A lithium battery health status prediction method based on feature screening technology

Abstract

The present invention discloses a method for predicting the health status of a lithium battery based on feature screening technology. The method for predicting the health status of a lithium battery based on feature screening technology includes the following steps: acquisition of a data set; extraction and preliminary processing of initial features; screening of feature sets; preliminary training of a machine learning model; evaluation and optimization of a machine learning model; prediction and verification of the health status of a lithium battery. The present invention utilizes an effective feature screening technology based on the Fisher information matrix to indirectly reveal the complex relationship between features and target outputs by evaluating the uncertainty of parameter estimation. It can not only effectively capture the linear and nonlinear interactions between features, but also reflect the joint influence of each feature in a multi-parameter model, providing a more comprehensive and accurate basis for feature selection. In addition, the feature screening technology based on the Fisher information matrix is ​​applicable to linear and nonlinear models, and especially exhibits strong robustness and accuracy when processing multi-parameter models with complex interactions.

Description

Technical Field

[0001] The present invention relates to the technical field of lithium battery health state prediction, and in particular to a lithium battery health state prediction method based on feature screening technology. Background Art

[0002] Lithium-ion batteries have become an important choice for clean energy in recent years due to their high energy density, high power density, long service life and low self-discharge characteristics. When lithium-ion batteries are used as power sources, accurate evaluation of their online operating status is crucial for battery management systems because it directly affects the service life and safety of electric vehicles. As one of the key indicators of battery performance, health status has become a research focus. However, since battery degradation runs through its entire life cycle and is affected by multiple factors such as temperature and charge and discharge rate, the prediction of lithium battery health status remains a challenging problem.

[0003] The key to improving the accuracy of lithium battery health state prediction lies in feature extraction and processing. Accurately extracting features related to lithium battery health state can not only improve the accuracy of the prediction model, but also enhance the model's ability to explain the battery health state. However, the correlation between the features and the lithium battery health state determines whether the extracted features can truly reflect the changes in the lithium battery health state, which directly affects the prediction performance. Traditional correlation analysis methods are less effective when processing high-dimensional data and have difficulty dealing with the dimensionality disaster problem, resulting in reduced effectiveness in complex data. Existing methods usually ignore the interactions between features, which play a key role in the prediction of lithium battery health state, especially in the case of multiple features. Ignoring them may significantly reduce the prediction accuracy of the model. Therefore, how to adopt more comprehensive and effective methods in feature selection and analysis to obtain more accurate and meaningful prediction results has become an important issue that needs to be solved urgently. Summary of the invention

[0004] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or be understood by implementing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures particularly pointed out in the description and other drawings of the description.

[0005] The purpose of the present invention is to overcome the above-mentioned shortcomings and provide a lithium battery health status prediction method based on feature screening technology.

[0006] The technical solution adopted by the present invention to solve the technical problem is: a lithium battery health status prediction method based on feature screening technology, comprising the following steps:

[0007] Acquisition of data sets: selecting lithium battery data in multiple different states to obtain data sets;

[0008] Extraction and preliminary processing of initial features: extract a set of initial features from the data set, and perform preliminary processing on the initial features to remove redundant and irrelevant features to generate a feature set;

[0009] Screening of feature sets: constructing an effective feature screening technology based on Fisher information matrix to screen feature sets and divide them into training sets and test sets;

[0010] Initial training of the machine learning model: input the training set into the machine learning model to predict the health status of lithium batteries, and select the best model and the best hyperparameter configuration through tuning and cross-validation of the machine learning model parameters to output preliminary prediction results;

[0011] Evaluation and optimization of machine learning models, using root mean square error and mean absolute error to evaluate preliminary prediction results, and optimizing hyperparameters and feature combinations in the training process based on the evaluation error to improve the prediction performance of machine learning models;

[0012] Prediction and verification of lithium battery health status: The optimized machine learning model is used to predict the health status of lithium batteries on the test set, and the prediction results are compared with those obtained by other feature screening methods to verify the accuracy and generalization ability of the lithium battery health status prediction results.

