Battery state of health estimation method and device fusing pattern recognition and inflection point features

Abstract

This invention provides a battery health state estimation method and apparatus that integrates pattern recognition and inflection point features, relating to the field of battery health management technology. The invention preprocesses collected time-series data and extracts multi-dimensional features; utilizes an improved heuristic clustering algorithm to identify degradation patterns; extracts inflection point cycle number and sharpness factor as augmented features through a two-stage nonlinear hybrid model; employs an improved snake optimization algorithm to collaboratively optimize the hyperparameters and feature subsets of a Gaussian process regression model; and introduces an adaptive position update strategy driven by inflection point features and a composite fitness function for optimization, resulting in an optimized model used for battery health state estimation. This application can accurately capture the nonlinear phase transition behavior in battery degradation, effectively isolate heterogeneous aging interference, significantly improve estimation accuracy and generalization performance throughout the entire lifespan, and achieve intelligent model construction throughout the entire process.

Description

Technical Field

[0001] This invention relates to the field of battery health management and state monitoring technology, and more specifically, to a method and apparatus for estimating battery health state by integrating pattern recognition and inflection point features. Background Technology

[0002] In battery management systems, accurately estimating the battery's health status is crucial for ensuring system safety, optimizing energy allocation, and predicting the battery's remaining lifespan.

[0003] Currently, data-driven estimation methods based on Gaussian process regression are one of the mainstream approaches for predicting battery health status. These methods typically extract statistical health features such as voltage, current, and time from historical charge-discharge time-series data of the battery. They then combine the feature data from different individual batteries to construct a unified training set. Subsequently, algorithms such as grid search or Bayesian optimization are used to fine-tune the hyperparameters of the regression model, ultimately using a single, fixed model to make unified predictions for all tested batteries.

[0004] However, existing estimation techniques still face significant challenges in handling the complexity and heterogeneity of battery degradation. First, because different individual batteries exhibit significant differences in aging patterns—for example, the degradation trajectories of long-life, medium-life, and short-life batteries differ—existing techniques often ignore this inter-individual heterogeneity. Training with all data forces the model to learn a compromised degradation pattern, leading to decreased estimation accuracy for batteries with specific aging patterns. Second, existing health feature extraction focuses primarily on static statistics such as mean or variance, lacking the ability to characterize the nonlinear inflection point behavior that marks accelerated degradation in the battery capacity decay curve. This makes it difficult for models to accurately depict and predict the rapid changes in health status before and after the inflection point, resulting in larger prediction errors in the later stages of battery life. Furthermore, feature selection and model hyperparameter optimization are typically two separate steps, and traditional optimization methods are inefficient in high-dimensional search spaces, failing to quickly find the globally optimal combination of feature subsets and model hyperparameters in complex data distributions. This limits the upper limit of model performance to human experience.

[0005] In view of the above, this application is hereby submitted. Summary of the Invention

[0006] The present invention aims to provide a battery health state estimation method, device, equipment and medium that integrates pattern recognition and inflection point features, so as to solve the defects in the existing lithium-ion battery health state estimation, such as the model accuracy being affected by heterogeneous data, insufficient feature representation of nonlinear degradation dynamics, and low and incomplete optimization efficiency.

[0007] To solve the above-mentioned technical problems, the present invention is achieved through the following technical solution:

[0008] A battery health state estimation method that integrates pattern recognition and inflection point features includes:

[0009] S1, acquire and preprocess the timing data of the lithium-ion battery charge-discharge cycle;

[0010] S2, extract initial battery health features from the charging segments of the preprocessed time-series data;

[0011] S3, using a heuristic clustering algorithm that combines genetic algorithm and simulated annealing algorithm, the initial battery health characteristics are divided into different degradation mode clusters;

[0012] S4, construct a two-stage linear-nonlinear hybrid model to fit and extract key inflection point features from the preprocessed time series data; wherein, the key inflection point features are the key feature points of the linear-nonlinear stage transition in the time series data where a trend change occurs and the slope changes significantly.

[0013] S5, based on the degradation mode cluster, evaluate the importance of the initial battery health features and the key inflection point features for each degradation mode, and select the optimal feature subset with high importance for each degradation mode;

[0014] S6. Based on the optimal feature subset, the snake optimization algorithm combined with Gaussian process regression is used to optimize the model hyperparameters.

[0015] S7 inputs the time series data to be tested into the hyperparameter-optimized model and outputs the battery health state estimation result.

[0016] Preferably, the preprocessing includes aligning and filtering the time-series data;

[0017] The initial battery health characteristics include: time-related characteristics, charge-related characteristics, voltage curve characteristics, and current curve characteristics.

[0018] Preferably, S3 specifically comprises:

[0019] Set the target number of clusters to K;

[0020] Based on the initial battery health characteristics, a genetic algorithm is used to encode each individual's chromosome into a set of vectors containing K cluster centers, serving as the initial population; wherein the dimension of the chromosome is determined by the dimension of the initial battery health characteristics; the expression for the initial population is:

[0021] ;

[0022] in, This is the initial population, i.e., the set of vectors representing the cluster centers; For the first Cluster center vectors; It is a real number; Dimensions of initial battery health characteristics;

[0023] The optimization is performed with the objective function being the minimization of the sum of weighted intra-class numerical attribute distances and categorical attribute dissimilarity. The formula is:

[0024] ;

[0025] in, This is the membership matrix; To belong to cluster The An initial battery health feature vector; For cluster centers The defined first A cluster of degenerate patterns; , These are the weight coefficients for numerical features and categorical features, respectively. The numerical feature is the Euclidean distance; It is used as a measure of dissimilarity in classification features; , The first Numerical health feature component and categorical health feature component of individual battery sample feature vectors; , The first Numerical health feature component and categorical health feature component of each cluster center vector;

[0026] The fitness value is calculated using a clustering quality quantification function, expressed as:

[0027] ;

[0028] in, This is the fitness value; For cluster centers The defined first A cluster of degenerate patterns; To belong to cluster The An initial battery health feature vector; For sample weights; It is an L2 norm;

[0029] The initial population is iteratively updated based on the fitness value to seek the optimal clustering scheme and obtain the globally optimal individual.

[0030] Decoding the globally optimal individual yields the final cluster center of the initial battery health characteristics, i.e., the degradation pattern cluster.

[0031] Preferably, iteratively updating the initial population includes:

[0032] Based on the fitness value, the initial population is screened for dominant solutions to obtain two parent prototype sets;

[0033] The prototype vectors of two parent prototype sets are recombined through convex combination to generate a new candidate prototype set, expressed as:

[0034] ;

[0035] in, For the new candidate prototype set; These are the weighting coefficients; , For two parent prototype sets;

[0036] Based on the new candidate prototype set, random perturbation is added to the cluster center vector to expand the exploration range of the solution space, find the globally optimal clustering scheme, and output the corresponding new fitness value;

[0037] Based on the new fitness value, a simulated annealing mechanism is introduced for local optimization, and temporary deterioration is accepted with dynamic acceptance probability until the algorithm meets the convergence condition or reaches the preset maximum number of iterations, at which point the iteration terminates and the globally optimal individual is obtained.

