A method for predicting state of health of a retired power battery

Abstract

The application provides a kind of decommissioned power battery SOH prediction method, comprising: obtaining the EIS test data of the decommissioned power battery to be measured;Data processing and feature extraction are carried out on the EIS test data to obtain input features;The input features are input into the decommissioned power battery SOH prediction model constructed, and the maximum capacity value of the decommissioned power battery to be measured is predicted by the decommissioned power battery SOH prediction model to determine the SOH of the decommissioned power battery to be measured.This scheme can solve the current problem of decommissioned power battery SOH prediction model, under the premise of considering prediction accuracy and application portability, using the decommissioned power battery SOH prediction technology based on multi-frequency EIS feature data and algorithm combination adaptive optimization function, a hierarchical and well-defined adaptive processing flow is constructed by dividing into four different stages, realizing high-precision prediction of EIS feature data of decommissioned power battery under any SOC state.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of battery SOH detection, in particular to a retired power battery SOH prediction method. BACKGROUND

[0002] Power batteries are the core components of new energy vehicles. Among them, lithium batteries are widely used in electric vehicles due to their high energy density, high charging efficiency, long cycle life, and lightness. However, with the rapid development of the new energy vehicle industry, the amount of retired power batteries has also increased year by year.

[0003] However, the current power battery recycling industry is still developing slowly, the main reason is that the technical barriers of cascade utilization are high, and the life prediction technology is the most critical problem. The capacity degradation mechanism of power batteries is influenced by multiple factors, resulting in differences in aging degree among retired battery monomers, which increases the difficulty of detecting the remaining life of battery cascade utilization. Therefore, developing a high-precision, fast, and low-cost health status evaluation method for retired power batteries is of great significance to promote the rapid development of the power battery recycling industry.

[0004] The health status (SOH) of power batteries refers to the ratio of the current maximum capacity to the rated capacity. The state of charge (SOC) of power batteries refers to the ratio of the current remaining capacity to the rated capacity. Referring to Figure 1 .

[0005] SOH represents the current battery's ability to store energy relative to a new battery, and SOC represents the available state of the remaining charge in the current state. SOH can measure the aging state of the battery, and accurate evaluation of SOH is the prerequisite and guarantee for the cascade utilization of retired power batteries.

[0006] Although there are many classification methods for SOH prediction, they can be mainly divided into the following three categories:

[0007] The first type is definition method

[0008] As a direct and simple method, the definition method strictly follows the regulations of the battery test manual under standard conditions and performs a complete charge and discharge test process. This process can obtain the discharge capacity of the battery, which is compared with the rated capacity of the new battery. The resulting ratio represents the health status (SOH) of the battery in the current state. Although this method has significant advantages in accuracy, it is time-consuming to operate, requires a complete and error-free charge and discharge process, and relies on high-precision equipment, making it difficult to popularize in practical application scenarios. However, in the research field, it can serve as a benchmark tool to evaluate the accuracy of other estimation methods and has important reference value.

[0009] The second type: model-based method (model method)

[0010] The model-based method mainly includes electrochemical model method, equivalent circuit model method (ECM) and mathematical model method.

[0011] The electrochemical model method is divided into two types based on aging mechanism and electrochemical impedance spectroscopy (EIS), the former simulates the changes of key parameters in the battery aging process, and the latter evaluates SOH through alternating current impedance.

[0012] The ECM method is based on the electrical characteristics of the battery, and uses resistors, capacitors and other components to build a model. Common models include Rint, Thevenin and first-order RC model, etc. The SOH is obtained through parameter identification and algorithm processing.

[0013] The mathematical model method is divided into empirical model and probability model. The empirical model relates the life decay to multiple decay factors, and data fitting improves the prediction accuracy. The probability model is based on probability theory, analyzes the change of voltage probability density in the charge and discharge curve, and establishes the corresponding relationship between peak voltage and SOH.

[0014] The above methods need to select appropriate models and parameter identification methods according to specific circumstances to achieve accurate SOH prediction.

[0015] The third type: data-driven method

[0016] The data-driven method is mainly divided into artificial intelligence-based, filtering-based and statistical data-based methods. The artificial intelligence-based method includes ANN, SVM, LSTM, RNN, etc., which improves the prediction effect by improving the algorithm, but is mainly used to process real-time data such as battery management system (BMS). The filtering-based method such as EKF, PF, etc. effectively handles nonlinear systems. The statistical data-based method such as GPR, WP is suitable for complex regression and random process research. The time series-based method such as ARMA, RNN, LSTM, etc. uses historical data to predict future trends. Various methods need to be selected and optimized according to actual circumstances to achieve accurate SOH prediction and provide support for lithium-ion battery health management.

[0017] The detailed classification of the above three methods and their existing problems are shown in the following table 1.

[0018] Table 1. Comparison of SOH prediction method classification

[0019]

[0020] In general, the data-driven method is more suitable for real-time monitoring data such as vehicle-mounted BMS, and cannot be applied to the SOH prediction scenario of retired power battery cells (after disassembly). The model method is suitable for SOH prediction of independent cells after disassembly, but the prediction accuracy is low due to the simple modeling of the model output data, and it is often based on some typical characteristics to directly predict SOH through binomial regression; at the same time from the perspective of electrochemistry, the stronger the data representation ability of the model, the worse the portability in the actual application stage. For example, a high-precision SOH of a certain type of battery is achieved based on a certain ECM model, but the same high-precision prediction cannot be achieved when the model is transplanted to another type of battery. SUMMARY

[0021] The purpose of the embodiments of the present application is to provide a retired power battery SOH prediction method, which solves the common problems of current retired power battery SOH prediction models, and under the premise of considering prediction accuracy and application portability, uses a retired power battery SOH prediction technology based on multi-frequency EIS characteristic data and algorithm combination adaptive optimization function, and constructs a hierarchical and clear adaptive processing flow through four different stages, to realize high-precision prediction of EIS characteristic data of retired power batteries in any SOC state.

[0022] In order to achieve the above purpose, the embodiments of the present application are implemented by the following ways:

[0023] In a first aspect, the embodiments of the present application provide a retired power battery SOH prediction method, comprising: obtaining EIS test data of a to-be-tested retired power battery; performing data processing and feature extraction on the EIS test data to obtain input features; inputting the input features into a constructed retired power battery SOH prediction model, and predicting the maximum capacity value of the to-be-tested retired power battery through the retired power battery SOH prediction model to determine the SOH of the to-be-tested retired power battery.

[0024] In a first possible implementation manner of the first aspect, a construction procedure of the retired power battery SOH prediction model comprises: obtaining an EIS dataset composed of EIS data of N retired battery samples, wherein the EIS dataset contains N sub-tables, each sub-table has M rows and L columns, M is a product of a sample point quantity p and a test number q, L is 4, and L respectively represents a frequency, a real part, a reciprocal of an imaginary part, and a current power of the retired battery sample in a test state; performing one-dimensional feature alignment processing on the N sub-tables in the EIS dataset to obtain primary feature data, wherein the primary feature data is total table data containing N×q rows and 2p+1 columns; performing data preprocessing on the primary feature data to obtain secondary feature data; based on the secondary feature data, performing adaptive optimization on a plurality of candidate feature processing algorithms to determine a target feature processing algorithm, so as to perform feature processing on the secondary feature data by using the target feature processing algorithm to obtain complete feature data; based on the complete feature data, performing optimization on a plurality of candidate modeling algorithms to determine a target modeling algorithm, so as to perform modeling by using the target modeling algorithm, and perform model training and testing by using the complete feature data, thereby constructing the retired power battery SOH prediction model.

[0025] In a second possible implementation manner of the first aspect, in the first possible implementation manner of the first aspect, the one-dimensional feature alignment processing on the N sub-tables in the EIS dataset to obtain the primary feature data comprises: traversing the N sub-tables in the EIS dataset, and for each sub-table: extracting parameters in a second column and parameters in a third column in a p row unit to process as a 1×2p matrix, determining a q×2p matrix corresponding to the current sub-table according to the matrix, and searching for a maximum value in a fourth column in the current sub-table to construct a q×1 matrix as a label value and add the label value to a rightmost side of the q×2p matrix corresponding to the current sub-table to obtain a q×(2p+1) matrix as feature data corresponding to the current sub-table; and splicing the feature data corresponding to the N sub-tables to obtain a (N×q)×(2p+1) matrix as the primary feature data corresponding to the EIS dataset.