[0013] The present invention utilizes the effective feature screening technology based on the Fisher information matrix to indirectly reveal the complex relationship between features and target output by evaluating the uncertainty of parameter estimation. It can not only effectively capture the linear and nonlinear interactions between features, but also reflect the joint influence of each feature in the multi-parameter model, providing a more comprehensive and accurate basis for feature selection. In addition, the feature screening technology based on the Fisher information matrix is ​​suitable for linear and nonlinear models, especially when dealing with multi-parameter models with complex interactions, it shows strong robustness and accuracy.

[0014] In some embodiments, the lithium battery data includes charging current, voltage, temperature and discharging current, voltage, temperature, and the initial features include the battery voltage, temperature, current waveform during charging and discharging, and number of cycles. Selecting appropriate lithium battery data as a data set is the basis for feature screening.

[0015] In some embodiments, the method for preliminary processing of the initial features includes a statistical analysis method, a frequency domain feature extraction method, and a time domain feature extraction method. The statistical analysis method processes the initial features to form statistical features, the frequency domain feature extraction method extracts the battery frequency characteristics, and the time domain feature extraction method extracts the battery time characteristics. The statistical analysis method includes but is not limited to mean, variance, and entropy, the frequency domain feature extraction method includes but is not limited to Fourier transform, and the time domain feature extraction method includes but is not limited to peak value and zero crossing point.

[0016] In some embodiments, the feature screening technique based on Fisher information matrix includes Fisher information matrix construction, convex optimization problem solving, and importance ranking.

[0017] In some embodiments, the Fisher information matrix is ​​constructed by using the partial derivative matrix calculated by back propagation in the neural network model to construct an approximate Fisher information matrix, converting the feature selection process into an optimization problem, and the specific steps are:

[0018] Input all the extracted features into the neural network for multiple trainings and save the training results;

[0019] The back propagation function is used to perform reverse transmission and pass it to the first layer of the neural network to obtain the sensitivity matrix, which is expressed as:

[0020] ,

[0021] Among them, L is the output of the first layer, is the corresponding parameter feature, is the symbol of partial derivative;

[0022] The average second-order statistics related to the influence of parameters on the output are obtained through the sensitivity matrix, and its expression is:

[0023] ,

[0024] Where n is the number of samples, Q is the measurement error covariance matrix, T is the transpose of the matrix, S is the sensitivity matrix of the output to the input features, and by calculating the product of S and the transpose of S, the average second-order statistics related to the impact of the features on the output can be obtained. The inverse matrix of F is used as an approximation of the error covariance matrix;

[0025] Convex optimization problem solving includes:

[0026] Consider feature selection as the optimal experimental design problem, select p features from n targets to retain, where n ≥ p, and the optimal experimental design problem expression is:

[0027] ,

[0028] in, represents the characteristic parameter, is the optimal feature selection set, Represents the characteristics The information matrix, is a row vector consisting of 1s in all entries, , is a non-negative weight factor that can be chosen to be 1 or 0, By all The collection composed of represents the scalar Fisher information matrix inverse function. argmin is used to represent the parameter value of a function that achieves the minimum value in the domain. For each data set, S is a positive definite matrix.

[0029] For designs involving a large number of weighting factors, the continuous optimization problem is converted into a convex optimization problem, expressed as:

[0030] ,

[0031] Among them, the weighted term It is defined as a continuous variable ranging from 0 to 1;

[0032] Applying the E-optimal design criterion to transform into an equivalent semidefinite programming problem to determine the key features is expressed as:

[0033] ,

[0034] in, is the minimum eigenvalue, I is the unit matrix, Indicates that for all ;

[0035] The importance ranking is obtained based on the equivalent semidefinite programming problem The value of Represents the importance of each feature obtained. Arrange in order of size to get the order of relevance. The features are retained for subsequent network training.