[0038] Preferably, the dynamic acceptance probability is generated based on the difference between the current annealing temperature and the intra-class dispersion to avoid the algorithm getting trapped in local optima. The expression is:

[0039] ;

[0040] in, For dynamic acceptance probability; The difference between the new and old fitness values; This is the annealing temperature for the current iteration.

[0041] An exponential cooling strategy is employed to reduce the annealing temperature until convergence, expressed as:

[0042] ;

[0043] in, This is the annealing temperature for the next round; This is the temperature cooling coefficient.

[0044] Preferably, the two-stage linear-nonlinear hybrid model uses a hyperbolic tangent cumulative distribution function as the transition function to describe the relationship between capacity and cycle number, expressed as:

[0045] ;

[0046] in, For the first Battery capacity during the next cycle; This refers to the number of battery cycles. These are the initial capacity parameters; This represents the linear decay slope in the first stage, i.e., the linear decay slope before the inflection point. This represents the linear decay slope in the second stage, i.e., the linear decay slope after the inflection point. The number of cycles at the inflection point; It is a hyperbolic tangent cumulative distribution function, used to achieve a smooth transition between two stages; For the transition rate coefficient; This is the random error term;

[0047] When extracting key inflection point features from the preprocessed time-series data, a nonlinear least squares method is used to optimize the fitting of the two-stage linear-nonlinear hybrid model to minimize the sum of squared residuals between the model's predicted values ​​and the actual observed values, thus obtaining the optimal parameter set. The formula is as follows:

[0048] ;

[0049] in, The set of parameters for the model to be fitted, including the initial capacity parameters. The linear decay slope in the first stage The linear decay slope in the second stage Inflection point cycle number With transition rate coefficient ; For minimization operation; This represents the total number of battery cycles. For the first The actual observation capacity of the next cycle; Predict capacity for the model;

[0050] Based on the optimal parameter set, key inflection point features, including the inflection point cycle number and the inflection point sharpness factor, are extracted; wherein, the inflection point sharpness factor is the first derivative of the transition function at zero, used to quantify the drastic change in the degradation rate at the inflection point, and its expression is:

[0051] ;

[0052] in, The sharpness factor at the inflection point; Indicates the number of loops Find the first derivative; The optimal linear decay slope for the second stage; The optimal inflection point cycle number; The optimal transition rate coefficient; For the transition function in The first derivative at that point.

[0053] Preferably, before assessing the importance of the features, the preprocessed time series data is standardized based on the degradation pattern cluster;

[0054] Then, based on the random forest tree model, the importance score of the standardized data features is calculated using the following formula:

[0055] ;

[0056] in, For the first 3D features Importance score; This represents the total number of decision trees in the random forest. This is the index of the b-th decision tree in the random forest; Let be the out-of-bag mean squared error of the b-th decision tree; For the b-th decision tree, the first... The new out-of-bag mean square error is obtained by randomly arranging the eigenvalues ​​of the dimensional feature.

[0057] Based on the calculated importance score, compare it with a set threshold, and select the feature data whose importance score is greater than the set threshold, which is the optimal feature subset.

[0058] Preferably, a snake optimization algorithm combining Gaussian process regression is used to optimize the model hyperparameters, specifically as follows:

[0059] Let the current iteration number be... The maximum number of iterations is The globally optimal individual position is ;

[0060] Inflection point cycle number Inflection point sharpness factor It is dual-driven, based on the current iteration number. and Based on the relative relationships, the optimization process is divided into the pre-inflection point stage, the inflection point region stage, and the post-inflection point stage:

[0061] when At this point, in the pre-inflection point stage, a global exploration is performed based on the center of the degradation mode cluster to adapt to the slow degradation stage before the battery inflection point. The formula is:

[0062] ;

[0063] in, This is the inflection point ratio control coefficient; , The first The individual snake was in the first , The position vector of the next iteration; For pattern exploration coefficients; It is the center of the degenerate pattern cluster; Historical experience weighting coefficient; A random number in the range of 0-1; This represents the average evolutionary direction vector of historically superior solutions under the degradation mode;

[0064] when At this point, during the inflection point region stage, refined local development is carried out to adapt to the abrupt change in degradation rate at the battery inflection point. The formula is:

[0065] ;

[0066] in, The radius of the neighborhood window at the inflection point; The mode convergence factor; It is a very small constant; The sharpness factor at the inflection point; for Sensitivity scaling factor; This is a sign function used to indicate the direction of position updates;

[0067] when At this point, in the post-inflection point stage, steady-state local development is carried out with the global optimal individual as the guide, in order to adapt to the accelerated degradation stage after the battery inflection point. The formula is:

[0068] ;

[0069] in, A random number in the range of 0-1;

[0070] A composite fitness function is constructed that integrates global trend consistency, inflection point region accuracy, and feature set quality. Through dynamic weighting, it adaptively focuses on the prediction accuracy of inflection point regions, outputting the globally optimal individual for each degradation mode. The formula for the composite fitness function is:

[0071] ;

[0072] in, This is the normalized correlation coefficient between the predicted and actual values, i.e., the global trend consistency term; This is the normalized average absolute error within the inflection point window, i.e., the accuracy term for the inflection point region. The similarity between the selected feature subset and the ideal feature subset in the pattern, i.e., the feature set quality term; , These are the dynamic weights of the global trend consistency term and the inflection point region accuracy term, respectively. The weights of the quality terms in the feature set;

[0073] The formula for the dynamic weight is:

[0074] ;

[0075] ;

[0076] in, The number of cycles at the inflection point Proximity response factor of location; This is the width coefficient for the Gaussian weights;

[0077] The globally optimal individual includes the optimal hyperparameters, optimal feature subset mask, optimal mode adjustment parameters, and optimal inflection point response parameters of the Gaussian process regression model; wherein, the optimal mode adjustment parameters include the optimal mode exploration coefficients. With mode convergence factor The optimal inflection point response parameters include the optimal sensitivity scaling factor. With proximity response factor .

[0078] Preferably, the standardization process is as follows:

[0079] The 3-Sigma criterion is used to identify and remove outlier samples;

[0080] For the feature data after removing outliers, Z-score standardization is performed to eliminate differences in feature dimensions;

[0081] Let the feature matrix of the training set after removing anomalies be... The standardized formula is:

[0082] ;

[0083] in, The standardized feature matrix; It is the feature mean vector; This is the standard deviation vector.

[0084] The present invention also provides a battery health state estimation device that integrates pattern recognition and inflection point features, comprising:

[0085] The data acquisition and preprocessing module is used to acquire and preprocess the time-series data of lithium-ion battery charge-discharge cycles.

