identifier cycle life labels to train the life estimation basic model. Using the target lead-acid battery model, the list of key impedance frequency points, and the list of high-frequency associated charge and discharge parameters, collect several corresponding second battery operating condition time-series data and second identifiers of cycle life. Perform transfer learning on the life estimation basic model to obtain the target lead-acid battery life estimation model and execute the lead-acid battery life estimation task.

2. The lead-acid battery life estimation method based on a multi-parameter data model as described in claim 1, characterized in that, The operational monitoring data of the target lead-acid battery is used to extract characteristic frequency ranges and obtain a list of key impedance frequency points, including: Constrained by battery electrode material parameters and electrolyte characteristic parameters, impedance spectrum test data are collected, impedance modulus-frequency curve characteristic point analysis is performed, and characteristic impedance frequency interval set is obtained. From the operational monitoring data, extract the battery electrochemical impedance spectroscopy data, and statistically analyze the frequency points in each interval of the characteristic impedance frequency interval set whose impedance modulus change rate is greater than the change rate threshold and the impedance characteristic trigger frequency set. The frequency points whose trigger frequencies are greater than or equal to the frequency threshold are added to the list of key impedance frequency points.

3. The lead-acid battery life estimation method based on a multi-parameter data model as described in claim 2, characterized in that, The process involves extracting battery electrochemical impedance spectroscopy data from the operational monitoring data, and statistically analyzing the set of frequency points within each characteristic impedance frequency range where the rate of change of impedance modulus is greater than a threshold, along with the set of impedance characteristic trigger frequencies. This includes: From the battery electrochemical impedance spectroscopy data, extract the first set of impedance spectroscopy data up to the Mth set of impedance spectroscopy data, where M represents the number of impedance spectroscopy data sets. The first group of impedance spectrum data is divided into frequency intervals up to the Mth group of impedance spectrum data to obtain the first frequency interval impedance dataset up to the Kth frequency interval impedance dataset, where K represents the number of frequency intervals. Delete the interval datasets with impedance data sample size less than the sample size threshold to obtain several retained frequency interval impedance datasets. Traverse the several retained frequency interval impedance datasets, and based on the characteristic impedance frequency interval set, count the frequency points in each interval where the rate of change of impedance magnitude is greater than the rate of change threshold and the impedance characteristic trigger frequency set.

4. The lead-acid battery life estimation method based on a multi-parameter data model as described in claim 3, characterized in that, The process of traversing the several retained frequency interval impedance datasets, based on the characteristic impedance frequency interval set, statistically analyzing the frequency point set and impedance characteristic trigger frequency set within each interval where the rate of change of impedance magnitude is greater than the rate of change threshold, includes: From the set of characteristic impedance frequency intervals, extract the first characteristic frequency interval up to the Pth characteristic frequency interval, where P represents the number of characteristic frequency intervals; Extract the impedance data of the first retention interval from the plurality of retention frequency interval impedance datasets; Calculate the rate of change of impedance modulus within the first characteristic frequency interval, filter out frequency points with a rate of change greater than the rate of change threshold, and count their trigger frequencies, which are set as the trigger frequencies of the first interval. Until the rate of change of impedance modulus within the Pth characteristic frequency interval is calculated, frequency points with a rate of change greater than the rate of change threshold are selected, and their trigger frequencies are statistically analyzed and set as the trigger frequencies of the Pth interval. The frequency points corresponding to the first interval trigger frequency up to the Pth interval trigger frequency are added to the frequency point set, and the trigger frequency value is added to the impedance characteristic trigger frequency set.

5. The lead-acid battery life estimation method based on a multi-parameter data model as described in claim 1, characterized in that, Based on the operational monitoring data, the process iterates through the list of key impedance frequency points, performs charge-discharge characteristic correlation analysis, and obtains a list of high-frequency correlated charge-discharge parameters, including: Extract the first key frequency point from the list of key impedance frequency points; Based on the first key frequency point, extract the first set of impedance-charge-discharge parameter correlation data from the operation monitoring data; Based on the Pearson correlation coefficient threshold, feature filtering is performed on the first set of associated data to obtain a set of charging and discharging parameters that are strongly correlated with the impedance modulus. The frequency of occurrence of the charging and discharging parameter set in multiple sets of associated data is statistically analyzed, and the charging and discharging parameters with a frequency greater than the frequency threshold are set as the first high-frequency associated charging and discharging parameters. The process continues until all key frequency points have undergone correlation analysis, high-frequency correlated charge and discharge parameters are summarized, and a list of high-frequency correlated charge and discharge parameters is output.