[0026] In a third possible implementation manner of the first aspect, in the first possible implementation manner of the first aspect, the data preprocessing on the primary feature data to obtain the secondary feature data comprises: performing data cleaning on the primary feature data, including processing of missing values and abnormal values; performing data scaling on the primary feature data after the data cleaning, wherein a data scaling method is Z-Score standardization; and performing data inspection on the primary feature data after the data scaling to obtain the secondary feature data, wherein a data inspection method is a two-sample KS test.

[0027] In a fourth possible implementation manner of the first aspect, based on the secondary feature data, the adaptive optimization is performed on the multiple candidate feature processing algorithms to determine the target feature processing algorithm, including: performing feature extraction on the secondary feature data by using each candidate feature processing algorithm to determine candidate feature data corresponding to each candidate feature processing algorithm; performing data set construction on the candidate feature data corresponding to each candidate feature processing algorithm to determine a theoretical experimental data set and an engineering application data set corresponding to each candidate feature processing algorithm; and based on each candidate feature processing algorithm, the corresponding engineering application data set and each candidate modeling algorithm, using an adaptive feature optimization module to determine the target feature processing algorithm, wherein the adaptive feature optimization module is based on the idea of a greedy algorithm, and uses a global sorting processing algorithm and a feature priority processing algorithm to integrate processing of the candidate feature processing algorithm.

[0028] In a fifth possible implementation manner of the first aspect, the data set construction is performed on the candidate feature data corresponding to each candidate feature processing algorithm to determine the theoretical experimental data set and the engineering application data set corresponding to each candidate feature processing algorithm, including: for the candidate feature data corresponding to each candidate feature processing algorithm: randomly sorting all rows in the candidate feature data, and dividing the candidate feature data into a training and verification set and a test set according to a ratio of 8:2, and then performing K-fold division on the training and verification set to obtain K pairs of training set and verification set combinations, combining the test set, the K pairs of training set and verification set combinations to obtain the corresponding theoretical experimental data set; and for the candidate feature data corresponding to each candidate feature processing algorithm: in the N×q rows of the candidate feature data, taking all q rows of data of one retired battery sample as a test set, and in the remaining (N-1)×q rows, grouping the q rows of each retired battery sample to split into (N-2)×q rows of training set of (N-2) retired battery samples and q rows of verification set of one retired battery sample, and taking different retired battery samples as the test set repeatedly N-1 times to obtain N groups of training set, verification set and test set as the engineering application data set.

[0029] In a sixth possible implementation manner of the first aspect, based on each alternative feature processing algorithm, the corresponding engineering application dataset, and each alternative modeling algorithm, the target feature processing algorithm is determined by using the adaptive feature optimization module, including: using the alternative feature processing algorithm to complete data construction in the feature processing algorithm dimension, using the alternative modeling algorithm to complete data construction in the modeling algorithm dimension, and based on the engineering application dataset and the idea of traversal loop test, data construction in the battery dimension is completed, forming a three-dimensional data containing three dimensions X, Y, and Z, denoted as a three-dimensional array F1, the three dimensions X, Y, and Z respectively represent the feature processing algorithm dimension, the modeling algorithm dimension, and the battery dimension, wherein F1(x, y, z) is a corresponding cell data list of the three-dimensional array F1, containing two floating-point value data of the test set precision mean and the test set precision minimum, represented by F1(x, y, z)[0] and F1(x, y, z)[1] respectively; in the case of z = i, all values of x and y are traversed, it is judged whether the minimum precision F1(x, y, i)[1] with x and y as independent variables is greater than a threshold value, if the condition is met, the average precision F1(x, y, i)[0] is recorded, all values of z in the z-axis direction are traversed, and the corresponding operation is repeated, and the recorded F1(x, y, i)[0] is sequentially filled into G1(x, y), that is, the two-dimensional conversion of the three-dimensional array is completed, and finally the two-dimensional array G1 representing the prediction effect of the algorithm combination is obtained, G1(x, y) is a corresponding cell data list of the two-dimensional array G1, which is composed of F1(x, y, i)[0] determined when z meets the condition at each value; the cell data list of G1(x, y) is averaged to obtain the average precision G1_acc(x, y) of the current algorithm combination on the engineering application dataset, all x and y are traversed to obtain G1_acc, and the length of the data list of G1(x, y) is recorded to obtain the number G1_num(x, y) of the current algorithm combination meeting the standard on the engineering application dataset, and all x and y are traversed to obtain G1_num; based on the priority order of G1_num and G1_acc, the first m1 algorithm combinations are screened out, and the corresponding feature processing algorithm x_list1 is recorded as the output result of the three-dimensional global sorting algorithm; for the G1_num two-dimensional array, all values of y are traversed to calculate , and the first m2 values with the largest value of the feature processing algorithm x_list2 are obtained, that is, the output result of the feature priority sorting algorithm; the highest precision mean in the intersection of the x_list1 and x_list2 sets is determined, if the intersection is empty, the highest precision mean in x_list1 and x_list2 is taken, as the output of the target feature processing algorithm determined by the adaptive feature optimization module.

[0030] In a seventh possible implementation of the first aspect, based on the complete feature data, the target modeling algorithm is determined by optimizing a plurality of alternative modeling algorithms, comprising: taking the theoretical experimental data set and the engineering application data set corresponding to the target feature processing algorithm as the complete feature data; and determining the target modeling algorithm and corresponding parameters by using an algorithm combination optimization module based on the target feature processing algorithm, the complete feature data, each alternative modeling algorithm, an algorithm parameter space, and a parameter optimization manner, wherein the algorithm combination optimization module adaptively selects grid search or random search for parameter optimization according to the number of parameters in the hyperparameter optimization process of the modeling algorithm.

[0031] In an eighth possible implementation of the first aspect, based on the target feature processing algorithm, the complete feature data, each alternative modeling algorithm, the algorithm parameter space, and the parameter optimization manner, the target modeling algorithm and corresponding parameters are determined by using the algorithm combination optimization module, comprising: taking the theoretical experimental data set and the engineering application data set corresponding to the target feature processing algorithm as the complete feature data; based on the combination of the target feature processing algorithm and each alternative modeling algorithm, different algorithm combinations are formed to complete data construction in the algorithm combination dimension, according to each algorithm combination, a corresponding parameter search space is generated, and according to the number of parameters in the parameter search space, a parameter optimization manner is adaptively selected to complete data construction in the algorithm parameter dimension, and according to the number of retired battery samples, the engineering application data construction method is used to complete data construction in the battery dimension to form a three-dimensional array F2 containing U, V, and Z dimensions, the three dimensions U, V, and Z respectively represent the algorithm combination dimension, the algorithm parameter dimension, and the battery dimension, wherein F2(u, v, z) is the model prediction result based on the current u, v, and z, F2(u, v, z)[0] is the test set accuracy mean value, F2(u, v, z)[1] is the test set accuracy minimum value, and F2(u, v, z)[2] is the R 2 Determination coefficient; by determining and compressing the z-axis data, two-dimensional conversion of three-dimensional data is realized to obtain G2_num, G2_acc, and G2_R 2 Three two-dimensional tables; based on the priority of G2_num, G2_acc, and G2_R 2 The optimal algorithm combination and the corresponding optimal parameter list are calculated to determine the target modeling algorithm and corresponding parameters as the output of the algorithm combination optimization module.

[0032] In a ninth possible implementation of the first aspect, the model optimization index comprises a loss function for optimizing the model training fitting effect, and an evaluation index for measuring the model training effect,

[0033] The loss function is designed as:

[0034] ,

[0035] wherein the weight hyperparameter is 0.6;

[0036] ,

[0037] ,

[0038] wherein the hyperparameter is 1, represents the true value of the i-th sample, represents the model prediction value of the i-th sample, represents the sample number, ; The evaluation index is designed as:

[0039]

[0040] ,

[0041] ,

[0042] wherein, represents the accuracy of the prediction value;

[0043] Let the set M={1,2,3,…,m}, M1={i∈M|Acc i >λ1}, M2={i∈M|R 2 i The selection formula of the optimal model i* is as follows:

[0044] ,

[0045] wherein the hyperparameter λ1 is 0.98, λ2 is 0.6, and when the optimal model is empty, λ2 is adjusted or the data is re-screened and processed to train the model.