[0086] The multi-dimensional feature extraction module is used to extract initial battery health features from the charging segments of the preprocessed time-series data.

[0087] The pattern recognition module is used to divide the initial battery health features into different degradation pattern clusters using a heuristic clustering algorithm that combines genetic algorithm and simulated annealing algorithm.

[0088] The inflection point dynamics extraction module is used to construct a two-stage linear-nonlinear hybrid model to fit and extract key inflection point features from preprocessed time series data; wherein, the key inflection point features are key feature points in the time series data where abrupt trend changes and significant slope changes occur during the linear-nonlinear stage transition;

[0089] The optimal feature selection module is used to evaluate the importance of the initial battery health features and the key inflection point features based on the degradation mode cluster, and select the optimal feature subset with high importance for each degradation mode;

[0090] The collaborative optimization modeling module is used to optimize the model hyperparameters based on the optimal feature subset using a snake optimization algorithm that combines Gaussian process regression.

[0091] The state estimation module is used to input the time series data to be tested into the hyperparameter-optimized model and output the battery health state estimation results.

[0092] The present invention also provides a battery health state estimation device that integrates pattern recognition and inflection point features, including a processor and a memory. The memory stores a computer program that can be executed by the processor to implement the battery health state estimation method that integrates pattern recognition and inflection point features as described above.

[0093] The present invention also provides a computer-readable storage medium storing computer-readable instructions, which, when executed by a processor of the device on which the computer-readable storage medium resides, implement the battery health state estimation method described above, which integrates pattern recognition and inflection point features.

[0094] In summary, compared with the prior art, the present invention has the following beneficial effects:

[0095] First, this invention significantly improves the nonlinear dynamic characterization capability and estimation accuracy. By introducing parameterized inflection point cycle number and inflection point sharpness factor, and combining them with a dynamic weight fitness function, it solves the technical problem that traditional statistical features are insensitive to capacity drop phenomena in the mid-to-late stages of battery life. The introduction of inflection point features enables the model to accurately capture the key time points where the degradation rate undergoes a phase transition. Combined with adaptive focusing on the accuracy of the inflection point region, it greatly reduces the prediction error of the battery in the mid-to-late stages of its life. Experiments show that its mean absolute error and root mean square error are both superior to existing technologies.

[0096] Secondly, this invention enhances the model's generalization ability and scenario adaptability. It employs a strategy of first clustering and identification, followed by customized modeling, effectively isolating heterogeneous aging interference between long-life, medium-life, and short-life batteries due to differences in material systems or operating conditions. The system automatically identifies the degradation mode of the battery under test and calls upon its specific optimal model, ensuring the estimation scheme is highly robust to individual batteries with different degradation rates, thus avoiding the trade-off errors of a single general-purpose model when handling heterogeneous data.

[0097] Third, this invention achieves full automation and intelligence from feature engineering to model construction. It innovatively integrates feature selection masks, model hyperparameters, mode adjustment parameters, and inflection point response parameters into a unified optimization framework, and performs efficient optimization in complex high-dimensional spaces through an improved snake optimization algorithm. Employing a three-stage adaptive position update strategy, it dynamically adjusts the search step size and direction based on the actual physical degradation stage of the battery, reducing reliance on manual feature engineering experience and improving the method's universality and engineering application efficiency across different battery systems. Attached Figure Description

[0098] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0099] Figure 1 This is a flowchart illustrating a battery health state estimation method that integrates pattern recognition and inflection point features, as provided in Example 1.

[0100] Figure 2 This is a battery health (SOH) feature diagram of a short-term battery provided in Example 1.

[0101] Figure 3 This is a battery health (SOH) characteristic diagram of a medium-sized battery provided in Example 1.

[0102] Figure 4 This is a battery health (SOH) characteristic diagram of a long-term battery provided in Example 1.

[0103] Figure 5 This is a visualization diagram for identifying the two-stage inflection point of lithium-ion battery capacity decay provided in Example 1.

[0104] Figure 6 The battery health characteristics distribution diagram of the three degradation mode clusters of short life (a), medium life (b) and long life (c) after standardization provided in Example 1.

[0105] Figure 7 The three-dimensional clustering distribution diagram of feature values ​​under three degradation modes—long lifetime, medium lifetime, and short lifetime—after feature selection is provided in Example 1.

[0106] Figure 8 The overall flowchart of the pattern adaptive collaborative optimization (ISO-GPR) provided in Example 1.

[0107] Figure 9 The graph shows the predicted and actual SOH values ​​of the short-life battery provided in Example 1.

[0108] Figure 10 This is a comparison curve of the predicted and actual SOH values ​​of the medium-life battery provided in Example 1.

[0109] Figure 11 This is a comparison curve of the predicted and actual SOH values ​​of the medium-life battery provided in Example 1.

[0110] Figure 12 This is a comparison chart of the results of the method of the present invention provided in Example 1 with other conventional methods on the MAE, MAPE, RMSE, and error value heatmap.

[0111] Figure 13 The graph shows a comparison of the results of the Gaussian regression prediction ISO-GPR algorithm of the present invention with other mainstream machine learning algorithms on the MAE, MAPE, RMSE, and error value heatmaps provided in Example 1.

[0112] Figure 14 This is a schematic diagram of a battery health state estimation device that integrates pattern recognition and inflection point features, as provided in Embodiment 2.

[0113] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. Detailed Implementation

[0114] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely represents selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0115] Example 1

[0116] Embodiment 1 of the present invention provides a battery health state estimation method that integrates pattern recognition and inflection point features. It can be implemented by a battery health state estimation device that integrates pattern recognition and inflection point features (hereinafter referred to as estimation device), and in particular, it can be executed by one or more processors within the estimation device.

[0117] In this embodiment, the estimation device may be an electronic device equipped with a processor, the processor having a computer program for the battery health state estimation method that integrates pattern recognition and inflection point features, and the computer program can be executed, such as a computer, smartphone, smart tablet, workstation, etc., which are not limited here.

[0118] This embodiment provides a method and system for estimating the health status of lithium-ion batteries based on degradation mode identification, nonlinear inflection point characterization, and multi-objective collaborative optimization. By systematically reconstructing the heterogeneity, nonlinearity, and optimality of model parameters in the battery degradation process, it achieves high-precision monitoring of the health status throughout the entire life cycle.

[0119] like Figure 1 As shown, a battery health state estimation method that integrates pattern recognition and inflection point features includes steps S1 to S7.

[0120] S1: Acquire and preprocess the timing data of the lithium-ion battery charge-discharge cycle.

[0121] The timing data of a lithium-ion battery during multiple charge-discharge cycles is acquired through a data acquisition device. The timing data includes voltage sequence, current sequence and time sequence.

