A performance detection method for retired lithium battery based on multi-information fusion

Abstract

The present application relates to a kind of performance detection methods of retired lithium battery based on multi-information fusion, belong to machine learning model field.The method of the present application is as follows: first, obtain retired battery electrochemical impedance spectrum data and carry out multidimensional analysis, respectively for the impedance real part and imaginary part of each frequency point, equivalent circuit fitting parameter and relaxation time distribution characteristics, independent dimension reduction screening is carried out using mutual information method, and multi-source feature set is constructed;Then, the multi-source feature set is respectively input into the corresponding base learning ware and independently mapped to obtain the preliminary estimate of each base model;Finally, the above preliminary estimate is spliced to construct secondary feature matrix, and input into bayesian ridge regression model for fusion, and output the probability distribution containing prediction mean and standard deviation.The present application improves the accuracy of SOH point estimation while realizing high reliability uncertainty quantification, and provides quantitative risk assessment basis for retired battery gradient utilization classification.

Description

Technical Field

[0001] This invention belongs to the field of battery energy storage technology, specifically relating to a method for testing the performance of retired lithium batteries based on multi-information fusion. Background Technology

[0002] Electrochemical impedance spectroscopy (EIS) is a non-destructive detection technique that, by applying a small AC voltage and measuring the battery's impedance response, can obtain rich information about the internal electrochemical processes of the battery, such as charge transfer resistance, diffusion impedance, and interfacial reactions. EIS data, characterized by high dimensionality and nonlinearity, can reflect the microscopic mechanisms involved in battery aging and is an important data source for assessing battery state of health (SOH).

[0003] However, existing multi-source fusion estimation methods still have the following drawbacks: On the one hand, retired batteries suffer from severe impedance spectrum distortion due to complex aging mechanisms, making it easy for them to get trapped in local optima or even fail to converge when directly fitting the fractional-order ECM across the entire frequency band. Simultaneously, existing methods often directly and brute-force concatenate multi-source physical features, lacking systematic and independent screening based on correlation with SOH, resulting in severe redundancy and noise interference in the input features. On the other hand, and more importantly, most existing SOH estimation models only output a single deterministic point estimate, failing to provide uncertainty quantification for the prediction results. The tiered utilization of retired batteries has extremely low requirements for safety sorting tolerance; predictions lacking uncertainty assessment (i.e., confidence intervals) will directly lead to sorting failures or even safety hazards if deviations occur. Therefore, a high-precision method for estimating the health status of retired batteries with uncertainty quantification capabilities is urgently needed. Summary of the Invention

[0004] In view of this, the present invention addresses the problem that existing models in estimating the state of health (SOH) of retired lithium batteries can only output deterministic point estimates and cannot provide quantitative assessments of uncertainty. It provides a performance testing method for retired lithium batteries based on multi-information fusion, aiming to improve the accuracy, robustness, and safety of engineering decisions in SOH estimation.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] In a first aspect, the present invention provides a method for performance testing of retired lithium batteries based on multi-information fusion, which includes the following steps:

[0007] S1. Obtain a training set consisting of electrochemical impedance spectroscopy (EIS) and state of health (SOH) labels of retired lithium batteries, and extract electrochemical impedance spectroscopy screening features from the EIS. ), equivalent circuit fitting parameter characteristics ( ) and relaxation time distribution characteristics ( The three types of features are combined with the health status (SOH) label to form three training subsets. The base learners are trained independently using the three training subsets to obtain three primary prediction models.

[0008] S2. Using a K-fold cross-validation strategy, three primary prediction models generate three out-of-fold prediction values ​​for each training sample in the training set, and these values ​​are concatenated to construct the secondary features of that training sample. Secondary features of all training samples in the training set After standardization, the samples are combined with their respective health status (SOH) labels to form new samples, which are then used to train the machine learning regression model to obtain a trained secondary fusion model.

[0009] S3. The three trained primary prediction models and the secondary fusion model are cascaded and combined to estimate the health status of the retired lithium battery under test.

[0010] As a preferred embodiment of the first aspect above, the electrochemical impedance spectroscopy screening features ( The extraction method is as follows: the real and imaginary parts of the impedance at each frequency point in the electrochemical impedance spectroscopy (EIS) are used as candidate features. The mutual information between each candidate feature and the health status (SOH) label is calculated. Based on the mutual information, the multiple candidate features with the highest correlation are selected as electrochemical impedance spectroscopy screening features.

[0011] As a preferred embodiment of the first aspect above, the equivalent circuit fitting parameter characteristics ( The extraction method is as follows: model the retired lithium battery as a component of inductance. Ohm resistance A second-order fractional equivalent circuit model is formed by connecting a first-order fractional RC parallel unit, a second-order fractional RC parallel unit, and a constant-phase element in series. The equivalent circuit parameters are fitted using electrochemical impedance spectroscopy (EIS). The mutual information between each equivalent circuit parameter and the state of health (SOH) label is calculated. Based on the mutual information, the most correlated equivalent circuit parameters are selected as the equivalent circuit fitting parameter features.

[0012] As a preferred embodiment of the first aspect above, when fitting the equivalent circuit parameters using electrochemical impedance spectroscopy (EIS), a two-stage fitting strategy is employed, wherein:

[0013] In the first stage, the full-band electrochemical impedance spectroscopy (EIS) data is divided into three subsets: high frequency, mid frequency, and low frequency. The local equivalent circuit topology of each of the three frequency bands in the second-order fractional equivalent circuit model is then preliminarily fitted piecewise.

[0014] In the second stage, all the equivalent circuit parameter values ​​obtained from the initial fitting in the first stage are used as the starting point for global fitting. The second-order fractional equivalent circuit model is then fitted again using full-band data to obtain the final equivalent circuit fitting parameters.