[0036] In some embodiments, the machine learning model includes a multi-layer perceptron, a convolutional neural network, or a graph neural network.

[0037] In some embodiments, in the configuration of the network, the multilayer perceptron consists of three layers, the convolutional neural network includes two convolutional layers, two pooling layers and one fully connected layer, the graph neural network consists of two graph convolutional layers and one fully connected layer, and the multilayer perceptron, convolutional neural network, and graph neural network all apply batch normalization and linear rectification function activation function, and set a random inactivation rate of 0.1 to prevent overfitting.

[0038] In some embodiments, 80% of the filtered feature set is used as a training set, and 20% of the filtered feature set is used as a test set.

[0039] In some embodiments, other feature screening methods include Spearman's method, Pearson's method, and Kendall's method.

[0040] In some embodiments, it also includes: model updating and adaptive adjustment, through regular model updating, online learning or incremental learning methods, the model parameters are gradually adjusted to ensure that the prediction accuracy is continuously improved as the battery aging process progresses.

[0041] By adopting the above technical solution, the beneficial effects of the present invention are:

[0042] 1. The present invention utilizes the effective feature screening technology based on the Fisher information matrix to indirectly reveal the complex relationship between features and target outputs by evaluating the uncertainty of parameter estimation. It can not only effectively capture the linear and nonlinear interactions between features, but also reflect the joint influence of each feature in the multi-parameter model, providing a more comprehensive and accurate feature selection basis; in addition, the feature screening technology based on the Fisher information matrix is ​​applicable to linear and nonlinear models, especially when dealing with multi-parameter models with complex interactions, it shows strong robustness and accuracy.

[0043] 2. The feature screening technology based on the Fisher information matrix of the present invention evaluates the importance of features by quantifying the sensitivity of the model output, rather than relying solely on traditional statistical significance tests or correlation-based metrics. In linear regression models, the diagonal elements based on the feature screening technology based on the Fisher information matrix are closely related to the accuracy of parameter estimation, and indirectly express the linear relationship between the features and the target variables. In nonlinear regression models, the nonlinear relationship between parameters often leads to changes in the non-diagonal elements based on the feature screening technology based on the Fisher information matrix. The non-diagonal elements based on the feature screening technology based on the Fisher information matrix reflect the covariance between the parameters, and further reveal how the parameters in the model jointly affect the model output. When two parameters are highly correlated, their covariance value is large, which further strengthens the expression of this relationship.

[0044] 3. The present invention can not only effectively process simple linear and nonlinear models, but is particularly suitable for multi-parameter models. It can more comprehensively reveal how multiple parameters work together on the model output. At the same time, the feature screening technology based on the Fisher information matrix provides a more in-depth analysis of the interaction between estimation accuracy and parameters.

[0045] It is to be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present disclosure.

[0046] Undoubtedly, these and other objects of the present invention will become more apparent after the following detailed description of the preferred embodiment described with reference to various figures and drawings.

[0047] In order to make the above and other purposes, features and advantages of the present invention more obvious and easy to understand, one or several preferred embodiments are given below and described in detail with reference to the accompanying drawings. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used together with the embodiments of the present invention to explain the present invention and do not constitute a limitation of the present invention.

[0049] In the drawings, the same reference numerals are used for the same components and the drawings are schematic and not necessarily drawn to scale.

[0050] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only one or several embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on such drawings without paying creative work.

[0051] Figure 1 It is an overall flow chart of a lithium battery health status prediction method based on feature screening technology of the present invention;

[0052] Figure 2 A flowchart of a lithium battery health status prediction method based on feature screening technology in some embodiments of the present invention;

[0053] Figure 3 This is an overall block diagram of lithium battery health status prediction in some embodiments of the present invention. DETAILED DESCRIPTION

[0054] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention, but are not used to limit the present invention.