[0122] The preprocessing includes aligning and filtering the time-series data. A preset filter can be used to perform noise reduction on the time-series data, and a linear interpolation algorithm can be used to time-align sampling points under different cycles, eliminating interference caused by sensor noise and inconsistencies in sampling frequency.

[0123] Alternatively, you can directly use publicly available datasets (such as the MIT-Stanford dataset).

[0124] S2 extracts initial battery health features from the charging segments of the preprocessed time-series data.

[0125] The initial battery health characteristics include: time-related characteristics, charge-related characteristics, voltage curve characteristics, and current curve characteristics.

[0126] From the preprocessed constant current charging and constant voltage charging segments, at least sixteen initial health features, encompassing time-related features, charge-related features, voltage curve features, and current curve features, are systematically extracted to construct an initial battery health feature vector set. The time-related features include constant current charging time and constant voltage charging time; the charge-related features include accumulated charge during the constant current phase and accumulated charge during the constant voltage phase; the voltage curve features include voltage mean, voltage standard deviation, voltage skewness, voltage kurtosis, voltage slope, and voltage curve information entropy; and the current curve features include current mean, current standard deviation, current skewness, current kurtosis, current slope, and current information entropy.

[0127] Table 1 shows the 16 initial health characteristics across four dimensions.

[0128] Table 1. Health characteristics under four dimensions

[0129]

[0130] S3. A heuristic clustering algorithm that combines genetic algorithm and simulated annealing algorithm is used to divide the initial battery health characteristics into different degradation mode clusters.

[0131] This step utilizes an improved heuristic clustering algorithm (IGSA-KP) that integrates the global exploration capabilities of a genetic algorithm with the probabilistic jump characteristics of simulated annealing to unsupervisedly divide the initial battery health features of all battery samples into multiple disjoint degradation pattern clusters. In the initialization phase, this invention uses the silhouette coefficient method to quantify the diversity of the battery feature space, and the three determined cluster centers correspond to different electrochemical degradation rates. The introduction of the genetic algorithm gives the clustering process powerful global search capabilities, while the Metropolis criterion in the simulated annealing mechanism effectively avoids cluster centers getting trapped in locally dense regions in the early stages of iteration by accepting the probability function of the difference in solutions. This fusion algorithm ensures that the obtained degradation pattern clusters have extremely high intra-class similarity and inter-class discriminability, laying a homogeneous data foundation for subsequent customized modeling.

[0132] Specifically, the target number of clusters is set to K. For example, the target number of clusters can be determined using the silhouette coefficient method. For example, the number of clusters can be set to three, corresponding to the long-lived mode, the medium-lived mode, and the short-lived mode, respectively.

[0133] Based on the initial battery health characteristics, a genetic algorithm is used to encode each individual's chromosome into a set of vectors containing K cluster centers, serving as the initial population; wherein the dimension of the chromosome is determined by the dimension of the initial battery health characteristics; the expression for the initial population is:

[0134] ;

[0135] in, This is the initial population, i.e., the set of vectors representing the cluster centers; For the first Cluster center vectors; It is a real number; Dimensions of initial battery health characteristics;

[0136] The algorithm's optimization objective is to minimize the sum of weighted intra-class numerical attribute distances and categorical attribute dissimilarity. Its objective function is... The formula is:

[0137] ;

[0138] in, This is the membership matrix; To belong to cluster The An initial battery health feature vector; For cluster centers The defined first A cluster of degenerate patterns; , These are the weight coefficients for numerical features and categorical features, respectively. The numerical feature is the Euclidean distance; It is used as a measure of dissimilarity in classification features; , The first Numerical health feature component and categorical health feature component of individual battery sample feature vectors; , The first Numerical health feature component and categorical health feature component of each cluster center vector.

[0139] The fitness value is calculated using a clustering quality quantification function, expressed as:

[0140] ;

[0141] in, This is the fitness value; For cluster centers The defined first A cluster of degenerate patterns; To belong to cluster The An initial battery health feature vector; For sample weights; It is an L2 norm;

[0142] The initial population is iteratively updated based on the fitness value to seek the optimal clustering scheme and obtain the globally optimal individual.

[0143] Decoding the globally optimal individual yields the final cluster center of the initial battery health characteristics, i.e., the degradation pattern cluster.

[0144] The iterative update of the initial population includes three main steps: selection of dominant solutions, recombination of prototype vectors, and exploration of solution space perturbation, as detailed below:

[0145] Dominant solution selection: Based on the fitness value, dominant solutions are selected from the initial population to obtain two parent prototype sets;

[0146] Prototype Vector Recombination: Two parent prototype sets are recombined using a convex combination to generate a new candidate prototype set. The expression is:

[0147] ;

[0148] in, For the new candidate prototype set; These are the weighting coefficients; , For two parent prototype sets;

[0149] Solution space perturbation exploration: Based on the new candidate prototype set, random perturbation is added to the cluster center vector to expand the exploration range of the solution space, find the globally optimal clustering scheme, and output the corresponding new fitness value;

[0150] Based on the new fitness value, a simulated annealing mechanism is introduced for local optimization, and temporary deterioration is accepted with dynamic acceptance probability until the algorithm meets the convergence condition or reaches the preset maximum number of iterations, at which point the iteration terminates and the globally optimal individual is obtained.

[0151] The dynamic acceptance probability is generated based on the difference between the current annealing temperature and the intra-class dispersion to avoid the algorithm getting trapped in local optima. The expression is:

[0152] ;

[0153] in, For dynamic acceptance probability; The difference between the new and old fitness values; This is the annealing temperature for the current iteration.

[0154] An exponential cooling strategy is employed to reduce the annealing temperature until convergence, expressed as:

[0155] ;

[0156] in, This is the annealing temperature for the next round; This is the temperature cooling coefficient.

[0157] After the algorithm converges, it decodes the globally optimal individual to obtain the final K=3 cluster centers, and assigns all battery samples to the corresponding clusters, thus recognizing three aging modes: long lifespan, medium lifespan, and short lifespan. Figures 2-4 The figures shown are characteristic distribution diagrams of the State of Health (SOH) of batteries with short, medium, and long lifespans, respectively.

[0158] S4. Construct a two-stage linear-nonlinear hybrid model to fit and extract key inflection point features from the preprocessed time series data; wherein, the key inflection point features are the key feature points of the linear-nonlinear stage transition in the time series data where a trend change occurs and the slope changes significantly.

[0159] Unlike existing technologies that only use static features of the charging segment, this step parametrically extracts key "inflection point dynamic features" from the battery's full-life capacity decay trajectory. These key inflection point features are used to characterize the core quantitative features that represent the abrupt change in battery capacity decay from a slow linear phase to an accelerated linear phase.