[0015] As a preferred embodiment of the first aspect mentioned above, when dividing the full-band electrochemical impedance spectroscopy (EIS) data into subsets, the frequency range greater than 100 Hz is defined as the high-frequency band, the frequency range from 1 to 100 Hz as the mid-frequency band, and the frequency range less than 1 Hz as the low-frequency band.

[0016] As a preferred embodiment of the first aspect above, the relaxation time distribution characteristics ( The extraction method is as follows: generalized regularized inversion analysis is performed on electrochemical impedance spectroscopy (EIS) to obtain relaxation time distribution (DRT) spectrum. The peak intensity, time constant, peak area and half-peak width of each polarization peak in the spectrum are extracted as candidate features. The mutual information between each candidate feature and the health state (SOH) label is calculated. Based on the mutual information, the multiple candidate features with the highest correlation are selected as relaxation time distribution features.

[0017] As a preferred option in the first aspect mentioned above, the base learner used to train the primary prediction model is the XGBoost model. On each training subset, the hyperparameters of the XGBoost model are automatically tuned through Bayesian optimization and K-fold cross-validation mechanism. The XGBoost model with the optimal hyperparameters is trained using the same training subset to obtain a primary prediction model with the feature type corresponding to the training subset as input and the health status (SOH) as output.

[0018] As a preferred embodiment of the first aspect mentioned above, the machine learning regression model employs a Bayesian ridge regression model, wherein during the training phase, the Bayesian ridge regression model uses secondary features... As input, with health status (SOH) as the output label, the model weights are solved using maximum a posteriori probability estimation on the new samples, and the Gaussian posterior distribution of the weights is inferred to form a secondary fusion model. During the inference phase, the secondary fusion model calculates the secondary features of the input based on the Gaussian posterior distribution. The corresponding predicted mean and standard deviation of health status (SOH).

[0019] As a preferred embodiment of the first aspect above, when estimating the health status of the retired lithium battery to be tested, electrochemical impedance spectroscopy screening features are extracted from the electrochemical impedance spectroscopy (EIS) of the retired lithium battery to be tested. ), equivalent circuit fitting parameter characteristics ( ) and relaxation time distribution characteristics ( The three predicted values ​​are obtained by inputting them into three primary prediction models respectively. The three predicted values ​​are then concatenated into a secondary feature and input into a secondary fusion model to obtain the predicted mean and standard deviation of the state of health (SOH) of the battery under test. The confidence interval is then calculated according to the preset confidence coefficient.

[0020] In a second aspect, the present invention provides a computer program product, including a computer program / instruction, which, when executed by a processor, can implement the multi-information fusion-based performance testing method for retired lithium batteries as described in any of the first aspects above.

[0021] Thirdly, the present invention provides a computer electronic device, which includes a memory and a processor;

[0022] The memory is used to store computer programs;

[0023] The processor is configured to, when executing the computer program, implement the multi-information fusion-based performance testing method for retired lithium batteries as described in any of the first aspects above.

[0024] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0025] 1. This invention improves prediction accuracy through multi-source feature fusion: mutual information theory is used to independently reduce and screen the original EIS frequency points, ECM parameters and DRT features to ensure the effectiveness of multi-source physical features, effectively eliminate redundant noise, and avoid the dimensionality curse caused by directly splicing multi-source features.

[0026] 2. This invention solves the problem of fitting fractional-order models and improves the real generalization ability: On the one hand, it adopts a two-stage fitting strategy to obtain equivalent circuit parameters, avoiding the problem of getting trapped in local optima by direct global fitting and improving the accuracy of equivalent parameter extraction; on the other hand, in the multi-model fusion stage, it constructs a secondary feature matrix by extrapolation prediction, avoiding data leakage and ensuring that the model's generalization ability and prediction accuracy are real and reliable when facing unseen retired battery samples.

[0027] 3. This invention enables high-reliability uncertainty quantification: Traditional health status estimation for decommissioned batteries can only output a single SOH point estimate, which cannot reflect the reliability of the prediction results. This invention uses Bayesian ridge regression as a secondary fusion model, directly calculating the prediction standard deviation for each sample while outputting the SOH prediction mean. Attached Figure Description

[0028] To make the objectives, technical solutions, and advantages of the present invention clearer, the preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings, wherein:

[0029] Figure 1A schematic diagram illustrating the steps of a performance testing method for retired lithium batteries based on multi-information fusion;

[0030] Figure 2 This is the second-order fractional-order equivalent circuit model in the embodiments of the present invention;

[0031] Figure 3 This is a schematic diagram of the structure of a computer electronic device;

[0032] Figure 4 The above are simulation fitting results of the second-order equivalent circuit in the embodiments of the present invention;

[0033] Figure 5 This is a schematic diagram of the DRT curve and feature extraction obtained after using DRTtools in an embodiment of the present invention, where the horizontal axis is log(τ) which is the logarithm of the time constant τ, and the vertical axis is γ(τ) which is the impedance intensity corresponding to the time constant τ.

[0034] Figure 6 This refers to the SOH prediction results and uncertainty distribution on the test set in this embodiment of the invention. Detailed Implementation

[0035] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Many specific details are set forth in the following description to provide a thorough understanding of the present invention. However, the present invention can be practiced in many other ways different from those described herein, and those skilled in the art can make similar modifications without departing from the spirit of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below. Technical features in various embodiments of the present invention can be combined accordingly without mutual conflict.

[0036] like Figure 1 As shown, to address the problems of poor generalization ability of single-model estimation and lack of safety risk assessment caused by the complex aging mechanism of retired batteries, this invention provides a performance testing method for retired lithium batteries based on multi-information fusion, which includes the following steps S1 to S3. The specific implementation of each step is described in detail below.