[0046] ​​Beneficial effects: According to the data characteristics of EIS, the engineering application data construction method is used to divide the training set, the verification set and the test set, so as to ensure that the model can learn the complete data characteristics of 0~100% SOC without deviation in the training process. Then, the leave-one-out cross-validation based on the cyclic traversal test is used to ensure that the model can achieve the accuracy standard when using any battery as the test set, which proves that the model has high robustness and strong generalization. Through the above improvements, the SOH of EIS characteristics at any SOC is predicted with high precision, which fills the practical significance of operability for the SOH prediction technology based on EIS measurement data. In this scheme, different stages are divided to construct a self-adaptive processing flow with rich levels and clear division. The feature algorithm, modeling algorithm, evaluation method and parameter selection are regarded as a whole variable, and the physical maximum prediction accuracy of battery data is approached through multivariate setting, which is close to the global optimal solution. At the same time, based on the greedy algorithm, an adaptive optimization module is constructed to effectively solve the dimension disaster and reduce the time complexity, and a high-precision analysis and prediction scheme is customized for battery data. Through the algorithm of this scheme, the average prediction accuracy of EIS data of four types of batteries at various SOC states reaches 99.11% (±0.16%), and the average deviation of the remaining capacity diagnosis is less than 1%.

[0047] For the two-dimensional data matrix that can be used for model training, the general construction method of the data set is to randomly sort all rows and then divide the training set and the test set according to a certain proportion. However, for the EIS feature data of N×q rows and 2p+1 columns in this scheme, the data is composed of N batteries of the same type with different SOH. Therefore, although the number of label values is N×q, there are only N different values. Therefore, randomly dividing the data set according to the conventional method will lead to inconsistent distribution of the training set and the test set, which will make the model only learn the EIS data characteristics at part of the SOC. In order to learn the EIS characteristics at any SOC and realize SOH prediction, on the basis of the conventional data set construction method, according to the characteristics that multiple features of EIS data correspond to one label, an engineering application data construction method is proposed to ensure that the model can learn the data characteristics at full SOC distribution evenly and without deviation during data division. At the same time, in order to adapt to this construction method in the model verification stage, a leave-one-out cross-validation based on cyclic traversal test method is proposed, which strictly verifies N groups of data sets to greatly ensure the robustness of the model.

[0048] To improve the prediction accuracy of the model as much as possible, and make the model prediction ability approach the physical upper limit of the data, the idea of constructing the optimal solution space of multiple variables is adopted. The feature algorithm, modeling algorithm, evaluation and verification method and various parameter selection are regarded as variables that need to be considered in the whole scheme. Through the setting of a large number of variables, it is ensured that the data distribution composed of various variables can approach the physical maximum prediction accuracy of the current battery data as much as possible, and the effect of approximate global optimal solution is achieved. On the basis of the above, in order to solve the problems of dimension disaster, time complexity and the like brought by multiple variables, a dimension reduction optimization method based on greedy algorithm is proposed, and adaptive feature optimization module and optimal algorithm combination optimization module are constructed based on different application data, realizing the adaptive optimization selection of key algorithms and parameters of the model under low time complexity.

[0049] In order to make the above-mentioned purposes, features and advantages of the present application more obvious and easy to understand, the following preferred embodiments are specifically described below, and the accompanying drawings are described in detail as follows. BRIEF DESCRIPTION OF DRAWINGS

[0050] In order to more clearly illustrate the technical solutions of the embodiments of the present application, the following will briefly introduce the drawings needed to be used in the embodiments of the present application. It should be understood that the following drawings only show some embodiments of the present application, and therefore should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can also be obtained without creative labor on the basis of these drawings.

[0051] Figure 1 The schematic diagram of battery SOC and SOH.

[0052] Figure 2 The overall idea diagram of the retired power battery SOH prediction method.

[0053] Figure 3 The flowchart of constructing the retired power battery SOH prediction model.

[0054] Figure 4 The schematic diagram of one-dimensional feature alignment processing of data.

[0055] Figure 5 The flowchart of the retired power battery SOH prediction method. DETAILED DESCRIPTION

[0056] The technical solutions in the embodiments of the present application will be described below in combination with the drawings in the embodiments of the present application.

[0057] Please refer to Figure 2 , Figure 2 The overall idea diagram of the retired power battery SOH prediction method.

[0058] In the present embodiment, the retired power battery SOH prediction method is roughly divided into four stages: data measurement stage, data processing stage, modeling and prediction stage, and evaluation and verification stage. Before running the retired power battery SOH prediction method, a retired power battery SOH prediction model needs to be constructed, and the process of constructing the retired power battery SOH prediction model is as shown in Figure 3

[0059] In the present embodiment, the process of constructing the retired power battery SOH prediction model includes steps S11, S12, S13, S14 and S15.

[0060] Firstly, step S11 can be run.

[0061] Step S11: Obtain an EIS data set composed of EIS data of N retired battery samples, wherein the EIS data set contains N sub-tables, and the data of each sub-table is M rows and L columns, M is the product of the number of samples p and the number of tests q, and L is 4, which respectively represents the frequency, the real part, the inverse of the imaginary part and the current power of the retired battery sample in the test state.

[0062] In the present embodiment, in the data measurement stage, an EIS data set composed of EIS data of N retired battery samples can be obtained.

[0063] EIS (Electrochemical Impedance Spectroscopy) can reflect the detailed information of the internal electrochemical reaction of the battery, and its detection technology can realize non-destructive detection, so in recent years it has become the best data acquisition method in the field of retired power battery research. The electrochemical workstation measurement stage is to measure the EIS data of the aged retired power battery using an electrochemical workstation. The processing steps of this stage are as follows:

[0064] ① Use the charge-discharge machine to perform a complete CC-CV (constant current-constant voltage) charge-discharge test to obtain the maximum capacity of the battery after aging, and calculate the SOH of the current battery.

[0065] ② After discharging to 0% again, then charging to 5%, and then using the electrochemical workstation to perform EIS test after standing for one hour. The test frequency interval is set to 0.1 Hz~10000 Hz, and the number of samples is set to 51.

[0066] ③ Charge 5% again, and perform EIS test after standing according to the above method. (Repeat this step until the battery power is 100%. )

[0067] ​④Complete the EIS data collection of the battery, export all EIS characteristic data of the same batch of batteries using the electrochemical workstation, and use it as the initial data for subsequent modeling, that is, the EIS data set composed of EIS data.

[0068] Thus, the EIS data set composed of EIS data of N retired battery samples can be obtained, and the EIS data set contains N sub-tables, and the data of each sub-table is M rows and L columns, M is the product of the number of samples p and the number of tests q, L is 4, respectively representing the frequency, the real part, the reciprocal of the imaginary part, and the current power of the retired battery sample in the test state. In this embodiment, p is 51 and q is 21, and other embodiments can take other values, for example, to obtain higher precision training data, and to measure once per one hundredth, then q is 101, and this embodiment is not limited.

[0069] After that, step S12 can be performed.

[0070] Step S12: One-dimensional feature alignment processing is performed on the N sub-tables in the EIS data set to obtain primary feature data, wherein the primary feature data is a total table data containing N×q rows and 2p+1 columns.

[0071] In this embodiment, the N sub-tables in the EIS data set can be traversed, and for each sub-table: the parameters in the second column and the parameters in the third column are extracted in p units, processed into a 1×2p matrix, and the q×2p matrix corresponding to the current sub-table is determined accordingly. In the current sub-table, find the maximum value in the fourth column, construct a q×1 matrix as a label value and add it to the rightmost side of the q×2p matrix corresponding to the current sub-table, obtain a q×(2p+1) matrix as the feature data corresponding to the current sub-table, and then splice the feature data corresponding to the N sub-tables to obtain a (N×q)×(2p+1) matrix as the primary feature data corresponding to the EIS data set.