[0160] The full-life capacity decay process of lithium-ion batteries can be divided into two linear stages, smoothly connected by a continuously differentiable nonlinear transition function, thus accurately depicting the complete degradation process of "slow decay → sudden inflection point → accelerated decay":

[0161] Phase 1 (before the inflection point): The battery is in a stable, slow degradation phase, and the capacity decreases approximately linearly with the number of cycles.

[0162] Nonlinear transition stage: Near the inflection point, the degradation rate changes abruptly, and the two stages are smoothly connected through a continuously differentiable function;

[0163] Second stage (after the inflection point): The battery enters the accelerated degradation stage, and the slope of the linear decrease in capacity with the number of cycles increases significantly.

[0164] The two-stage linear-nonlinear hybrid model uses a hyperbolic tangent cumulative distribution function as the transition function to describe the relationship between capacity and cycle number, expressed as:

[0165] ;

[0166] in, For the first Battery capacity during the next cycle; This refers to the number of battery cycles. These are the initial capacity parameters; This represents the linear decay slope in the first stage, i.e., the linear decay slope before the inflection point. This represents the linear decay slope in the second stage, i.e., the linear decay slope after the inflection point. The number of cycles at the inflection point; It is a hyperbolic tangent cumulative distribution function, used to achieve a smooth transition between two stages; For the transition rate coefficient; This is the random error term.

[0167] When extracting key inflection point features from the preprocessed time-series data, a nonlinear least squares method is used to optimize the fitting of the two-stage linear-nonlinear hybrid model to minimize the sum of squared residuals between the model's predicted values ​​and the actual observed values, thus obtaining the optimal parameter set. The formula is as follows:

[0168] ;

[0169] in, The set of parameters for the model to be fitted, including the initial capacity parameters. The linear decay slope in the first stage The linear decay slope in the second stage Inflection point cycle number With transition rate coefficient ; For minimization operation; This represents the total number of battery cycles. For the first The actual observation capacity of the next cycle; Predict capacity for the model;

[0170] Based on the optimal parameter set, key inflection point features, including the inflection point cycle number and the inflection point sharpness factor, are extracted; wherein, the inflection point sharpness factor is the first derivative of the transition function at zero, used to quantify the drastic change in the degradation rate at the inflection point, and its expression is:

[0171] ;

[0172] in, The sharpness factor at the inflection point; Indicates the number of loops Find the first derivative; The optimal linear decay slope for the second stage; The optimal inflection point cycle number; The optimal transition rate coefficient; For the transition function in The first derivative at that point.

[0173] like Figure 5 The visualization of the two-stage inflection point identification of lithium-ion battery capacity decay shown demonstrates the heterogeneity of different battery samples (dispersion of capacity curves) and reveals the common law of battery degradation (two-stage decay + inflection point mutation) through two-stage fitting and inflection point identification. This provides an intuitive dynamic basis for subsequent "mode modeling" and "inflection point adaptive optimization".

[0174] This step employs a two-stage piecewise linear-nonlinear hybrid degradation model for fitting and feature extraction, which can accurately capture the inflection point phase transition behavior characterizing the abrupt change in degradation kinetics in the capacity decay curve of lithium-ion batteries. This model smoothly connects two linear degradation stages with different slopes through a continuously differentiable transition function, thereby accurately identifying the inflection point location and phase transition intensity.

[0175] In the later stages of cycling, lithium-ion batteries exhibit a significant nonlinear acceleration trend in capacity decay due to accelerated side reactions such as lithium-ion loss and active material shedding. This invention's two-stage hybrid degradation model utilizes... The parameters precisely pinpoint the number of cycles in which this transition occurs. And... The parameters, calculated by the rate of change of the slope in the transition region, reflect the intensity of the phase transition in the battery's internal chemical state. Using these two kinetic parameters with clear physical meanings as feature inputs enables the data-driven model to sense the acceleration of battery degradation, thus compensating for the shortcomings of traditional statistical features in dynamic characterization.

[0176] S5. Based on the degradation mode cluster, evaluate the importance of the initial battery health features and the key inflection point features for each degradation mode, and select the optimal feature subset with high importance for each degradation mode.

[0177] Before assessing the importance of features, the preprocessed time-series data is standardized based on the degradation pattern cluster to reduce noise and redundancy and improve model efficiency.

[0178] The standardization process is as follows:

[0179] The 3-Sigma criterion is used to identify and remove outlier samples;

[0180] For the feature data after removing outliers, Z-score standardization is performed to eliminate differences in feature dimensions;

[0181] Let the feature matrix of the training set after removing anomalies be... The standardized formula is:

[0182] ;

[0183] in, The standardized feature matrix; It is the feature mean vector; This is the standard deviation vector.

[0184] like Figure 6 The figure shows the distribution of battery health characteristics for the three modes after standardization.

[0185] Then, based on the random forest tree model, the importance score of the standardized data features is calculated using the following formula:

[0186] ;

[0187] in, For the first 3D features Importance score; This represents the total number of decision trees in the random forest. This is the index of the b-th decision tree in the random forest; Let be the out-of-bag mean squared error of the b-th decision tree; For the b-th decision tree, the first... The new out-of-bag mean square error is obtained by randomly arranging the eigenvalues ​​of the dimensional feature.

[0188] Based on the calculated importance score, compare it with a set threshold (e.g., 0.32), and select the feature data whose importance score is greater than the set threshold, which is the optimal feature subset.

[0189] As shown in Table 2 and Figure 7 As shown, after feature selection, batteries with different lifespans retain different effective features.

[0190] Table 2. Characteristic Importance Scoring Table under Different Aging Modes

[0191]

[0192] S6. Based on the optimal feature subset, the snake optimization algorithm combined with Gaussian process regression is used to optimize the model hyperparameters.

[0193] This step uses the snake optimization algorithm (ISO-GPR) combined with Gaussian process regression to optimize the model hyperparameters.

[0194] In existing technologies, the snake optimization (SO) algorithm has been used to optimize neural network hyperparameters. However, its optimization process does not consider the differences in battery degradation modes and the specific dynamics of inflection points, resulting in limited generalization ability of the optimization results in heterogeneous data scenarios. This invention proposes a mode-adaptive collaborative optimization framework based on degradation mode identification and inflection point feature extraction steps. An improved ISO-GPR algorithm is proposed to dynamically adjust the optimization objective and search strategy according to the specific degradation behavior of the battery, achieving customized modeling with a "one model, one optimization" approach.

[0195] Detailed steps for ISO-GPR co-optimization are as follows: Figure 8 As shown, the core lies in embedding degradation mode labels and inflection point dynamic parameters into the optimization process to form a closed-loop feedback mechanism of "pattern recognition → feature extraction → collaborative optimization".

[0196] Specifically, a solution space integrating problem cognition is first constructed. The position vector of the individual snake encodes all optimization variables, that is, it simultaneously encodes the kernel function hyperparameters, binary feature selection mask vector, mode adjustment parameters, and inflection point response parameters of the Gaussian process regression model. The mode adjustment parameters include the mode convergence factor and the mode exploration coefficient; the inflection point response parameters include the inflection point proximity response factor and the inflection point sharpness sensitivity scaling factor.