[0055] In the present invention, unless otherwise clearly specified and limited, the first feature "on" or "under" the second feature may be that the first and second features are in direct contact, or the first and second features are in indirect contact through an intermediate medium. In the description of this specification, the description with reference to the terms "one embodiment", "some embodiments", "example", "specific example", or "some examples" means that the specific features, structures, materials or characteristics described in conjunction with the embodiment or example are included in at least one embodiment or example of the present invention. In this specification, the schematic representation of the above terms does not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials or characteristics described may be combined in an appropriate manner in any one or more embodiments or examples.

[0056] Reference Figure 1-Figure 3 , Figure 1 It is an overall flow chart of a lithium battery health status prediction method based on feature screening technology of the present invention; Figure 2 A flowchart of a lithium battery health status prediction method based on feature screening technology in some embodiments of the present invention; Figure 3 This is an overall block diagram of lithium battery health status prediction in some embodiments of the present invention.

[0057] According to some embodiments of the present invention, the present invention provides a method for predicting the health status of a lithium battery based on feature screening technology, comprising the following steps:

[0058] Acquisition of data sets: selecting lithium battery data in multiple different states to obtain data sets;

[0059] Extraction and preliminary processing of initial features: extract a set of initial features from the data set, and perform preliminary processing on the initial features to remove redundant and irrelevant features to generate a feature set;

[0060] Screening of feature sets: constructing an effective feature screening technology based on Fisher information matrix to screen feature sets and divide them into training sets and test sets;

[0061] Initial training of the machine learning model: input the training set into the machine learning model to predict the health status of lithium batteries, and select the best model and the best hyperparameter configuration through tuning and cross-validation of the machine learning model parameters to output preliminary prediction results;

[0062] Evaluation and optimization of machine learning models, using root mean square error and mean absolute error to evaluate preliminary prediction results, and optimizing hyperparameters and feature combinations in the training process based on the evaluation error to improve the prediction performance of machine learning models;

[0063] Prediction and verification of lithium battery health status: The optimized machine learning model is used to predict the health status of lithium batteries on the test set, and the prediction results are compared with those obtained by other feature screening methods to verify the accuracy and generalization ability of the lithium battery health status prediction results.

[0064] The present invention utilizes the effective feature screening technology based on the Fisher information matrix to indirectly reveal the complex relationship between features and target output by evaluating the uncertainty of parameter estimation. It can not only effectively capture the linear and nonlinear interactions between features, but also reflect the joint influence of each feature in the multi-parameter model, providing a more comprehensive and accurate basis for feature selection. In addition, the feature screening technology based on the Fisher information matrix is ​​suitable for linear and nonlinear models, especially when dealing with multi-parameter models with complex interactions, it shows strong robustness and accuracy.

[0065] According to some embodiments of the present invention, optionally, the lithium battery data includes charging current, voltage, temperature and discharging current, voltage, temperature, and the initial features include the battery voltage, temperature, current waveform during charging and discharging, and the number of cycles. Selecting appropriate lithium battery data as a data set is the basis for feature screening.

[0066] According to some embodiments of the present invention, optionally, the method for preliminary processing of initial features includes a statistical analysis method, a frequency domain feature extraction method, and a time domain feature extraction method. The statistical analysis method is to process the initial features to form statistical features, the frequency domain feature extraction method is to extract the battery frequency characteristics, and the time domain feature extraction method is to extract the battery time characteristics. The statistical analysis method includes but is not limited to mean, variance, and entropy, the frequency domain feature extraction method includes but is not limited to Fourier transform, and the time domain feature extraction method includes but is not limited to peak value and zero crossing point.

[0067] According to some embodiments of the present invention, optionally, the feature screening technology based on the Fisher information matrix includes Fisher information matrix construction, convex optimization problem solving and importance arrangement.