[0037] S1. Obtain a training set consisting of electrochemical impedance spectroscopy (EIS) and state of health (SOH) labels of retired lithium batteries, and extract electrochemical impedance spectroscopy screening features from the EIS. ), equivalent circuit fitting parameter characteristics ( ) and relaxation time distribution characteristics ( The three types of features are combined with the health status (SOH) label to form three training subsets. The base learners are trained independently using the three training subsets to obtain three primary prediction models.

[0038] It should be noted that the above training set can be obtained by performing capacity calibration experiments and EIS tests on a large number of retired lithium batteries, extracting EIS data (including frequency, imaginary part, and real part) for each group of batteries, and calculating SOH (State of Health) through capacity calibration. ,in To calibrate capacity, (Rated capacity). In an embodiment of the present invention, when performing EIS testing on retired lithium batteries, 100% SOC is selected, the frequency is between 10k and 0.01Hz, and 10 points are collected for every 10 times the frequency, for a total of 61 points. The EIS data and SOH tags of each group of retired lithium batteries constitute a training sample, thereby constructing a training set containing EIS data and SOH tags.

[0039] It should be noted that the above-mentioned electrochemical impedance spectroscopy screening characteristics ( ), equivalent circuit fitting parameter characteristics ( ) and relaxation time distribution characteristics ( The three types of features are obtained by performing multi-dimensional analysis on the electrochemical impedance spectroscopy data of retired batteries. Specifically, for each frequency point, the real and imaginary parts of the impedance, the equivalent circuit fitting parameters, and the distributed relaxation time characteristics are selected using mutual information methods for independent dimensionality reduction. The specific selection methods for these three types of features are described below.

[0040] In embodiments of the present invention, electrochemical impedance spectroscopy screening features ( The extraction method is as follows: the real and imaginary parts of the impedance at each frequency point in the electrochemical impedance spectroscopy (EIS) are used as candidate features. The mutual information between each candidate feature and the health status (SOH) label is calculated. Based on the mutual information, the multiple candidate features with the highest correlation are selected as electrochemical impedance spectroscopy screening features.

[0041] It should be noted that in EIS data, each frequency point contains both real and imaginary parts of the impedance. It is assumed that the real and imaginary parts of the impedance at each frequency point need to be used as candidate features in mutual information calculation and correlation screening. Electrochemical impedance spectroscopy feature screening ( The specific number of candidate features included in ) is denoted as Therefore, after calculating the mutual information between each candidate feature and the SOH label, the candidate features are sorted in descending order according to their mutual information values, and the features with the highest mutual information values ​​are extracted. Several candidate features were used as screening features for electrochemical impedance spectroscopy. The value can be optimized and adjusted according to the actual situation.

[0042] In embodiments of the present invention, the equivalent circuit fitting parameter characteristics ( The extraction method is as follows: model the retired lithium battery as a component of inductance. Ohm resistance First fractional-order RC parallel unit (composed of polarization resistors) and constant phase element (connected in parallel), the second fractional-order RC parallel unit (composed of polarization resistors) and constant phase element Parallel connection), constant phase element The second-order fractional equivalent circuit model, which is formed by sequentially connecting the two circuits, is fitted with the equivalent circuit parameters using electrochemical impedance spectroscopy (EIS). The mutual information between each equivalent circuit parameter and the state of health (SOH) label is calculated. Based on the mutual information, the multiple equivalent circuit parameters with the highest correlation are selected as the equivalent circuit fitting parameter features.

[0043] like Figure 2 The diagram shows the circuit topology of the second-order fractional-order equivalent circuit model, and its theoretical equations are as follows:

[0044]

[0045] in, It is an inductor (H). It is an ohmic resistor (Ω). , f is the polarization resistance (Ω), and f is the frequency (Hz). It is an imaginary number. , , It is a constant phase element , , The corresponding admittance values ​​( ), , , It is a constant phase element , , Each has its corresponding diffusion index.

[0046] Theoretically, electrochemical impedance spectroscopy (EIS) can be directly used to analyze the equivalent circuit parameters in the theoretical equation. , , , f , , , , , Global fitting can be performed, but direct global fitting is prone to getting trapped in local optima. Therefore, in the embodiments of this invention, when fitting the equivalent circuit parameters using electrochemical impedance spectroscopy (EIS), a two-stage fitting strategy is adopted to obtain the equivalent circuit parameters. Specifically, in the two-stage fitting strategy, the fitting process is divided into two stages, and the fitting methods of each stage are as follows:

[0047] In the first stage, the full-band electrochemical impedance spectroscopy (EIS) data is divided into three subsets: high frequency, mid frequency, and low frequency. Then, the local equivalent circuit topology of each of the three frequency bands in the second-order fractional equivalent circuit model is initially fitted piecewise.

[0048] The specific frequency ranges for the above-mentioned high-frequency, mid-frequency, and low-frequency bands can be slightly adjusted. In the embodiments of this invention, when dividing the full-band electrochemical impedance spectroscopy (EIS) data into subsets, frequencies greater than 100Hz are considered high-frequency, 1-100Hz are considered mid-frequency, and less than 1Hz are considered low-frequency. Therefore, EIS spectrum data greater than 100Hz is used to fit the local equivalent circuit topology corresponding to the high-frequency band in the second-order fractional equivalent circuit model, EIS spectrum data from 1-100Hz is used to fit the local equivalent circuit topology corresponding to the mid-frequency band in the second-order fractional equivalent circuit model, and EIS spectrum data less than 1Hz is used to fit the local equivalent circuit topology corresponding to the low-frequency band in the second-order fractional equivalent circuit model. The local equivalent circuit topologies for each of the three frequency bands in the second-order fractional equivalent circuit model can be based on... Figure 1 The circuit topology shown is divided into the following sections:

[0049] For high-frequency bands, inductors are used. Ohm resistance The local equivalent circuit structure, formed by connecting the first fractional-order RC parallel unit in series, is fitted using high-frequency subset data to obtain the first part of the equivalent circuit parameters. , , , , The initial value.