[0197] Then let the current iteration number be... The maximum number of iterations is The globally optimal individual position is .

[0198] Inflection point cycle number Inflection point sharpness factor It is dual-driven, based on the current iteration number. and Based on the relative relationships, the optimization process is divided into the pre-inflection point stage, the inflection point region stage, and the post-inflection point stage:

[0199] Phase 1: When At this point, in the pre-inflection point stage, a global exploration is performed based on the center of the degradation mode cluster to adapt to the slow degradation stage before the battery inflection point. The formula is:

[0200] ;

[0201] in, This is the inflection point ratio control coefficient; , The first The individual snake was in the first , The position vector of the next iteration; For pattern exploration coefficients; It is the center of the degenerate pattern cluster; Historical experience weighting coefficient; A random number in the range of 0-1; This represents the average evolutionary direction vector of historically superior solutions under the degradation mode;

[0202] Phase Two: When At this point, during the inflection point region stage, refined local development is carried out to adapt to the abrupt change in degradation rate at the battery inflection point. The formula is:

[0203] ;

[0204] in, The radius of the neighborhood window at the inflection point; The mode convergence factor; It is a very small constant; The sharpness factor at the inflection point; for Sensitivity scaling factor; This is a sign function used to indicate the direction of position updates;

[0205] Phase Three: When At this point, in the post-inflection point stage, steady-state local development is carried out with the global optimal individual as the guide, in order to adapt to the accelerated degradation stage after the battery inflection point. The formula is:

[0206] ;

[0207] in, A random number in the range of 0-1;

[0208] A composite fitness function is constructed that integrates global trend consistency, inflection point region accuracy, and feature set quality. Through dynamic weighting, it adaptively focuses on the prediction accuracy of inflection point regions, outputting the globally optimal individual for each degradation mode. The formula for the composite fitness function is:

[0209] ;

[0210] in, This is the normalized correlation coefficient between the predicted and actual values, i.e., the global trend consistency term; This is the normalized average absolute error within the inflection point window, i.e., the accuracy term for the inflection point region. The similarity between the selected feature subset and the ideal feature subset in the pattern, i.e., the feature set quality term; , These are the dynamic weights of the global trend consistency term and the inflection point region accuracy term, respectively. The weights of the quality terms in the feature set;

[0211] The formula for the dynamic weight is:

[0212] ;

[0213] ;

[0214] in, The number of cycles at the inflection point Proximity response factor of location; is the width coefficient of the Gaussian weight.

[0215] The globally optimal individual includes the optimal hyperparameters, optimal feature subset mask, optimal mode adjustment parameters, and optimal inflection point response parameters of the Gaussian process regression model; wherein, the optimal mode adjustment parameters include the optimal mode exploration coefficients. With mode convergence factor The optimal inflection point response parameters include the optimal sensitivity scaling factor. With proximity response factor .

[0216] Traditional optimization algorithms often treat feature selection and hyperparameter optimization as two separate processes, which can easily lead to the loss of local optimal combinations. The ISO algorithm in this step achieves simultaneous optimization of the input dimension and model structure through despace encoding. During the optimization process, the three-stage position update strategy dynamically adjusts the search step size according to the actual physical degradation stage of the battery. For example, when the battery is detected to be near an inflection point region, the algorithm automatically reduces the search radius and increases the number of iterations, thereby finding configuration parameters in the complex hyperparameter space that can minimize the prediction error in the inflection point region. This adaptive focusing mechanism is the core guarantee for the high-precision estimation achieved in this invention.

[0217] S7 inputs the time series data to be tested into the hyperparameter-optimized model and outputs the battery health state estimation result.

[0218] For the time-series data of lithium-ion batteries under test, initial battery health features (such as 16-dimensional features) and key inflection point features are extracted, and the feature importance scores are evaluated using a tree model. Then, the optimal feature subsets under different aging modes are selected, and these selected features are input into the optimized ISO-GPR model for the corresponding category, outputting SOH point estimates and confidence intervals. For each mode, training, validation, and test sets are proportionally divided. The SOH estimation prediction results are as follows: Figure 9 (Short-life batteries) Figure 10 (Medium lifespan battery) Figure 11 As shown in the image (long-life battery).

[0219] In practical implementation, this invention features a highly automated operational logic. The system first monitors the external characteristic parameters of each individual battery cell in real time through a sensor network deployed within the battery management system. The pre-processed data stream is then sent to an embedded processing unit for pattern recognition and rapid classification based on pre-stored clustering prototypes. For newly added battery cells, the system initiates an inflection point kinetic extraction process, updating the parameters of the hybrid degradation model using its historical operational data. Simultaneously, the collaborative optimization modeling component performs parallel computation on a cloud server, periodically iteratively fine-tuning the model parameters for each mode based on accumulated operational data. Finally, the state estimation component distributes the optimized lightweight model to the local terminal, achieving low-latency, high-precision real-time online monitoring of the battery's health status.

[0220] Furthermore, in a preferred embodiment, compared with traditional methods that lack aging pattern recognition, outlier detection, and inflection point feature supplementation, the present invention, such as Figure 12As shown, the present invention significantly reduces the mean absolute error (MAE) and root mean square error (RMSE) of short-life, medium-life, and long-life batteries. Experimental data indicate that the present invention improves the estimation accuracy by more than 30% over the entire battery life cycle, especially in the inflection point region where battery capacity rapidly declines, where the predicted trajectory accurately matches the actual value.

[0221] In another preferred embodiment, such as Figure 13 As shown, the ISO-GPR model used in this invention is compared with the performance of other machine learning algorithms. Other machine learning algorithms include mainstream machine learning algorithms such as regression algorithms (GAM), support vector machines (SVM), convolutional neural networks (CNN), and long short-term memory networks (LSTM); the metrics include MAE (mean absolute error), RMSE (root mean square error), and MAPE (mean absolute percentage error). Among them, the smaller the value of the metric, the higher the prediction accuracy of the model, all of which show better convergence speed and estimation accuracy.

[0222] In summary, compared with the prior art, the present invention has the following beneficial effects:

[0223] This invention constructs a closed-loop, end-to-end lithium-ion battery health state estimation system through precise identification of degradation modes, physical characterization of nonlinear inflection points, and collaborative optimization based on swarm intelligence. This system not only effectively isolates the interference of heterogeneous data but also significantly improves the predictive reliability of the model in the mid-to-late stages of battery life by capturing degradation kinetic characteristics. The methods described in this invention possess strong universality and can be transferred to battery state monitoring of various material systems such as lithium iron phosphate and ternary lithium, providing solid technical support for the intelligent upgrading of battery management systems.