[0072] Specifically, as shown in Figure 4 Data featureization refers to processing the initial data file exported by the electrochemical workstation into a data format that meets the model training specification. The one-dimensional feature alignment processing steps used in this scheme are as follows:

[0073] ①Traverse each sub-table of the initial file, and for each sub-table, extract the real part in the second column and the reciprocal of the imaginary part in the third column in p (such as 51) units.

[0074] ②The obtained p×2 (such as 51×2) matrix is processed into a 1×2p (such as 1×102) matrix as the feature of each row.

[0075] ③Repeat the above operation in the sub-table, and the same processing is performed on the remaining q-1 (such as 21-1) different EIS data under different electric quantities to obtain a qx2p (such as 21x102) matrix capable of representing the first sub-table (battery) data.

[0076] ④Find the maximum value of the fourth column in the current sub-table of the initial file, and construct a qx1 matrix as a label value and add it to the rightmost side of the qx2p matrix. At this time, the qx (2p+1) matrix is the all tuples of the current sub-table of the initial file after data characterization.

[0077] ⑤Repeat the above operation for the other remaining sub-tables, and the qx (2p+1) matrix obtained each time is sequentially appended below the matrix. Finally, a (Nxq)x (2p+1) matrix capable of representing the data of all sub-tables is obtained. The data after data characterization has N x q rows and 2p+1 columns.

[0078] Accordingly, the one-dimensional feature alignment processing of the N sub-tables in the EIS data set is completed, and the obtained data is recorded as the primary feature data.

[0079] After that, step S13 can be performed.

[0080] Step S13: data preprocessing is performed on the primary feature data to obtain secondary feature data.

[0081] In this embodiment, the preprocessing of the primary feature data can include data cleaning (including processing of missing values and abnormal values), data scaling (Z-Score standardization can be used), data testing (such as two-sample KS test), etc. to obtain secondary feature data. It should be noted that the data testing and the model evaluation in the following text are two independent processes. The two-sample KS test here is used to determine whether each row of data is on the same mathematical distribution to ensure the significance of subsequent training set and test set division.

[0082] The purpose of data preprocessing is to improve data quality, ensure the consistency and comparability of model input, and verify the statistical significance of data, thereby enhancing the performance and reliability of the machine learning model. Specifically, the data preprocessing methods used in this embodiment include data cleaning, data scaling, and data testing, as shown in Table 2 below:

[0083] Table 2 Data preprocessing content table

[0084]

[0085] Data cleaning successfully eliminates meaningless data in EIS samples through processing of missing values and abnormal values in EIS data, ensuring the effectiveness of the data.

[0086] The data scaling selects the Z-Score standardization method, realizes the unification of different dimensional features in the dimension, and optimizes the efficiency of subsequent model solving. At the same time, compared with other standardization and normalization methods, this method has strong universality, preserves the distance information between EIS samples, and can better handle abnormal maximum values in measurement.

[0087] The data inspection adopts double-sample KS inspection, verifies the distribution consistency of the EIS data training set and test set on any feature through traversal, and ensures the generalization ability of the model. Compared with other inspection methods, this method does not require the data to present a similar normal distribution, and can process EIS data sets with a large number of samples, and has stronger robustness.

[0088] After completing the data preprocessing, the secondary feature data can be obtained.

[0089] At this time, step S14 can be run.

[0090] Step S14: Based on the secondary feature data, an adaptive optimization of a plurality of alternative feature processing algorithms is performed to determine a target feature processing algorithm, so as to perform feature processing on the secondary feature data by using the target feature processing algorithm to obtain complete feature data.

[0091] In the embodiment, the secondary feature data can be extracted by using each alternative feature processing algorithm to determine the alternative feature data corresponding to each alternative feature processing algorithm. Here, it is equivalent that each alternative feature processing algorithm corresponds to a feature processing scheme, and the secondary feature data is processed by the alternative feature processing algorithm to obtain the corresponding alternative feature data.

[0092] The purpose of feature processing is to construct features with high information content through feature dimension reduction and feature dimension increase. For EIS features, all column features are composed of real and imaginary parts of electrochemical impedance spectroscopy at different frequencies. For such features composed of the same type of data, the correlation between the features is high, and there is no clear optimal dimension reduction or dimension increase algorithm.

[0093] Therefore, based on common feature processing algorithms, a feature processing algorithm library suitable for EIS data is constructed, and an adaptive feature optimization module is designed to realize the optimal feature processing customized for the current battery and automation. The feature processing standby algorithm library is shown in Table 3.

[0094] Table 3. Content table of feature processing standby algorithm library

[0095]

[0096] In order to realize the optimization of the alternative feature processing algorithm, the data set corresponding to each alternative feature processing algorithm needs to be constructed, and the theoretical experimental data set and the engineering application data set corresponding to each alternative feature processing algorithm are determined.

[0097] For a two-dimensional data matrix that can be used for model training, the general construction method of the data set is to randomly sort all rows and then divide the training set and the test set according to a certain proportion. However, for the EIS feature data of N×q rows and 2p+1 columns in the present scheme, the data is composed of N SOH inconsistent batteries of the same type. Therefore, although the number of label values is N×q, there are only N different numerical values. Therefore, randomly dividing the data set in the conventional way will lead to inconsistent distribution of the training set and the test set.

[0098] For example, the construction of the theoretical experimental data set and the engineering application data set is taken as an example.

[0099] The theoretical experimental data set construction method is that, for the alternative feature data corresponding to each alternative feature processing algorithm, all rows (i.e. N×q rows) of the alternative feature data can be randomly sorted, and the training and validation set and the test set can be divided according to the ratio of 8:2. Then, the training and validation set is divided into K parts, and K pairs of training and validation set combinations are obtained. The test set, the K pairs of training and validation set combinations are combined, and the corresponding theoretical experimental data set is obtained.

[0100] The engineering application data construction method is that, for the alternative feature data corresponding to each alternative feature processing algorithm, in the N×q rows of the alternative feature data, all q rows of data of 1 retired battery sample are taken as the test set, and in the remaining (N-1)×q rows, the q rows of each retired battery sample are grouped and divided into (N-2)×q rows of (N-2) training sets of retired battery samples and q rows of 1 retired battery sample of the validation set. Different retired battery samples are taken as the test set, and the operation is repeated N-1 times to obtain N groups of training sets, validation sets and test sets as the engineering application data set.

[0101] After the construction of the engineering application data set is completed, the target feature processing algorithm can be determined based on each alternative feature processing algorithm, the corresponding engineering application data set and each alternative modeling algorithm, using the adaptive feature optimization module. The adaptive feature optimization module uses the global sorting processing algorithm and the feature priority processing algorithm to integrate the alternative feature processing algorithm based on the idea of the greedy algorithm.

[0102] For example, the data construction in the feature processing algorithm dimension is completed using an alternative feature processing algorithm, the data construction in the modeling algorithm dimension is completed using an alternative modeling algorithm, and the data construction in the battery dimension is completed based on the idea of an engineering application data set and a traversal loop test. Here, there is a three-dimensional array with three coordinate axes x, y, and z, which are respectively the feature algorithm, the modeling algorithm, and the battery dimension. The value of each dimension is a prediction result value with the three as variables, such as x[feature algorithm 1, feature algorithm 2], y[modeling algorithm 1, modeling algorithm 2], and z[test battery 1, test battery 2], and the prediction data of F(x[0]y[0]z[0]) indicates the prediction value obtained by using the feature algorithm 1 + the modeling algorithm 1 and the battery 1 for testing.

[0103] Accordingly, a three-dimensional data including the three dimensions X, Y, and Z can be constructed, denoted as a three-dimensional array F1, and the three dimensions X, Y, and Z respectively refer to the feature processing algorithm dimension, the modeling algorithm dimension, and the battery dimension. F1(x, y, z) is a corresponding cell data list of the three-dimensional array F1, containing two floating-point value data of a test set accuracy mean value and a test set accuracy minimum value, represented by F1(x, y, z)[0] and F1(x, y, z)[1] respectively.

[0104] Then, in the case of z = i, all the values of x and y are traversed to determine whether the minimum accuracy F1(x, y, i)[1] with x and y as independent variables is greater than a threshold value (for example, the threshold value threshold1 = 98%), and if the condition is met, the average accuracy F1(x, y, i)[0] is recorded. All the values of the z-axis are traversed, and the corresponding operation is repeated. The recorded F1(x, y, i)[0] is sequentially filled into G1(x, y), and the two-dimensional conversion of the three-dimensional array is completed, and finally the two-dimensional array G1 representing the prediction effect of the algorithm combination is obtained. G1(x, y) is a corresponding cell data list of the two-dimensional array G1, which is composed of F1(x, y, i)[0] determined when z meets the condition at each value.