[0050] When fitting high-frequency circuits, the theoretical formula used for fitting the local equivalent circuit structure can be expressed as:

[0051]

[0052] in, It is the impedance in the high-frequency range. It is an ohmic resistance (Ω).

[0053] For the mid-frequency range, an ohmic resistor is used. The local equivalent circuit structure, formed by connecting the second fractional-order RC parallel unit in series, is used to fit the mid-frequency subset data to obtain the second part of the equivalent circuit parameters. , , , The initial value.

[0054] When fitting in the mid-frequency band, the theoretical formula used for fitting the local equivalent circuit structure can be expressed as:

[0055]

[0056] in, It is the impedance in the mid-frequency range. It is an ohmic resistance (Ω).

[0057] For low-frequency bands, ohmic resistors are used. With constant phase element The third part of the equivalent circuit parameters is obtained by fitting the low-frequency subset data with the series-connected local equivalent circuit structure. , , The initial value.

[0058] When fitting in the low-frequency band, the theoretical formula used for fitting the local equivalent circuit structure can be expressed as:

[0059]

[0060] in, It is the impedance in the mid-frequency range. It is an ohmic resistance (Ω).

[0061] In the second stage, all equivalent circuit parameter values ​​obtained from the initial fitting in the first stage are then applied. , , , f , , , , , As the starting point for the parameters of the global fit, where The initial values ​​were obtained by fitting the high-frequency band. The second-order fractional equivalent circuit model was refitted using full-band data to obtain the final equivalent circuit fitting parameters.

[0062] It is important to note that during the second-stage fitting, it is not necessary to use subsets for fitting. Instead, the entire frequency band data from EIS is used to fit all equivalent circuit parameters in the theoretical model formula of the entire second-order fractional equivalent circuit model. , , , f , , , , , A global fit is performed, but the initial values ​​of each parameter during the fit are determined by the first stage.

[0063] In the two-stage fitting process described above, the fitting algorithm for each fitting step can be implemented using the Levenberg-Marquardt algorithm.

[0064] Equivalent circuit fitting parameter characteristics ( The specific number of parameters included in ) is denoted as Therefore, after calculating the mutual information between each equivalent circuit parameter and the SOH label, all equivalent circuit parameters are sorted in descending order according to their mutual information values, and the parameters with the highest mutual information values ​​are extracted. Each equivalent circuit parameter serves as a feature of the equivalent circuit fitting parameters. The value can be optimized and adjusted according to the actual situation.

[0065] In embodiments of the present invention, the relaxation time distribution characteristics ( The extraction method is as follows: generalized regularized inversion analysis is performed on electrochemical impedance spectroscopy (EIS) to obtain relaxation time distribution (DRT) spectrum. The peak intensity, time constant, peak area and half-peak width of each polarization peak in the spectrum are extracted as candidate features. The mutual information between each candidate feature and the health state (SOH) label is calculated. Based on the mutual information, the multiple candidate features with the highest correlation are selected as relaxation time distribution features.

[0066] It should be noted that the aforementioned Generalized Regularization Inversion (GRI) is an existing mathematical solution technique that can be used to solve ill-conditioned inverse problems in electrochemical impedance spectroscopy (EIS) analysis to extract the relaxation time distribution (DRT) spectrum. In the embodiments of this invention, the generalized regularization inversion analysis is implemented using the open-source DRTtools based on Python. The peak intensity, time constant, peak area, and half-peak width of each polarization peak in the DRT spectrum are used as four independent candidate features in the mutual information calculation. Relaxation time distribution features ( The specific number of candidate features included in ) is denoted as Therefore, after calculating the mutual information between each candidate feature and the SOH label, the candidate features are sorted in descending order according to their mutual information values, and the features with the highest mutual information values ​​are extracted. Each candidate feature serves as a relaxation time distribution feature. The value can be optimized and adjusted according to the actual situation.

[0067] It should also be noted that, in this invention, the formula for calculating the mutual information between any feature and the target variable is prior art, as detailed below:

[0068]

[0069] Where X represents the feature and Y represents the target variable. This represents the joint probability distribution of feature X and target variable Y. and Let X and Y represent the marginal probability distributions of feature X and target variable Y, respectively.

[0070] Based on the extraction of the above three types of features, electrochemical impedance spectroscopy screening features can be extracted from electrochemical impedance spectroscopy (EIS) for each training sample in the training set. ), equivalent circuit fitting parameter characteristics ( ) and relaxation time distribution characteristics ( The three types of features are respectively matched with the health status labels corresponding to this training sample. By combining these, three new binary sample pairs can be formed and added to the training subset. The binary sample pairs in the first training subset are as follows: The sample pairs of the second training subset are in the form of The sample pairs of the third training subset are in the form of .

[0071] In this invention, based on these three training subsets, base learners can be independently trained using each subset to obtain three primary prediction models. The base learners used to train the primary prediction models can be selected from various machine learning models. Each base learner is an independent and single prediction model in the ensemble model. To ensure the reliability of subsequent secondary feature calculations, the base learners used in the three primary prediction models need to be as consistent as possible.

[0072] In an embodiment of the present invention, the base learner used to train the primary prediction model is an XGBoost model. On each training subset, the hyperparameters of the XGBoost model are automatically tuned using Bayesian optimization and K-fold cross-validation. The XGBoost model with the optimal hyperparameters is then trained using the same training subset, resulting in a primary prediction model that takes the feature type corresponding to that training subset as input and the state of health (SOH) as output. It should be noted that the three primary training models ultimately require different input features. The input to the first primary prediction model is an electrochemical impedance spectroscopy screening feature (…). The input to the second primary prediction model is the equivalent circuit fitting parameter features. The input to the third primary prediction model is the relaxation time distribution characteristics ( The outputs of the three primary prediction models are all health status (SOH) labels.