[0224] Example 2

[0225] like Figure 14 As shown, the second embodiment of the present invention also provides a battery health state estimation device that integrates pattern recognition and inflection point features, comprising:

[0226] The data acquisition and preprocessing module is used to acquire and preprocess the time-series data of lithium-ion battery charge-discharge cycles.

[0227] The multi-dimensional feature extraction module is used to extract initial battery health features from the charging segments of the preprocessed time-series data.

[0228] The pattern recognition module is used to divide the initial battery health features into different degradation pattern clusters using a heuristic clustering algorithm that combines genetic algorithm and simulated annealing algorithm.

[0229] The inflection point dynamics extraction module is used to construct a two-stage linear-nonlinear hybrid model to fit and extract key inflection point features from preprocessed time series data; wherein, the key inflection point features are key feature points in the time series data where abrupt trend changes and significant slope changes occur during the linear-nonlinear stage transition;

[0230] The optimal feature selection module is used to evaluate the importance of the initial battery health features and the key inflection point features based on the degradation mode cluster, and select the optimal feature subset with high importance for each degradation mode;

[0231] The collaborative optimization modeling module is used to optimize the model hyperparameters based on the optimal feature subset using a snake optimization algorithm that combines Gaussian process regression.

[0232] The state estimation module is used to input the time series data to be tested into the hyperparameter-optimized model and output the battery health state estimation results.

[0233] Example 3

[0234] The third embodiment of the present invention also provides a battery health state estimation device that integrates pattern recognition and inflection point features, which includes a memory and a processor. The memory stores a computer program that can be executed by the processor to realize the battery health state estimation method that integrates pattern recognition and inflection point features as described above.

[0235] Example 4

[0236] The fourth embodiment of the present invention also provides a computer-readable storage medium storing computer-readable instructions. When the computer-readable instructions are executed by the processor of the device where the computer-readable storage medium is located, the battery health status estimation method based on the fusion pattern recognition and inflection point features described above is implemented.

[0237] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A battery health state estimation method integrating pattern recognition and inflection point features, characterized in that, include: S1, acquire and preprocess the timing data of the lithium-ion battery charge-discharge cycle; S2, extract initial battery health features from the charging segments of the preprocessed time-series data; S3, using a heuristic clustering algorithm that combines genetic algorithm and simulated annealing algorithm, the initial battery health characteristics are divided into different degradation mode clusters; S4, construct a two-stage linear-nonlinear hybrid model to fit and extract key inflection point features from the preprocessed time series data; wherein, the key inflection point features are the key feature points of the linear-nonlinear stage transition in the time series data where a trend change occurs and the slope changes significantly; the key inflection point features include the inflection point cycle number and the inflection point sharpness factor. S5, based on the degradation mode cluster, evaluate the importance of the initial battery health characteristics by degradation mode, and select the optimal feature subset with high importance for each degradation mode; S6. Based on the optimal feature subset and the key inflection point features, the snake optimization algorithm combined with Gaussian process regression is used to optimize the model hyperparameters. S7: Input the time series data to be tested into the hyperparameter-optimized model and output the battery health state estimation result; In the process of optimizing model hyperparameters using a snake optimization algorithm combined with Gaussian process regression, the number of inflection point iterations is used as the key factor. Inflection point sharpness factor It is dual-driven, based on the current iteration number. and Based on the relative relationships, the optimization process is divided into the pre-inflection point stage, the inflection point region stage, and the post-inflection point stage: when At this point, in the pre-inflection point stage, a global exploration is performed based on the center of the degradation mode cluster to adapt to the slow degradation stage before the battery inflection point. The formula is: ； in, This is the inflection point ratio control coefficient; , The first The individual snake was in the first , The position vector of the next iteration; For pattern exploration coefficients; It is the center of the degenerate pattern cluster; Historical experience weighting coefficient; A random number in the range of 0-1; This represents the average evolutionary direction vector of historically superior solutions under the degradation mode; This represents the current iteration number; when At this point, during the inflection point region stage, refined local development is carried out to adapt to the abrupt change in degradation rate at the battery inflection point. The formula is: ； in, The radius of the neighborhood window at the inflection point; The mode convergence factor; It is a very small constant; The sharpness factor at the inflection point; for Sensitivity scaling factor; This is a sign function used to indicate the direction of position updates; The globally optimal individual position; when At this point, in the post-inflection point stage, steady-state local development is carried out with the global optimal individual as the guide, in order to adapt to the accelerated degradation stage after the battery inflection point. The formula is: ； in, A random number in the range of 0-1.

2. The battery health state estimation method based on the fusion of pattern recognition and inflection point features according to claim 1, characterized in that... The preprocessing includes aligning and filtering the time-series data; The initial battery health characteristics include: time-related characteristics, charge-related characteristics, voltage curve characteristics, and current curve characteristics.

3. The battery health state estimation method based on the fusion of pattern recognition and inflection point features according to claim 1, characterized in that... S3 specifically refers to: Set the target number of clusters to K; Based on the initial battery health characteristics, a genetic algorithm is used to encode each individual's chromosome into a set of vectors containing K cluster centers, serving as the initial population; wherein the dimension of the chromosome is determined by the dimension of the initial battery health characteristics; the expression for the initial population is: ； in, This is the initial population, i.e., the set of vectors representing the cluster centers; For the first Cluster center vectors; It is a real number; Dimensions of initial battery health characteristics; The optimization is performed with the objective function being the minimization of the sum of weighted intra-class numerical attribute distances and categorical attribute dissimilarity. The formula is: ； in, This is the membership matrix; To belong to cluster The An initial battery health feature vector; For cluster centers The defined first A cluster of degenerate patterns; , These are the weight coefficients for numerical features and categorical features, respectively. The numerical feature is the Euclidean distance; It is used as a measure of dissimilarity in classification features; , The first Numerical health feature component and categorical health feature component of individual battery sample feature vectors; , The first Numerical health feature component and categorical health feature component of each cluster center vector; The fitness value is calculated using a clustering quality quantification function, expressed as: ； in, This is the fitness value; For cluster centers The defined first A cluster of degenerate patterns; To belong to cluster The An initial battery health feature vector; For sample weights; It is an L2 norm; The initial population is iteratively updated based on the fitness value to seek the optimal clustering scheme and obtain the globally optimal individual. Decoding the globally optimal individual yields the final cluster center of the initial battery health characteristics, i.e., the degradation pattern cluster.