[0105] The cell data list of G1(x, y) is averaged to obtain the average accuracy G1_acc(x, y) of the current algorithm combination on the engineering application data set, and all x and y are traversed to obtain G1_acc. The data list length of G1(x, y) is recorded to obtain the number G1_num(x, y) of the current algorithm combination that meets the standard on the engineering application data set, and all x and y are traversed to obtain G1_num.

[0106] Afterwards, based on the priority order of G1_num and G1_acc, the first m1 algorithm combinations are screened out, and the corresponding feature processing algorithm x_list1 is recorded as the output result of the three-dimensional global sorting algorithm.

[0107] For the G1_num two-dimensional array, for the current x, the value of y is traversed, and the following is calculated:

[0108] (1)

[0109] The prediction result array List(x) is sorted from large to small, and the first m2 largest value feature processing algorithm x_list2 is obtained, which is the output result of the feature priority sorting algorithm. The highest average precision in the intersection of x_list1 and x_list2 sets is determined. If the intersection is empty, the highest average precision in x_list1 and x_list2 is taken as the output of the target feature processing algorithm determined by the adaptive feature optimization module.

[0110] Accordingly, the determination of the target feature processing algorithm can be completed.

[0111] After the target feature processing algorithm is determined, step S15 can be further run.

[0112] Step S15: Based on the complete feature data, the optimization of multiple candidate modeling algorithms is performed to determine the target modeling algorithm. The target modeling algorithm is used for modeling, and the complete feature data is used for model training and testing to construct a retired power battery SOH prediction model.

[0113] In this embodiment, since the target feature processing algorithm has been determined, the theoretical experimental data set and the engineering application data set corresponding to the target feature processing algorithm can be used as the complete feature data, Figure 2 The modeling and prediction stage in the above formula (2) shows the data set construction process, which essentially can use the theoretical experimental data set and the engineering application data set constructed in the process of determining the target feature processing algorithm. Of course, the theoretical experimental data set and the engineering application data set corresponding to the target feature processing algorithm can also be constructed separately after the target feature processing algorithm is determined, which is not limited here. The theoretical experimental data set and the engineering application data set corresponding to the target feature processing algorithm are determined, which are used as the complete feature data for the subsequent process of determining the target modeling algorithm.

[0114] For the case where both the independent variable and the dependent variable are continuous numerical data, the regression models commonly used for modeling include linear regression, polynomial regression, and ridge regression, etc. In addition, after selecting the above regression model matching the data type, the model library further incorporates the current mainstream and excellent algorithms to enrich the algorithm system. The main algorithms are shown in the following Table 4.

[0115] Table 4. Modeling prediction of the model library content table

[0116]

[0117]

[0118] Then, after obtaining the complete feature data, the target modeling algorithm and the corresponding parameters can be determined based on the target feature processing algorithm, the complete feature data, each candidate modeling algorithm, the algorithm parameter space, and the parameter optimization method, using an algorithm combination optimization module. In the hyperparameter optimization process of the modeling algorithm, grid search or random search is selected for parameter optimization according to the number of parameters.

[0119] Specifically, different algorithm combinations can be formed based on the matching of the target feature processing algorithm and each candidate modeling algorithm, to complete data construction in the algorithm combination dimension. According to each algorithm combination, a corresponding parameter search space is generated, and then a parameter optimization method is selected according to the number of parameters in the parameter search space, to complete data construction in the algorithm parameter dimension. In addition, data construction in the battery dimension can also be completed using an engineering application data construction method according to the number of retired battery samples, to form a three-dimensional array F2 containing the U, V, and Z dimensions. The three dimensions U, V, and Z respectively represent the algorithm combination dimension, the algorithm parameter dimension, and the battery dimension. F2(u, v, z) is the model prediction result based on the current u, v, and z, F2(u, v, z)[0] is the test set accuracy mean, F2(u, v, z)[1] is the test set accuracy minimum, and F2(u, v, z)[2] is the R 2 Determination coefficient.

[0120] After that, similar to the adaptive feature optimization module, the algorithm combination optimization module can realize two-dimensional conversion of three-dimensional data by judging and compressing the z-axis data, to obtain G2_num, G2_acc, and G2_R 2Three two-dimensional tables. For example, in the case of z = i, traverse all the values of u, v, judge whether the minimum accuracy of F2(u, v, i) [1] with u, v as the independent variable is greater than the threshold value, if the condition is met, record the average accuracy of F2(u, v, i) [0], traverse all the values of the z axis, repeat the corresponding operation, and fill the recorded F2(u, v, i) [0] into G2(u, v) in turn, that is, complete the two-dimensional conversion of the three-dimensional array, and finally obtain the two-dimensional array G2 for representing the prediction effect of the algorithm combination, G2(u, v) is the corresponding cell data list of the two-dimensional array G2, which is composed of F2(u, v, i) [0] determined when z meets the condition at each value; take the average of the cell data list of G2(u, v) to obtain the average accuracy G2_acc(u, v) of the current algorithm combination on the engineering application data set, traverse all u, v, that is, obtain G2_acc, and record the data list length of G2(u, v) to obtain the number of standards G2_num(u, v) of the current algorithm combination on the engineering application data set, and traverse all u, v to obtain G2_num.

[0121] Note that three tables are obtained here, and one more G2_R is obtained compared with the adaptive feature optimization module 2 Here G2_R 2 The processing method of the table is consistent with the above two tables, that is, one more G2_R 2 Because the overall flow of the algorithm is later, more data is needed to evaluate the accuracy of the model.

[0122] num represents the number of batteries that meet the 98% high accuracy constraint on the z axis with u, v as variables. For example, when G2_num(u = 0, v = 0) = 12, it means that in the case where u and v take the first value, 15 batteries all output prediction accuracy, and 12 of them meet the 98% constraint. This index can prove the robustness and generalization of the model to a large extent.

[0123] acc represents the average value of the accuracy on the z axis with u, v as variables. G2_acc and G2_num are obtained as follows: acc only retains data with a value greater than 98%, and num directly obtains the length of the valid data of acc. Under the selected conditions of u and v, G2_acc = [0.985, 0.99] after threshold screening, and G2_num = 2. (In this example, z has only 3 values).

[0124] R 2 represents the average value of the "determination coefficient" on the z axis with u, v as variables. R 2For regression prediction tasks in machine learning, a classic indicator is used to evaluate the goodness of a model.

[0125] And for the converted two-dimensional array, based on G2_num, G2_acc, G2_R 2 The priority ranking of the optimal algorithm combination and its corresponding optimal parameter list is calculated. If there are multiple maximum values of G2_num, further selection judgment can be made (described later, the direct embodiment of this selection judgment process is mainly in formula 7 below), and the target modeling algorithm and the corresponding parameters are determined as the output of the algorithm combination optimization module. That is, if G2_num has only one maximum value, the algorithm combination composed of u and v corresponding to the maximum value is adopted. If G2_num has multiple maximum values (most cases), formula (7) is used for judgment and selection. If G2_num has no maximum value, parameter adjustment and re-optimization are performed according to the adjustment parameter mode in formula (7) below.

[0126] Thus, the target feature processing algorithm, target modeling algorithm and corresponding parameters can be determined as the basis for constructing the prediction model. Then, using the EIS test data of N retired battery samples, after data preprocessing, feature extraction using the target feature processing algorithm, data construction, training set and test set division, etc., the training and test data are formed, and the combined model is trained, and finally the trained retired power battery SOH prediction model is obtained.

[0127] In order to realize the evaluation and optimization of the model, in the evaluation and verification stage, in terms of evaluation and verification methods, the verification method of "engineering application data set + leave-one-out cross-validation" as the main method, and "theoretical experimental data set + K-fold cross-validation" as the auxiliary method is adopted, as shown below.

[0128] 1. Theoretical experimental data set evaluation and verification

[0129] K-fold cross-validation refers to dividing the training and validation set excluding the test set into K non-overlapping subsets, and using K-1 subsets for model training and the remaining 1 subset for model validation.