[0073] S2. Using a K-fold cross-validation strategy, three primary prediction models generate three out-of-fold prediction values ​​for each training sample in the training set, and these values ​​are concatenated to construct the secondary features of that training sample. Secondary features of all training samples in the training set After standardization, the samples are combined with their respective health status (SOH) labels to form new samples, which are then used to train the machine learning regression model to obtain a trained secondary fusion model.

[0074] It should be noted that K-fold cross-validation is an existing technology, and the specific steps of cross-validation are as follows:

[0075] Divide the training set into A set of non-overlapping subsets;

[0076] Each iteration uses The model is trained on one subset, and predictions are made on the remaining one subset to obtain the out-of-place (OOF) prediction value for each sample in the subset. This out-of-place prediction value can be regarded as a secondary feature of each sample.

[0077] cycle Subsequently, the outlier prediction value for each sample in the training set is obtained. To facilitate matrix operations in the model, all outlier prediction values ​​are concatenated according to sample index to form a leak-free secondary feature matrix. The secondary feature matrix is ​​then Z-score standardized. The standardized outlier prediction value of each sample, together with its respective health status (SOH) label, forms a new sample, which can then be used to train the machine learning regression model, resulting in a trained secondary fusion model.

[0078] In an embodiment of the present invention, the machine learning regression model employs a Bayesian Ridge regression model. During the training phase, the Bayesian Ridge regression model uses secondary features... As input, with health status (SOH) as the output label, the model weights are solved using maximum a posteriori probability estimation on the new samples, and the Gaussian posterior distribution of the weights is inferred to form a secondary fusion model. During the inference phase, the secondary fusion model calculates the secondary features of the input based on the Gaussian posterior distribution. The corresponding predicted mean and standard deviation of health status (SOH). The predicted mean of SOH for each sample can be used as a point estimate, while the corresponding standard deviation can be used as a measure of uncertainty.

[0079] In addition, in the embodiments of the present invention, based on the predicted mean and predicted standard deviation of the health status (SOH) output by the secondary fusion model, the upper and lower limits of the confidence interval can also be calculated according to the preset confidence coefficient Z: upper limit = predicted mean + Z × predicted standard deviation, lower limit = predicted mean - Z × predicted standard deviation; wherein, when a 95% confidence level is set, it is recommended that the coefficient Z be set to 1.96.

[0080] S3. The three trained primary prediction models and the secondary fusion model are cascaded and combined to estimate the health status of the retired lithium battery under test.

[0081] It should be noted that the method for estimating the health status of the retired lithium battery under test is actually the inference process of the model. The feature extraction, processing, and input-output transformation essentially remain the same as the training samples. Specifically, in the embodiments of this invention, when estimating the health status of the retired lithium battery under test, electrochemical impedance spectroscopy screening features are extracted from the electrochemical impedance spectroscopy (EIS) of the retired lithium battery under test. ), equivalent circuit fitting parameter characteristics ( ) and relaxation time distribution characteristics ( The three predicted values ​​are obtained by inputting them into three primary prediction models respectively. The three predicted values ​​are then concatenated into a secondary feature and input into a secondary fusion model to obtain the predicted mean and standard deviation of the state of health (SOH) of the battery under test. The confidence interval is then calculated according to the preset confidence coefficient.

[0082] Furthermore, based on the same inventive concept, such as Figure 3 As shown, the present invention also provides a computer electronic device corresponding to the multi-information fusion-based performance testing method for retired lithium batteries provided in the above embodiments, which includes a memory and a processor;

[0083] The memory is used to store computer programs;

[0084] The processor is used to implement the multi-information fusion-based performance testing method for retired lithium batteries as described above when executing the computer program.

[0085] Furthermore, the logical instructions in the aforementioned memory can be implemented as software functional units and, when sold or used as independent products, can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention.

[0086] Therefore, based on the same inventive concept, the present invention provides a computer-readable storage medium corresponding to a multi-information fusion-based method for testing the performance of retired lithium batteries. The storage medium stores a computer program, which, when executed by a processor, can realize the multi-information fusion-based method for testing the performance of retired lithium batteries as described above.

[0087] Therefore, based on the same inventive concept, the present invention provides a computer program product, including a computer program / instruction, which, when executed by a processor, can realize the aforementioned method for performance testing of retired lithium batteries based on multi-information fusion.

[0088] Specifically, in the computer-readable storage medium of the above three embodiments, the stored computer program is executed by a processor, which can perform the aforementioned steps S1 to S3.

[0089] It is understood that the aforementioned storage media may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Furthermore, the storage media may also be various media capable of storing program code, such as USB flash drives, external hard drives, magnetic disks, or optical discs.

[0090] It is understood that the processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0091] It should also be noted that those skilled in the art will understand that, for the sake of convenience and brevity, the specific working process of the system described above can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here. In the embodiments provided in this application, the division of steps or modules in the system and method is merely a logical functional division, and there may be other division methods in actual implementation. For example, multiple modules or steps may be combined or integrated together, and a module or step may also be split.

[0092] The present invention will further demonstrate the detailed implementation process and technical effects of the multi-information fusion-based performance testing method for retired lithium batteries shown in steps S1 to S3 on a specific dataset through a specific embodiment, so as to facilitate understanding of the essence of the present invention.

[0093] Example

[0094] In this embodiment, the specific implementation steps are the same as those of the multi-information fusion-based performance testing method for retired lithium batteries shown in steps S1 to S3 above. The method principle and steps will not be repeated here. The main focus is on the specific dataset, some specific parameter settings and implementation results of this embodiment.