[0068] According to some embodiments of the present invention, optionally, the Fisher information matrix is ​​constructed by using the partial derivative matrix calculated by back propagation in the neural network model to construct an approximate Fisher information matrix, and the feature selection process is converted into an optimization problem. The specific steps are:

[0069] Input all the extracted features into the neural network for multiple trainings and save the training results;

[0070] The back propagation function is used to perform reverse transmission and pass it to the first layer of the neural network to obtain the sensitivity matrix, which is expressed as:

[0071] ,

[0072] Among them, L is the output of the first layer, is the corresponding parameter feature, is the symbol of partial derivative;

[0073] The average second-order statistics related to the influence of parameters on the output are obtained through the sensitivity matrix, and its expression is:

[0074] ,

[0075] Where n is the number of samples, Q is the measurement error covariance matrix, T is the transpose of the matrix, S is the sensitivity matrix of the output to the input features, and by calculating the product of S and the transpose of S, the average second-order statistics related to the impact of the features on the output can be obtained. The inverse matrix of F is used as an approximation of the error covariance matrix.

[0076] Convex optimization problem solving includes:

[0077] Consider feature selection as the optimal experimental design problem, select p features from n targets to retain, where n ≥ p, and the optimal experimental design problem expression is:

[0078] ,

[0079] in, represents the characteristic parameter, is the optimal feature selection set, Represents the characteristics The information matrix, is a row vector consisting of 1s in all entries, , is a non-negative weight factor that can be chosen to be 1 or 0, By all The collection composed of represents the scalar Fisher information matrix inverse function. argmin is used to represent the parameter value of a function that achieves the minimum value in the domain. For each data set, S is a positive definite matrix.

[0080] For designs involving a large number of weighting factors, the continuous optimization problem is converted into a convex optimization problem, expressed as:

[0081] ,

[0082] Among them, the weighted term It is defined as a continuous variable ranging from 0 to 1;

[0083] Applying the E-optimal design criterion to transform into an equivalent semidefinite programming problem to determine the key features is expressed as:

[0084]

[0085] in, is the minimum eigenvalue, I is the unit matrix, Indicates that for all ;

[0086] The importance ranking is obtained based on the equivalent semidefinite programming problem The value of Represents the importance of each feature obtained. Arrange in order of size to get the order of relevance. The features are retained for subsequent network training.

[0087] According to some embodiments of the present invention, optionally, the machine learning model includes a multi-layer perceptron, a convolutional neural network, or a graph neural network.

[0088] According to some embodiments of the present invention, optionally, in the configuration of the network, the multilayer perceptron consists of three layers, the convolutional neural network includes two convolutional layers, two pooling layers and one fully connected layer, the graph neural network consists of two graph convolutional layers and one fully connected layer, and the multilayer perceptron, convolutional neural network, and graph neural network all apply batch normalization and linear rectification function activation function, and set a random inactivation rate of 0.1 to prevent overfitting.

[0089] According to some embodiments of the present invention, optionally, 80% of the filtered feature set is used as a training set, and 20% of the filtered feature set is used as a test set.

[0090] According to some embodiments of the present invention, optionally, other feature screening methods include Spearman method, Pearson method, and Kendall method.

[0091] According to some embodiments of the present invention, optionally, it also includes: model updating and adaptive adjustment, through regular model updating, online learning or incremental learning methods, the model parameters are gradually adjusted to ensure that the prediction accuracy is continuously improved as the battery aging process progresses.

[0092] Example 1

[0093] Reference Figure 1 , Figure 3 This embodiment provides a method for predicting the health status of a lithium battery based on feature screening technology, comprising the following steps:

[0094] Acquisition of data sets: selecting lithium battery data in multiple different states to obtain data sets;

[0095] Extraction and preliminary processing of initial features: extract a set of initial features from the data set, and perform preliminary processing on the initial features to remove redundant and irrelevant features to generate a feature set;

[0096] Screening of feature sets: constructing an effective feature screening technology based on Fisher information matrix to screen feature sets and divide them into training sets and test sets;

[0097] Initial training of the machine learning model: input the training set into the machine learning model to predict the health status of lithium batteries, and select the best model and the best hyperparameter configuration through tuning and cross-validation of the machine learning model parameters to output preliminary prediction results;

[0098] Evaluation and optimization of machine learning models, using root mean square error and mean absolute error to evaluate preliminary prediction results, and optimizing hyperparameters and feature combinations in the training process based on the evaluation error to improve the prediction performance of machine learning models;

[0099] Prediction and verification of lithium battery health status: The optimized machine learning model is used to predict the health status of lithium batteries on the test set, and the prediction results are compared with those obtained by other feature screening methods to verify the accuracy and generalization ability of the lithium battery health status prediction results.