6. The lead-acid battery life estimation method based on a multi-parameter data model as described in claim 1, characterized in that, The process involves collecting a series of one-to-one corresponding first battery operating condition time-series data and first-identifying cycle life labels, constrained by battery electrode material parameters, electrolyte characteristic parameters, the key impedance frequency point list, and the high-frequency correlated charge-discharge parameter list, to train a basic lifespan estimation model. This includes: Based on the battery electrode material parameters and electrolyte characteristic parameters, a first-level constraint is constructed; Based on the list of key impedance frequency points and the list of high-frequency associated charge and discharge parameters, a secondary constraint is constructed. Load the battery sample set to be analyzed, wherein the battery sample set to be analyzed has the same electrode material to be analyzed, electrolyte characteristics to be analyzed, list of key frequency points to be analyzed, related charge and discharge parameters to be analyzed, battery operating impedance time series data, and charge and discharge dynamic monitoring time series information; When the compositional similarity between the electrode material to be analyzed and the battery electrode material is greater than or equal to the compositional similarity threshold, and the conductivity deviation between the electrolyte characteristics to be analyzed and the electrolyte characteristic parameters is less than the deviation threshold, it is considered to satisfy the first-level constraint. When the first-level constraint is met, and the frequency overlap between the list of key impedance frequency points and the list of key frequency points to be analyzed is greater than or equal to the overlap threshold, and the parameter matching degree between the list of high-frequency associated charge and discharge parameters and the associated charge and discharge parameters to be analyzed is greater than or equal to the matching threshold, it is considered that the second-level constraint is met. When the secondary constraint is met, the battery operating impedance timing data and the charge / discharge dynamic monitoring timing information are set as the first battery operating condition timing data. The cycle life termination time is statistically analyzed for the battery sample set to be analyzed to obtain the first label identifying the cycle life.

7. The lead-acid battery life estimation method based on a multi-parameter data model as described in claim 1, characterized in that, The method involves collecting a number of corresponding second battery operating condition time-series data and second cycle life labels using the target lead-acid battery model, the key impedance frequency point list, and the high-frequency associated charge and discharge parameter list. Transfer learning is then applied to the life estimation basic model to obtain the target lead-acid battery life estimation model. This life estimation basic model is a weighted ensemble of the outputs of multiple sub-models with different feature dimensions, including: Obtain multiple life estimation sub-models of the aforementioned life estimation basic model; An attention mechanism fusion layer is constructed to map the output features of multiple lifetime estimation sub-models to the input of the fusion layer; Using the second battery operating condition time series data as input to multiple life estimation basic sub-models, and using the second label identifying cycle life as output of the attention mechanism fusion layer, the life estimation basic model is transferred to the life estimation basic model. By dynamically adjusting the sub-model weights to optimize the fusion strategy, the target lead-acid battery life estimation model is obtained.

8. A lead-acid battery life estimation system based on a multi-parameter data model, characterized in that, The system includes: The feature extraction module is used to load the operation monitoring data of the target lead-acid battery, perform feature frequency range extraction, and obtain a list of key impedance frequency points. The correlation analysis module is used to traverse the list of key impedance frequency points based on the operation monitoring data, perform charge and discharge characteristic correlation analysis, and obtain a list of high-frequency correlated charge and discharge parameters. The model training module is used to collect several one-to-one corresponding first battery operating condition time series data and first identifier cycle life labels, with battery electrode material parameters, electrolyte characteristic parameters, the list of key impedance frequency points and the list of high-frequency related charge and discharge parameters as constraints, and to train the life estimation basic model. The transfer modeling module is used to collect several corresponding second battery operating condition time-series data and second identifier cycle life labels based on the target lead-acid battery model, the list of key impedance frequency points, and the list of high-frequency associated charge and discharge parameters. It then performs transfer learning on the life estimation basic model to obtain the target lead-acid battery life estimation model and executes the lead-acid battery life estimation task.

9. A lead-acid battery life estimation device based on a multi-parameter data model, characterized in that, The device includes: a memory, a processor, and a lead-acid battery life estimation program based on a multi-parameter data model stored in the memory and executable on the processor, the lead-acid battery life estimation program based on the multi-parameter data model being configured to implement the steps of the lead-acid battery life estimation method based on a multi-parameter data model as described in any one of claims 1 to 7.

10. A medium, characterized in that, The medium stores a lead-acid battery life estimation program based on a multi-parameter data model. When the lead-acid battery life estimation program based on the multi-parameter data model is executed by the processor, it implements the steps of the lead-acid battery life estimation method based on a multi-parameter data model as described in any one of claims 1 to 7.

Images