[0095] Step 1: Data Acquisition and Preprocessing

[0096] In this embodiment, several retired lithium batteries of the same model were collected and capacity calibration experiments were conducted to calculate their State of Health (SOH) data (SOH = current calibrated capacity / rated capacity × 100%). Under a static state with a State of Charge (SOC) of 100%, the batteries underwent an Electro-Insulation (EIS) test. The excitation frequency range of the EIS test was 10kHz-0.01Hz, with 10 data points collected every 10 octaves. All collected battery data were randomly divided into a training set and a test set in a 7:3 ratio.

[0097] Step 2: Extraction of original EIS features and screening of features by electrochemical impedance spectroscopy (EIS) )choose

[0098] Refer to the electrochemical impedance spectroscopy screening characteristics in step S1 above ( The extraction method involves calculating the mutual information values ​​between the real and imaginary parts of the impedance at each frequency point of the EIS and the SOH label, based on the data in the training set. The data are then sorted in descending order of mutual information values, and the top 10 frequency points with the strongest correlations are selected. Specifically: , , , , , , , , , ], where Z' represents the real part, Z'' represents the imaginary part, and the subscript indicates the corresponding frequency point. The real or imaginary parts of these 10 frequency points are the electrochemical impedance spectroscopy screening features used in this embodiment ( )

[0099] Step 3: ECM parameter identification and equivalent circuit fitting parameter characteristics ( )choose

[0100] Referring to the equivalent circuit fitting parameter characteristics in step S1 above ( The extraction method is based on a second-order fractional equivalent circuit model, employing a two-stage fitting strategy to extract parameters.

[0101] Phase 1: Based on the local equivalent circuit topology and the Levenberg-Marquardt algorithm, the impedance spectrum is initially fitted in segments for different frequency bands to obtain the initial parameter values ​​of each equivalent element in the second-order fractional model.

[0102] For the high-frequency region, a local structure consisting of ohmic resistors and fractional RC parallel units connected in series (i.e., R0-(CPE / / Rp) structure) is used for fitting to extract the initial values ​​of ohmic resistance and high-frequency polarization parameters.

[0103] For the mid-frequency region, a local structure consisting of ohmic resistors and fractional-order RC parallel units connected in series (i.e., R0-(CPE / / Rp) structure) is used for fitting to extract the initial values ​​of the mid-frequency polarization parameters.

[0104] For the low-frequency diffusion region, a local structure consisting of a resistor and a constant-phase element connected in series (i.e., R0-CPE structure) is used for fitting to extract the initial values ​​of the diffusion parameters;

[0105] The second stage: Using all the initial parameter values ​​obtained in the first stage as the starting point for global fitting, the Levenberg-Marquardt algorithm is used again to perform second-order fractional equivalent circuit fitting on the full-band impedance spectrum data to extract the final parameters. Figure 4 This demonstrates an exemplary second-order equivalent circuit simulation of the final global fitting result. In this embodiment, based on the training set data, mutual information is used to calculate the first 8 ECM features related to SOH: [R2,Q1, Q3, Q2, n3, n2, R0,n1]. These 8 features are the equivalent circuit fitting parameter features used in this embodiment. ).

[0106] Step 4: DRT variation and relaxation time distribution characteristics ( )choose

[0107] Referring to the relaxation time distribution characteristics in step S1 above ( The extraction method employed the open-source DRTtool tool, written in Python, to convert the EIS data from the training and test sets into relaxation time distribution (DRT) spectra. For the four feature peaks in the DRT spectra, relaxation time (X), peak intensity (H), peak area (W), and half-peak width (A) were extracted, resulting in a total of 16 original DRT features. Figure 5This document demonstrates an exemplary DRT curve and feature extraction diagram obtained using DRTtools. Based on the training set data and mutual information, the top 10 DRT features strongly correlated with SOH are calculated: [X3, X4, A3, W1, W3, H3, H2, X2, W2, A1], where letters represent feature physical quantities and numbers represent peak numbers. These DRT features are the relaxation time distribution features used in this embodiment. ).

[0108] Step 5: Base Model Training

[0109] Referring to the training method of the primary prediction model in step S1 above, the three types of features obtained in steps 2, 3, and 4 are combined with SOH labels to construct three independent training subsets (i.e., EIS feature subset, ECM feature subset, and DRT feature subset). For each subset, an XGBoost regression model is constructed as a base learner. Using a Bayesian optimization algorithm combined with 5-fold cross-validation, the hyperparameters (including learning rate, maximum tree depth, subsample ratio, etc.) of each XGBoost base model are automatically optimized based on their respective training subsets. After fixing their optimal hyperparameters, they are retrained on their respective training subsets, finally obtaining three primary prediction models (EIS-XGBoost, ECM-XGBoost, and DRT-XGBoost).

[0110] Step 6: Construct the secondary feature matrix

[0111] Referring to the secondary feature matrix training method in step S2 above, a 5-fold cross-validation strategy is used to generate secondary training features to prevent data leakage and overfitting during training. The specific process is as follows: the training set is divided into 5 folds. Each time, 4 folds of data are used to train the three base models in step 5, and predictions are made for the 1-fold data that did not participate in the current training round. Each out-of-fold battery sample yields 3 out-of-fold predicted values. This process is repeated 5 times until all training samples have been predicted once as out-of-fold data. The 3 out-of-fold predicted values ​​of the same battery sample are concatenated column-wise to construct a secondary feature matrix of dimension N×3 (N is the number of training set samples). Similarly, using the three base models trained on the complete training set, we can directly predict the test set and concatenate them to obtain the secondary feature matrix of the test set. .

[0112] It is important to note that the secondary features are constructed in matrix form here. and This is only to facilitate batch parallel computation of the model, but for each sample, it is still a feature vector composed of three out-of-bounds predicted values.