4. The battery health state estimation method based on the fusion of pattern recognition and inflection point features according to claim 3, characterized in that... Iterative updates to the initial population include: Based on the fitness value, the initial population is screened for dominant solutions to obtain two parent prototype sets; The prototype vectors of two parent prototype sets are recombined through convex combination to generate a new candidate prototype set, expressed as: ； in, For the new candidate prototype set; These are the weighting coefficients; , For two parent prototype sets; Based on the new candidate prototype set, random perturbation is added to the cluster center vector to expand the exploration range of the solution space, find the globally optimal clustering scheme, and output the corresponding new fitness value; Based on the new fitness value, a simulated annealing mechanism is introduced for local optimization, and temporary deterioration is accepted with dynamic acceptance probability until the algorithm meets the convergence condition or reaches the preset maximum number of iterations, at which point the iteration terminates and the globally optimal individual is obtained.

5. The battery health state estimation method based on the fusion of pattern recognition and inflection point features according to claim 4, characterized in that... The dynamic acceptance probability is generated based on the difference between the current annealing temperature and the intra-class dispersion to avoid the algorithm getting trapped in local optima. The expression is: ； in, For dynamic acceptance probability; The difference between the new and old fitness values; This is the annealing temperature for the current iteration. An exponential cooling strategy is employed to reduce the annealing temperature until convergence, expressed as: ； in, This is the annealing temperature for the next round; This is the temperature cooling coefficient.

6. The battery health state estimation method based on the fusion of pattern recognition and inflection point features according to claim 1, characterized in that... The two-stage linear-nonlinear hybrid model uses a hyperbolic tangent cumulative distribution function as the transition function to describe the relationship between capacity and cycle number, expressed as: ； in, For the first Battery capacity during the next cycle; This refers to the number of battery cycles. These are the initial capacity parameters; This represents the linear decay slope in the first stage, i.e., the linear decay slope before the inflection point. This represents the linear decay slope in the second stage, i.e., the linear decay slope after the inflection point. The number of cycles at the inflection point; It is a hyperbolic tangent cumulative distribution function, used to achieve a smooth transition between two stages; For the transition rate coefficient; This is the random error term; When extracting key inflection point features from the preprocessed time-series data, a nonlinear least squares method is used to optimize the fitting of the two-stage linear-nonlinear hybrid model to minimize the sum of squared residuals between the model's predicted values ​​and the actual observed values, thus obtaining the optimal parameter set. The formula is as follows: ； in, The set of parameters for the model to be fitted, including the initial capacity parameters. The linear decay slope in the first stage The linear decay slope in the second stage Inflection point cycle number With transition rate coefficient ; For minimization operation; This represents the total number of battery cycles. For the first The actual observation capacity of the next cycle; Predict capacity for the model; Based on the optimal parameter set, key inflection point features, including the inflection point cycle number and the inflection point sharpness factor, are extracted; wherein, the inflection point sharpness factor is the first derivative of the transition function at zero, used to quantify the drastic change in the degradation rate at the inflection point, and its expression is: ； in, The sharpness factor at the inflection point; Indicates the number of loops Find the first derivative; The optimal linear decay slope for the second stage; The optimal inflection point cycle number; The optimal transition rate coefficient; For the transition function in The first derivative at that point.

7. The battery health state estimation method based on the fusion of pattern recognition and inflection point features according to claim 6, characterized in that... Before assessing the importance of features, the preprocessed time series data is standardized based on the degradation pattern cluster. Then, based on the random forest tree model, the importance score of the standardized data features is calculated using the following formula: ； in, For the first 3D features Importance score; This represents the total number of decision trees in the random forest. This is the index of the b-th decision tree in the random forest; Let be the out-of-bag mean squared error of the b-th decision tree; For the b-th decision tree, the first... The new out-of-bag mean square error is obtained by randomly arranging the eigenvalues ​​of the dimensional feature. Based on the calculated importance score, compare it with a set threshold, and select the feature data whose importance score is greater than the set threshold, which is the optimal feature subset.

8. The battery health state estimation method based on the fusion of pattern recognition and inflection point features according to claim 7, characterized in that... , based on the inflection point cycle number Inflection point sharpness factor After optimizing the model hyperparameters for the pre-inflection point stage, inflection point region stage, and post-inflection point stage for dual-drive, a composite fitness function is constructed that integrates global trend consistency, inflection point region accuracy, and feature set quality. Through dynamic weight adaptive focusing on the prediction accuracy of the inflection point region, the global optimal individual for each degradation mode is output. The formula for the composite fitness function is: ； in, This is the normalized correlation coefficient between the predicted and actual values, i.e., the global trend consistency term; This is the normalized average absolute error within the inflection point window, i.e., the accuracy term for the inflection point region. The similarity between the selected feature subset and the ideal feature subset in the pattern, i.e., the feature set quality term; , These are the dynamic weights of the global trend consistency term and the inflection point region accuracy term, respectively. The weights of the quality terms in the feature set; The formula for the dynamic weight is: ； ； in, The number of cycles at the inflection point Proximity response factor of location; This is the width coefficient for the Gaussian weights; The globally optimal individual includes the optimal hyperparameters, optimal feature subset mask, optimal mode adjustment parameters, and optimal inflection point response parameters of the Gaussian process regression model; wherein, the optimal mode adjustment parameters include the optimal mode exploration coefficients. With mode convergence factor The optimal inflection point response parameters include the optimal sensitivity scaling factor. With proximity response factor .

9. The battery health state estimation method based on the fusion of pattern recognition and inflection point features according to claim 7, characterized in that... The standardization process is as follows: The 3-Sigma criterion is used to identify and remove outlier samples; For the feature data after removing outliers, Z-score standardization is performed to eliminate differences in feature dimensions; Let the feature matrix of the training set after removing anomalies be... The standardized formula is: ； in, The standardized feature matrix; It is the feature mean vector; This is the standard deviation vector.

10. A battery health state estimation device integrating pattern recognition and inflection point features, used to implement the battery health state estimation method integrating pattern recognition and inflection point features as described in any one of claims 1-9, characterized in that, include: The data acquisition and preprocessing module is used to acquire and preprocess the time-series data of lithium-ion battery charge-discharge cycles. The multi-dimensional feature extraction module is used to extract initial battery health features from the charging segments of the preprocessed time-series data. The pattern recognition module is used to divide the initial battery health features into different degradation pattern clusters using a heuristic clustering algorithm that combines genetic algorithm and simulated annealing algorithm. The inflection point dynamics extraction module is used to construct a two-stage linear-nonlinear hybrid model to fit and extract key inflection point features from preprocessed time series data. The key inflection point features are key feature points in the time series data where a trend change occurs or the slope changes significantly during the linear-nonlinear stage transition. The key inflection point features include the inflection point cycle number and the inflection point sharpness factor. The optimal feature selection module is used to evaluate the importance of the initial battery health features based on the degradation mode cluster, and select the optimal feature subset with high importance for each degradation mode; The collaborative optimization modeling module is used to optimize the model hyperparameters based on the optimal feature subset and the key inflection point features, using a snake optimization algorithm that combines Gaussian process regression. The state estimation module is used to input the time series data to be tested into the hyperparameter-optimized model and output the battery health state estimation results.

Images