[0100] Specifically, lithium battery data includes charging current, voltage, temperature and discharging current, voltage, and temperature. The initial features include the battery voltage, temperature, current waveform during charging and discharging, and number of cycles. The initial features are preliminarily processed using statistical analysis methods, frequency domain feature extraction, and time domain feature extraction. Among them, the statistical analysis method is to process the initial features to form statistical features, including but not limited to mean, variance, and entropy. The frequency domain feature extraction method is to extract the frequency characteristics of the battery, including but not limited to Fourier transform. The time domain feature extraction method is to extract the time characteristics of the battery, including but not limited to peak value and zero crossing point.

[0101] Furthermore, the feature screening technology based on the Fisher information matrix includes the construction of the Fisher information matrix, the solution of the convex optimization problem and the importance arrangement. The feature screening technology is different from the traditional feature screening method based on statistical significance or correlation. It indirectly reveals the complex relationship between the features and the target output by evaluating the uncertainty of parameter estimation. The feature screening technology based on the Fisher information matrix can not only effectively capture the linear and nonlinear interactions between features, but also reflect the joint influence of each feature in the multi-parameter model, providing a more comprehensive and accurate basis for feature selection. In addition, the feature screening technology based on the Fisher information matrix is ​​suitable for linear and nonlinear models, especially when dealing with multi-parameter models with complex interactions, it shows strong robustness and accuracy.

[0102] Furthermore, the Fisher information matrix is ​​constructed by using the partial derivative matrix calculated by back propagation in the neural network model to construct an approximate Fisher information matrix, converting the feature selection process into an optimization problem. The specific steps are:

[0103] Input all the extracted features into the neural network for multiple trainings and save the training results;

[0104] The back propagation function is used to perform reverse transmission and pass it to the first layer of the neural network to obtain the sensitivity matrix, which is expressed as:

[0105] ,

[0106] Among them, L is the output of the first layer, is the corresponding parameter feature, is the symbol of partial derivative;

[0107] The average second-order statistics related to the influence of parameters on the output are obtained through the sensitivity matrix, and its expression is:

[0108] ,

[0109] Where n is the number of samples, Q is the measurement error covariance matrix, T is the transpose of the matrix, S is the sensitivity matrix of the output to the input features, and by calculating the product of S and the transpose of S, the average second-order statistics related to the impact of the features on the output can be obtained. The inverse matrix of F is used as an approximation of the error covariance matrix;

[0110] Convex optimization problem solving includes:

[0111] Consider feature selection as the optimal experimental design problem, select p features from n targets to retain, where n ≥ p, and the optimal experimental design problem expression is:

[0112] ,

[0113] in, represents the characteristic parameter, is the optimal feature selection set, Represents the characteristics The information matrix, is a row vector consisting of 1s in all entries, , is a non-negative weight factor that can be chosen to be 1 or 0, By all The collection composed of represents the scalar Fisher information matrix inverse function. argmin is used to represent the parameter value of a function that achieves the minimum value in the domain. For each data set, S is a positive definite matrix.

[0114] For designs involving a large number of weighting factors, the continuous optimization problem is converted into a convex optimization problem, expressed as:

[0115] ,

[0116] Among them, the weighted term It is defined as a continuous variable ranging from 0 to 1;

[0117] Applying the E-optimal design criterion to transform into an equivalent semidefinite programming problem to determine the key features is expressed as:

[0118] ,

[0119] in, is the minimum eigenvalue, I is the unit matrix, Indicates that for all ;

[0120] The importance ranking is obtained based on the equivalent semidefinite programming problem The value of Represents the importance of each feature obtained. Arrange in order of size to get the order of relevance. The features are retained for subsequent network training.