[0113] Step 7: Integrate model training with uncertainty quantization output

[0114] For secondary characteristic matrix and Z-score standardization is performed. A Bayesian ridge regression model is used as the secondary fusion model, taking secondary features as input and the state of health (SOH) as the output label. During the training phase, the standardized features are used... The model is trained by fitting the SOH labels. On new binary samples with secondary features as input and SOH labels as output, the model weights are solved using maximum a posteriori probability estimation, and the Gaussian posterior distribution of the weights is inferred. After fixing the Gaussian posterior distribution, a secondary fusion model is formed. During the inference phase, the secondary feature matrix of the test set is used... Secondary features in the model are used as test objects, and the standardized features are used as test objects. The input is fed into a trained Bayesian ridge regression model, which outputs the predicted mean SOH and the corresponding predicted standard deviation for each test sample.

[0115] In this embodiment, based on the predicted mean and predicted standard deviation of the output SOH, the upper and lower limits of the confidence interval are calculated with a 95% confidence level: upper limit = predicted mean + 1.96 × predicted standard deviation, lower limit = predicted mean - 1.96 × predicted standard deviation.

[0116] Finally, the verification results of this embodiment on the test set are as follows: Figure 6 As shown, the mean absolute error (MAE) of the SOH point estimate is 1.168%; mean prediction standard deviation for all samples in the test set is 0.89%, and the average width of the 95% confidence interval constructed based on this is only 7.54%. Statistically, more than 99% of the true SOH values ​​in the test set fall within this calculated confidence interval.

[0117] To further demonstrate the advantages of the estimation framework shown in S1-S3 of the present invention, a series of comparative examples are introduced below through ablation experiments to compare and demonstrate the advantages of each module in the framework of the present invention. In each comparative example, the specific feature forms and extraction methods of the three types of features, EIS, ECM, and DRT, are respectively compared with the electrochemical impedance spectroscopy screening features in this embodiment ( ), equivalent circuit fitting parameter characteristics ( ) and relaxation time distribution characteristics ( The models, training sets, and test sets used are completely identical, which facilitates horizontal comparison.

[0118] Comparative Example 1: XGBoost Prediction Based on a Single EIS Feature

[0119] The difference between this comparative example and the embodiment is that only electrochemical impedance spectroscopy features are used: [ , , , , , , , , , As input to the model, an XGBoost model is trained on the training set, and this XGBoost is used to perform SOH prediction on the test set.

[0120] Comparative Example 2: XGBoost Prediction Based on a Single ECM Feature

[0121] The difference between this comparative example and the embodiment is that only the equivalent circuit model parameter features: [R2,Q1, Q3,Q2, n3, n2, R0, n1] are used as model input to train an XGBoost model on the training set, and this XGBoost is used to perform SOH prediction on the test set.

[0122] Comparative Example 3: XGBoost Prediction Based on a Single DRT Feature

[0123] The difference between this comparative example and the embodiment is that it only uses the relaxation time distribution features [X3, X4, A3, W1, W3, H3, H2, X2, W2, A1] as model input, trains an XGBoost model on the training set, and uses this XGBoost to perform SOH prediction on the test set.

[0124] Comparative Example 4: XGBoost Prediction Based on Early Fusion of EIS+ECM+DRT

[0125] Compared with the previous example, the difference in this comparative example is that the 28 dimensions of the three types of features, EIS, ECM and DRT, are directly concatenated into a single feature vector. An XGBoost model is trained on the training set and this XGBoost is used to perform SOH prediction on the test set.

[0126] Comparative Example 5: Bayesian Ridge Regression Model Prediction Based on Early Fusion of EIS+ECM+DRT

[0127] The difference between this comparative example and the previous example is that only the three types of features, EIS, ECM, and DRT, totaling 28 dimensions, are directly concatenated into a single feature vector. A Bayesian Ridge regression model is trained on the training set, and this Bayesian Ridge is used to make predictions on the test set.

[0128] Comparative Example 6: Multi-source prediction based on late-stage simple average fusion

[0129] The difference between this comparative example and the previous example is that XGBoost was used to independently train base models for three types of features, EIS, ECM, and DRT, on the training set. After obtaining three predicted values, the arithmetic mean (Simple Average) was used as the final SOH calculation result.

[0130] Comparative Example 7: Two-stage prediction based on a single base learner

[0131] Compared with the previous example, the difference in this comparative example is that all 28-dimensional features of EIS, ECM, and DRT are directly concatenated and then input into a single XGBoost base learner for training. Five-fold cross-validation is used to generate an out-of-fold prediction value (OOF) for each out-of-fold battery sample. The OOFs of all N samples are recombined to form an N×1 dimensional secondary feature matrix. This single out-of-fold prediction value is then input into the Bayesian Ridge secondary model for fusion to obtain the predicted mean and predicted standard deviation of SOH.

[0132] Finally, the performance metrics of the above pairs of examples and embodiments are compared as follows:

[0133] Table 1 Performance comparison of different SOH prediction methods

[0134]

[0135] As shown in Table 1, the present invention has a series of advantages over Comparative Example 1, specifically:

[0136] (1) Comparative Examples 1-3 show that all single information sources have limitations, and even the optimal single source (EIS) cannot achieve the accuracy of the present invention;

[0137] (2) Comparative examples 4-5 show that the early fusion strategy has inherent defects. Regardless of whether XGBoost or BayesianRidge is used, directly splicing heterogeneous features leads to performance degradation, proving that the early fusion route is not feasible.

[0138] (3) Comparative Example 6 shows that although late-stage fusion avoids heterogeneous interference, the fixed weights cannot be adaptively adjusted according to the reliability of each source, and the low-precision source of ECM directly drags down the overall performance.

[0139] (4) Comparative Example 7 shows that the two-stage strategy of early direct fusion cannot avoid the confusion of physical meaning and feature interference caused by early splicing of heterogeneous features, and completely loses the advantage of adaptive weighted fusion of multi-source information because the secondary model only receives 1-dimensional input.