[0130] For the K sets of data sets obtained by the theoretical experimental data construction method, the modeling effect of EIS features is evaluated and verified using the K-fold cross-validation method, which effectively reduces the differences introduced by different data division and the risk of model overfitting to training data.

[0131] Although the robustness of the model and the reliability of the evaluation results are theoretically improved, the special nature of the EIS dataset label value, i.e., the same label value for the same battery, causes the data distribution to be inconsistent when the data set is directly divided, thereby making it impossible for the model to establish a complete mapping relationship of the current data set due to the lack of feature data of some batteries during the training process.

[0132] 2. Engineering application dataset evaluation and verification

[0133] Leave-one-out cross-validation refers to a method in which each sample in the training and validation set except the test set is individually used as a subset, and all other samples except the current sample are used for model training, and then the current sample is used for model verification.

[0134] The specific method of using leave-one-out cross-validation on the engineering application dataset is as follows: taking the battery as the unit, in N batteries, the q-row data of 1 battery is the test set, the q-row data of another battery is the validation set, and the remaining (N-2) x q data is the training set. Repeat the processing cycle test until all batteries are used as the test set, and then the prediction effect of such batteries on different battery (SOH) distributions can be comprehensively verified.

[0135] In order to achieve optimization, the following two aspects are improved in this embodiment: the model optimization index includes the loss function used to optimize the fitting effect of the model training, and the evaluation index used to measure the training effect of the model.

[0136] The loss function quantifies the gap between the predicted value and the true value of the model through its continuous and derivable characteristics, guides the optimization algorithm to adjust the model parameters by calculating the gradient, minimizes the error, and thus improves the fitting effect of the model. Accordingly, the loss function in this embodiment is designed as:

[0137] , (2)

[0138] wherein the weight hyperparameter is 0.6.

[0139] The loss function uses a weighted combination of MSE and HuberLoss, which amplifies small errors through the square error mechanism of MSE to accelerate convergence, and uses the linear growth characteristics of Huber loss to weaken the influence of outliers, realizes flexible processing of complex error distribution, and balances high sensitivity and outlier robustness, balancing accurate fitting and optimization stability. The formulas of MSE and HuberLoss are as follows:

[0140] , (3)

[0141] , (4)

[0142] Among them, hyperparameters We set it to 1 because the data has already been standardized using Z-Score. Indicates the first The true value of each sample Indicates the first The model prediction value for each sample. Indicates the number of samples. .

[0143] The evaluation metric primarily uses a user-defined AccRatio, supplemented by the coefficient of determination R². The former measures the relative deviation between predicted and actual values; a smaller deviation results in a higher value. The latter, a classic evaluation metric for regression tasks, reflects the proportion of the total variance in the dependent variable that can be explained by the independent variables through the regression relationship; a higher value indicates a better model fit. By measuring the proportion of the total variance in the target variable explained by the model, R²... 2 This allows for an intuitive evaluation of the goodness of fit of the regression model. Therefore, in this embodiment, the evaluation index is designed as follows:

[0144] (5)

[0145] (6)

[0146] in, This indicates the accuracy of the predicted value; for example, if the actual SOH of a battery is 0.99 and the predicted value is 0.98, then the AccRatio on a sample is 1 - (0.99 - 0.98) / 0.99 = 98.99%.

[0147] Given a set M = {1, 2, 3, ..., m}, M1 = {i ∈ M | Acc i >λ1}, M2={i∈M|R 2 i >λ2}. The formula for selecting the optimal model i* is as follows:

[0148] (7)

[0149] Among them, the hyperparameters λ1 are set to 0.98 and λ2 are set to 0.6. When the optimal model is empty, λ2 is lowered or the data is re-selected and processed to train the model.

[0150] Set M represents a total of m data combinations, numbered sequentially from the first to the mth, each with a corresponding feature algorithm, modeling algorithm, hyperparameters, etc.; set M1 represents the data combinations in which data with a precision exceeding λ1 are retained, and those with a precision below λ1 are deleted; set M2 is similar to M1, except that the object of judgment is R. 2Coefficient of determination; if M1 is not empty set, it means there are still many data satisfying the threshold constraint of λ1 (e.g. reaching 98%), at this time, in order to ensure the model generalization and robustness, find R 2 The highest one is taken as the optimal combination of the model; if M1 is empty set, it means that all cases cannot reach the threshold constraint. At this time, in the corresponding R 2 Find the highest Acc in the data that can reach the requirement of λ2 (M2 set), and take it as the optimal combination of the model. If both are empty sets, appropriately reduce the constraint of λ2 and reprocess the training data. Among them, Argmax is a fixed usage, which means finding the data corresponding to the maximum value in a set.

[0151] After obtaining the trained retired power battery SOH prediction model, the retired power battery SOH prediction method can be run. Please refer to Figure 5 The retired power battery SOH prediction method can include steps S21, S22 and S23.

[0152] When it is necessary to predict the SOH of the retired power battery, first, step S21 can be run.

[0153] Step S21: Obtain the EIS test data of the to-be-tested retired power battery.

[0154] In this embodiment, the EIS test data of the to-be-tested retired power battery is obtained without going through cumbersome charge and discharge operations, but the EIS test data can be tested at any SOC.

[0155] After obtaining the EIS test data of the to-be-tested retired power battery, step S22 can be run.

[0156] Step S22: Perform data processing and feature extraction on the EIS test data to obtain input features.

[0157] In this embodiment, the EIS test data can be processed, such as one-dimensional feature processing, data cleaning, data scaling, etc., and then the target feature processing algorithm is used for feature extraction to obtain the input features.

[0158] Then step S23 can be run.

[0159] Step S23: Input the input features into the constructed retired power battery SOH prediction model, and predict the maximum capacity value of the to-be-tested retired power battery through the retired power battery SOH prediction model to determine the SOH of the to-be-tested retired power battery.

[0160] In this embodiment, the input features can be input into the constructed retired power battery SOH prediction model, the maximum capacity value of the to-be-tested retired power battery is predicted through the retired power battery SOH prediction model, and then the SOH of the to-be-tested retired power battery is calculated in combination with the rated capacity value of the to-be-tested retired power battery. Thus, the SOH prediction of the to-be-tested retired power battery is completed.

[0161] Part of the experimental data is provided here:

[0162] (1) Dataset

[0163] The experimental part will perform prediction analysis on four types of batteries of different specifications and models, i.e., lithium iron phosphate 12.5 Ah, ternary lithium 5.8 Ah, lithium iron phosphate 20 Ah, and lithium iron phosphate 22 Ah. The basic information of the four types of batteries is shown in Table 5 as follows:

[0164] Table 5. Basic information of battery data

[0165]

[0166] (2) Data measurement

[0167] Label value SOH measurement:

[0168] The complete charge and discharge measurement of all the above batteries is performed using a charge and discharge machine, and the SOH value of each battery is calculated. The maximum capacity Capacity (mah) and battery health state SOH distribution of each type of battery are shown in Table 6 as follows.

[0169] Table 6. Residual capacity and SOH information table of battery data

[0170]

[0171] Feature value EIS measurement:

[0172] Using the EIS feature data measurement method in this embodiment, the final Excel table data amount of all the above batteries is shown in Table 7 as follows.

[0173] Table 7. Battery EIS data information table

[0174]

[0175] (3) Modeling prediction

[0176] Optimal prediction result:

[0177] The EIS characteristic values and SOH label values of the four types of batteries obtained by the above measurement are sequentially processed, modeled, predicted, and evaluated using the method of the present application. The optimal prediction accuracy results of each type of battery are shown in the following table. The prediction accuracy value is calculated using the Acc index in formula 5.

[0178] Table 8. Optimal prediction accuracy and model parameter selection table for batteries

[0179]

[0180] The optimal algorithm combination corresponding to the optimal prediction accuracy is mainly composed of feature algorithms and modeling algorithms. Among them, the feature parameter variable only takes the dimension number dimension after feature transformation, and the modeling parameter sets its parameter space for each modeling algorithm, such as the alpha, tol, and max_iter of lasso regression, which are the regularization strength, convergence tolerance, and maximum iteration number, respectively.