[0140] (5) The present invention adopts the technical route of "feature modeling + late probability fusion". Bayesian Ridge automatically infers the optimal weights, and outputs the prediction uncertainty while achieving the highest prediction accuracy. This solves the technical problems of low accuracy of multi-source information fusion and inability to assess risk in the prior art.

[0141] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the invention. Those skilled in the art can make various changes and modifications without departing from the spirit and scope of the invention. Therefore, all technical solutions obtained through equivalent substitution or transformation fall within the protection scope of the present invention.

Claims

1. A method for performance testing of retired lithium batteries based on multi-information fusion, characterized in that, Includes the following steps: S1. Obtain a training set consisting of electrochemical impedance spectra and health status labels of retired lithium batteries, and extract electrochemical impedance spectral screening features, equivalent circuit fitting parameter features, and relaxation time distribution features from the electrochemical impedance spectra. Combine the three types of features with the health status labels to form three training subsets, and use the three training subsets to independently train the base learner to obtain three primary prediction models. S2. Using the K-fold cross-validation strategy, three out-of-fold prediction values ​​are generated for each training sample in the training set by three primary prediction models and concatenated to construct the secondary features of the training sample. The secondary features of all training samples in the training set are standardized and combined with their respective health status labels to form new samples. The machine learning regression model is then trained to obtain the trained secondary fusion model. S3. The three trained primary prediction models and the secondary fusion model are cascaded and combined to estimate the health status of the retired lithium battery under test.

2. The method for performance testing of retired lithium batteries based on multi-information fusion as described in claim 1, characterized in that, The method for extracting the electrochemical impedance spectroscopy screening features is as follows: using the real and imaginary parts of the impedance at each frequency point in the electrochemical impedance spectrum as candidate features, calculating the mutual information between each candidate feature and the health status label, and selecting multiple candidate features with the highest correlation as electrochemical impedance spectroscopy screening features based on the mutual information.

3. The method for performance testing of retired lithium batteries based on multi-information fusion as described in claim 1, characterized in that, The method for extracting the equivalent circuit fitting parameter features is as follows: the retired lithium battery is modeled as a second-order fractional equivalent circuit model consisting of an inductor, an ohmic resistor, a first fractional-order RC parallel unit, a second fractional-order RC parallel unit, and a constant-phase element connected in series. The equivalent circuit parameters are fitted using electrochemical impedance spectroscopy, and the mutual information between each equivalent circuit parameter and the health status label is calculated. Based on the mutual information, the multiple equivalent circuit parameters with the highest correlation are selected as the equivalent circuit fitting parameter features.

4. The method for performance testing of retired lithium batteries based on multi-information fusion as described in claim 3, characterized in that, When fitting equivalent circuit parameters using electrochemical impedance spectroscopy, a two-stage fitting strategy is employed, wherein: In the first stage, the full-band electrochemical impedance spectroscopy data is divided into three subsets: high frequency, mid frequency, and low frequency. Then, the local equivalent circuit topology of each of the three frequency bands in the second-order fractional equivalent circuit model is preliminarily fitted piecewise. In the second stage, all the equivalent circuit parameter values ​​obtained from the initial fitting in the first stage are used as the starting point for global fitting. The second-order fractional equivalent circuit model is then fitted again using full-band data to obtain the final equivalent circuit fitting parameters.

5. The method for performance testing of retired lithium batteries based on multi-information fusion as described in claim 1, characterized in that, The method for extracting the relaxation time distribution features is as follows: generalized regularized inversion analysis is performed on the electrochemical impedance spectroscopy to obtain the relaxation time distribution spectrum. The peak intensity, time constant, peak area and half-peak width of each polarization peak in the spectrum are extracted as candidate features. The mutual information between each candidate feature and the health status label is calculated. Based on the mutual information, the multiple candidate features with the highest correlation are selected as the relaxation time distribution features.

6. The method for performance testing of retired lithium batteries based on multi-information fusion as described in claim 1, characterized in that, The base learner used to train the primary prediction model is the XGBoost model. On each training subset, the hyperparameters of the XGBoost model are automatically tuned through Bayesian optimization and K-fold cross-validation. The XGBoost model with the optimal hyperparameters is trained on the same training subset to obtain a primary prediction model with the feature type corresponding to the training subset as input and the health status as output.

7. The method for performance testing of retired lithium batteries based on multi-information fusion as described in claim 1, characterized in that, The machine learning regression model adopts the Bayesian Ridge Regression model. During the training phase, the Bayesian Ridge Regression model takes secondary features as input and health status as output label. On the new sample, it solves the model weights through maximum a posteriori probability estimation and infers the Gaussian posterior distribution of the weights to form a secondary fusion model. During the inference phase, the secondary fusion model calculates the predicted mean and standard deviation of the health status corresponding to the input secondary features based on the Gaussian posterior distribution.

8. The method for performance testing of retired lithium batteries based on multi-information fusion as described in claim 1, characterized in that, When estimating the health status of a retired lithium battery under test, electrochemical impedance spectroscopy screening features, equivalent circuit fitting parameter features, and relaxation time distribution features are extracted from the electrochemical impedance spectroscopy of the retired lithium battery under test. These features are then input into three primary prediction models to obtain three predicted values. The three predicted values ​​are then concatenated into secondary features and input into a secondary fusion model to obtain the predicted mean and standard deviation of the health status of the battery under test. Finally, a confidence interval is calculated according to a preset confidence coefficient.

9. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instruction is executed by the processor, it can realize the multi-information fusion-based performance testing method for retired lithium batteries as described in any one of claims 1 to 8.

10. A computer electronic device, characterized in that, Including memory and processor; The memory is used to store computer programs; The processor is configured to, when executing the computer program, implement the multi-information fusion-based performance testing method for retired lithium batteries as described in any one of claims 1 to 8.

Images