[0181] Prediction result comparison:

[0182] The prediction results of each type of battery under different algorithm combinations are compared as follows. The unit in the table is %, and data with accuracy less than 98% is not recorded.

[0183] Battery1: Lithium Iron Phosphate_12.5Ah:

[0184] Table 9. Battery1_Lithium Iron Phosphate_12.5Ah prediction result comparison table

[0185]

[0186]

[0187] Battery2: Ternary Lithium_5.8Ah

[0188] Table 10. Battery2_Ternary Lithium_5.8Ah prediction result comparison table

[0189]

[0190]

[0191] Battery3: Lithium Iron Phosphate_20Ah

[0192] Table 11. Battery3_Lithium Iron Phosphate_20Ah prediction result comparison table

[0193]

[0194]

[0195] Battery4: Lithium Iron Phosphate_22Ah

[0196] Table 12. Comparison of Prediction Results for Battery4_Lithium Iron Phosphate_22Ah

[0197]

[0198]

[0199] In Tables 10-12, the missing data points were those that did not meet the threshold accuracy requirements and were therefore omitted. The experimental results above show that, overall, for EIS features, the LDA feature algorithm has good applicability to various types of batteries, and there is no obvious bias in the modeling algorithm.

[0200] Through the aforementioned iterative tests, a comprehensive accuracy assessment of the current battery category under different SOH distributions was achieved. Furthermore, for any SOH label value, high-precision prediction can be achieved based on the EIS features under any SOC state from 0% to 100%. This significantly increases the model's robustness and lowers the threshold requirements for input data.

[0201] In summary, this application provides a method for predicting the State of Health (SOH) of retired power batteries. Based on the characteristics of EIS data, an engineering application data construction method is used to divide the training, validation, and test sets, ensuring that the model can learn complete data features from 0% to 100% SOC without bias during training. Furthermore, a loop-based leave-one-out cross-validation test ensures that the model achieves the required accuracy even when using any battery as the test set, demonstrating the model's high robustness and strong generalization ability. Through these improvements, high-precision SOH prediction of EIS features under any SOC is achieved, filling a practical gap in the operational feasibility of SOH prediction based on EIS measurement data. In this scheme, a richly layered and clearly defined adaptive processing flow is constructed using different stage divisions. Feature algorithms, modeling algorithms, evaluation methods, and parameter selection are treated as overall variables, and multi-variable settings approximate the physical maximum prediction accuracy of the battery data, approaching the global optimum. Simultaneously, an adaptive optimization module based on a greedy algorithm effectively solves the curse of dimensionality and reduces time complexity, providing a customized high-precision analysis and prediction solution for battery data. The algorithm of this scheme achieves an average prediction accuracy of 99.11% (±0.16%) for EIS data under various SOC states of four types of batteries, and the average deviation of remaining capacity diagnosis is less than 1%.

[0202] For the two-dimensional data matrix available for model training, the general construction method of the data set is to randomly sort all rows and then divide the training set and the test set according to a certain proportion. However, for the EIS characteristic data of N x q rows and 2p+1 columns in the present scheme, the data is composed of N SOH inconsistent batteries of the same type. Therefore, although the number of label values is N x q, there are only N different numerical values. Therefore, randomly dividing the data set according to the conventional method will lead to inconsistent distribution of the training set and the test set, and thus the model can only learn the EIS data characteristics at a partial SOC. In order to learn the EIS characteristics at any SOC and realize SOH prediction, on the basis of the conventional data set construction method, according to the characteristic that multiple features of EIS data correspond to one label, an engineering application data construction method is proposed, which ensures that the model can learn the data characteristics at the full SOC distribution in a balanced and unbiased manner during data division. At the same time, in order to adapt to the present construction method during model validation, a cyclic traversal test method based on leave-one-out cross-validation is proposed, which greatly guarantees the robustness of the model through strict verification on N data sets.

[0203] In order to improve the prediction accuracy of the model as much as possible and make the prediction ability of the model approach the physical upper limit of the data, the idea of constructing an optimal solution space with multiple variables is adopted. The feature algorithm, modeling algorithm, evaluation and verification method and various parameter selection are all regarded as variables that need to be considered in the present scheme. Through the setting of a large number of variables, it is ensured that the data distribution formed by various variables can approach the physical maximum prediction accuracy of the current battery data as much as possible, achieving the effect of approximate global optimal solution. On the basis of the above, in order to solve the problems of dimension disaster, time complexity and the like caused by multiple variables, a dimension reduction optimization method based on the greedy algorithm is proposed, and adaptive feature optimization modules and optimal algorithm combination optimization modules are constructed based on different application data, realizing adaptive optimization selection of key algorithms and parameters of the model under low time complexity.

[0204] The above only describes the embodiments of the present application and does not limit the protection scope of the present application. For those skilled in the art, the present application can have various changes and modifications. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.

Claims

1. A method for predicting the state of harm (SOH) of retired power batteries, characterized in that, include: Obtain EIS test data of the retired power battery to be tested; Data processing and feature extraction are performed on the EIS test data to obtain the input features; The input features are fed into the constructed SOH prediction model for retired power batteries. The maximum capacity value of the retired power battery under test is predicted by the retired power battery SOH prediction model, so as to determine the SOH of the retired power battery under test. The construction process of the SOH prediction model for retired power batteries is as follows: Obtain an EIS dataset consisting of EIS data from N retired battery samples. The EIS dataset contains N sub-tables, each with M rows and L columns. M is the product of the number of samples p and the number of tests q, and L is 4, representing the frequency, the opposite of the real part, the imaginary part, and the current charge of the retired battery sample under test conditions, respectively. One-dimensional feature alignment is performed on the N sub-tables in the EIS dataset to obtain primary feature data, which is the total table data containing N×q rows and 2p+1 columns. The primary feature data is preprocessed to obtain secondary feature data; Based on secondary feature data, multiple alternative feature processing algorithms are adaptively optimized to determine the target feature processing algorithm, which is then used to process the secondary feature data to obtain complete feature data. Based on complete feature data, multiple alternative modeling algorithms are optimized to determine the target modeling algorithm. The target modeling algorithm is then used for modeling, and the complete feature data is used for model training and testing to construct a SOH prediction model for retired power batteries. Based on secondary feature data, adaptive optimization is performed on multiple candidate feature processing algorithms to determine the target feature processing algorithm, including: Each alternative feature processing algorithm is used to extract features from the secondary feature data to determine the alternative feature data corresponding to each alternative feature processing algorithm. For each candidate feature processing algorithm, a dataset is constructed to determine the theoretical experimental dataset and engineering application dataset corresponding to each candidate feature processing algorithm. Based on each candidate feature processing algorithm, the corresponding engineering application dataset, and each candidate modeling algorithm, the target feature processing algorithm is determined using an adaptive feature optimization module. The adaptive feature optimization module is based on the idea of ​​a greedy algorithm and integrates a global sorting algorithm and a feature priority processing algorithm to process the candidate feature processing algorithms. Based on complete feature data, multiple candidate modeling algorithms were optimized to determine the target modeling algorithm, including: The theoretical experimental dataset and engineering application dataset corresponding to the target feature processing algorithm are used as complete feature data; Based on the target feature processing algorithm, complete feature data, each alternative modeling algorithm, algorithm parameter space, and parameter optimization method, the target modeling algorithm and its corresponding parameters are determined by the algorithm combination optimization module. In the hyperparameter optimization process within the modeling algorithm, the algorithm combination optimization module will adaptively select grid search or random search for parameter optimization based on the number of parameters.

2. The method for predicting the state of harm (SOH) of decommissioned power batteries according to claim 1, characterized in that, One-dimensional feature alignment is performed on N sub-tables in the EIS dataset to obtain primary feature data, including: Iterate through the N sub-tables in the EIS dataset, and for each sub-table: Extract the parameters of the second and third columns in units of p rows, process them into a 1×2p matrix, and determine the q×2p matrix corresponding to the current sub-table. Also, find the maximum value of the fourth column in the current sub-table, construct a q×1 matrix as the label value and add it to the rightmost side of the q×2p matrix corresponding to the current sub-table, and obtain a q×(2p+1) matrix as the feature data corresponding to the current sub-table. The feature data corresponding to the N sub-tables are concatenated to obtain a (N×q)×(2p+1) matrix, which serves as the primary feature data corresponding to the EIS dataset.

3. The SOH prediction method for decommissioned power batteries according to claim 1, characterized in that, Data preprocessing is performed on the primary feature data to obtain secondary feature data, including: Data cleaning is performed on the primary feature data, including handling missing and outlier values; The primary feature data after data cleaning is scaled, and the data scaling method adopts Z-Score normalization. The primary feature data of the scaled data is tested to obtain secondary feature data. The data testing method is a two-sample KS test.

4. The SOH prediction method for decommissioned power batteries according to claim 1, characterized in that, For each candidate feature processing algorithm, a dataset is constructed to identify the theoretical experimental dataset and engineering application dataset for each algorithm, including: For each candidate feature processing algorithm, the candidate feature data is randomly sorted and divided into training and validation sets and test sets in an 8:2 ratio. The training and validation sets are then divided into K-fold partitions to obtain K pairs of training and validation set combinations. The test set and the K pairs of training and validation set combinations are combined to obtain the corresponding theoretical experimental dataset. For each alternative feature processing algorithm, the alternative feature data is as follows: In the N×q rows of the alternative feature data, all q rows of data of one retired battery sample are taken as the test set. In the remaining (N-1)×q rows, the data is grouped into (N-2) ×q rows of retired battery samples as the unit, and split into a training set of (N-2) retired battery samples and a validation set of q rows of one retired battery sample. Different retired battery samples are taken as the test set and the operation is repeated N-1 times to obtain N sets of training sets, validation sets and test sets, which are used as the engineering application dataset.

5. The method for predicting the state of harm (SOH) of decommissioned power batteries according to claim 1, characterized in that, Based on each candidate feature processing algorithm, the corresponding engineering application dataset, and each candidate modeling algorithm, the target feature processing algorithm is determined using an adaptive feature optimization module, including: The alternative feature processing algorithm is used to complete the data construction in the feature processing algorithm dimension, the alternative modeling algorithm is used to complete the data construction in the modeling algorithm dimension, and the battery dimension data is constructed based on the engineering application dataset and the idea of ​​traversal loop testing, forming a three-dimensional data containing three dimensions X, Y, and Z, denoted as a three-dimensional array F1. The three dimensions X, Y, and Z respectively refer to the feature processing algorithm dimension, the modeling algorithm dimension, and the battery dimension. Among them, F1(x,y,z) is the data list of the corresponding cell of the three-dimensional array F1, containing two floating-point numerical data: the mean of the test set precision and the minimum of the test set precision, which are represented by F1(x,y,z)[0] and F1(x,y,z)[1] respectively. When z=i, iterate through all x and y values ​​and determine whether the minimum precision of F1(x,y,i)[1] with x and y as independent variables is greater than the threshold. If the condition is met, record the average precision of F1(x,y,i)[0]. Iterate through all values ​​in the z-axis direction and repeat the corresponding operation. Fill the recorded F1(x,y,i)[0] into G1(x,y) in sequence. The two-dimensional transformation of the three-dimensional array can be completed. Finally, the two-dimensional array G1 is obtained to represent the prediction effect of the algorithm combination. G1(x,y) is the data list of the corresponding cells of the two-dimensional array G1, which is composed of F1(x,y,i)[0] determined when z meets the condition at each value. Calculate the mean of the cell data list of G1(x,y) to obtain G1_acc(x,y), which represents the average accuracy of the current algorithm combination on the engineering application dataset. Iterate through all x and y to obtain G1_acc. Record the length of the data list of G1(x,y) to obtain G1_num(x,y), which represents the number of times the current algorithm combination meets the standard on the engineering application dataset. Iterate through all x and y to obtain G1_num. Based on the priority order of G1_num and G1_acc, the top m1 algorithm combinations are selected, and their corresponding feature processing algorithm x_list1 is recorded as the output result of the three-dimensional global sorting algorithm. Given the two-dimensional array G1_num, iterate through all values ​​of y and calculate... The feature processing algorithm x_list2, which yields the first m2 largest values, is the output of the feature priority sorting algorithm. The algorithm identifies the feature with the highest mean precision in the intersection of x_list1 and x_list2. If the intersection is empty, the feature with the highest mean precision in x_list1 and x_list2 is selected as the output of the target feature processing algorithm determined by the adaptive feature optimization module.

6. The SOH prediction method for decommissioned power batteries according to claim 1, characterized in that, Based on the target feature processing algorithm, complete feature data, each alternative modeling algorithm, algorithm parameter space, and parameter optimization method, the target modeling algorithm and its corresponding parameters are determined using the algorithm combination optimization module, including: The theoretical experimental dataset and engineering application dataset corresponding to the target feature processing algorithm are used as complete feature data; Based on the combination of target feature processing algorithm and each alternative modeling algorithm, different algorithm combinations are formed to complete the data construction in the algorithm combination dimension. According to each algorithm combination, the corresponding parameter search space is generated. Then, according to the number of parameters in the parameter search space, the parameter optimization method is adaptively selected to complete the data construction in the algorithm parameter dimension. In addition, according to the number of retired battery samples, the data construction in the battery dimension is completed using the engineering application data construction method. A three-dimensional array F2 containing three dimensions U, V, and Z is formed. The three dimensions U, V, and Z refer to the algorithm combination dimension, algorithm parameter dimension, and battery dimension, respectively. Among them, F2(u,v,z) is the model prediction result based on the current u,v,z, F2(u,v,z)[0] is the mean accuracy of the test set, F2(u,v,z)[1] is the minimum accuracy of the test set, and F2(u,v,z)[2] is the R 2 Coefficient of determination; By judging and compressing the z-axis data, the two-dimensional conversion of the three-dimensional data is achieved, resulting in G2_num, G2_acc, and G2_R. 2 Three two-dimensional tables, from which G2_num, G2_acc, and G2_R are obtained. 2 The three two-dimensional tables are as follows: When z=i, iterate through all the values ​​of u and v, and determine whether the minimum precision of F2(u,v,i)[1] with u and v as independent variables is greater than the threshold. If the condition is met, record the average precision of F2(u,v,i)[0]. Iterate through all the values ​​in the z-axis direction and repeat the corresponding operation. Fill the recorded F2(u,v,i)[0] into G2(u,v) in turn. This completes the two-dimensional transformation of the three-dimensional array and finally obtains the two-dimensional array G2 used to represent the prediction effect of the algorithm combination. G2(u,v) is the two-dimensional array G2. The corresponding cell data list consists of F2(u,v,i)[0] determined when z satisfies the condition for each value; the average value of the cell data list of G2(u,v) is calculated to obtain G2_acc(u,v) which represents the average accuracy of the current algorithm combination on the engineering application dataset; all u and v are traversed to obtain G2_acc; the length of the data list of G2(u,v) is recorded to obtain G2_num(u,v) which represents the number of times the current algorithm combination meets the standard on the engineering application dataset; and all u and v are traversed to obtain G2_num and G2_R. 2 The table processing method is consistent with that of the G2_num and G2_acc tables; For the transformed two-dimensional array, based on G2_num, G2_acc, and G2_R 2 The algorithm is prioritized and the optimal combination of algorithms and its corresponding optimal parameter list are calculated. Based on this, the target modeling algorithm and its corresponding parameters are determined as the output of the algorithm combination optimization module.

7. The SOH prediction method for decommissioned power batteries according to claim 6, characterized in that, Model optimization metrics include the loss function used to optimize the model's training fit, and evaluation metrics used to measure the model's training performance. The loss function is designed as follows: ， Among them, the weight hyperparameter Take 0.6; ， ， Among them, hyperparameters Take 1, Indicates the first The true value of each sample Indicates the first The model prediction value for each sample. Indicates the number of samples. ; The evaluation indicators are designed as follows: ， ， in, Indicates the accuracy of the predicted value; Given a set M = {1, 2, 3, ..., m}, M1 = {i ∈ M | Acc i >λ1}, M2={i∈M|R 2 i >λ2}; The formula for selecting the optimal model i* is as follows: ， Among them, the hyperparameters λ1 are set to 0.98 and λ2 are set to 0.

6. When the optimal model is empty, λ2 is lowered or the data is re-selected and processed to train the model.

Images