[0121] Furthermore, the machine learning model includes a multilayer perceptron, a convolutional neural network, and a graph neural network. In the network configuration, the multilayer perceptron consists of three layers, the convolutional neural network contains two convolutional layers, two pooling layers, and one fully connected layer, and the graph neural network consists of two graph convolutional layers and one fully connected layer. The multilayer perceptron, convolutional neural network, and graph neural network all apply batch normalization and linear rectification function activation function, and set a random dropout rate of 0.1 to prevent overfitting.

[0122] Furthermore, other feature screening methods including Spearman method, Pearson method and Kendall method were used.

[0123] Example 2

[0124] Reference Figure 2 This embodiment provides a lithium battery health status prediction method based on feature screening technology. The difference between this embodiment and embodiment 1 is that:

[0125] In practical applications, the working state of the battery may change, so the machine learning model needs to have a certain degree of adaptive ability. Therefore, the present invention provides a lithium battery health state prediction method based on feature screening technology, which also includes: model updating and adaptive adjustment. Through regular model updating, online learning or incremental learning methods, the model parameters are gradually adjusted to ensure that the prediction accuracy is continuously improved as the battery aging process progresses.

[0126] Example 3

[0127] Reference Figure 1-Figure 3 This embodiment provides a lithium battery health status prediction method based on feature screening technology. The difference between this embodiment and embodiment 1 is that:

[0128] In this embodiment, 80% of the filtered feature set is used as a training set, and 20% of the filtered feature set is used as a test set. The features obtained by the Spearman method, the Pearson method, and the Kendall method and the feature set obtained based on the Fisher information matrix feature screening technology are respectively input into the multi-layer perceptron, the convolutional neural network, and the graph neural network machine learning model for training. The results obtained are as follows:

[0129] ,

[0130]

[0131] Among them, MAE is the mean absolute error, RMSE is the root mean square error, SP, PE and KE are the Spearman method, Pearson method and Kendall method, FIFS is the feature screening technology based on the Fisher information matrix, NASA, UofM and MIT are lithium-ion battery charging and discharging data sets, MLP is a multi-layer perceptron, CNN is a convolutional neural network, and GNN is a graph neural network. The average absolute error and root mean square error obtained by using the Fisher information matrix feature screening technology as the correlation analysis are smaller than those of the Spearman method, Pearson method and Kendall method. The average absolute error and root mean square error obtained by the multi-layer perceptron are 0.39% and 0.49%; the average absolute error obtained by the convolutional neural network is 1.69% and the average root mean square error is 2.56%; the average absolute error on the graph neural network is 0.35% and the average root mean square error is 0.47%. Among the three networks used, the graph neural network has the best effect, and the lowest root mean square error can reach 0.11%.

[0132] It should be understood that the embodiments disclosed in the present invention are not limited to the specific processing steps or materials disclosed herein, but should be extended to equivalent substitutions of such features understood by ordinary technicians in the relevant field. It should also be understood that the terms used herein are only used for the purpose of describing specific embodiments and are not meant to be limiting.

[0133] The "embodiment" mentioned in the specification means that the specific features or characteristics described in conjunction with the embodiment are included in at least one embodiment of the present invention. Therefore, the phrases or "embodiment" appearing in various places throughout the specification do not necessarily refer to the same embodiment.

[0134] In addition, the described features or characteristics may be combined in one or more embodiments in any other suitable manner. In the above description, some specific details, such as thickness, quantity, etc., are provided to provide a comprehensive understanding of the embodiments of the present invention. However, those skilled in the relevant art will understand that the present invention can be implemented without one or more of the above specific details or can also be implemented using other methods, components, materials, etc.

Claims

1. A lithium battery health status prediction method based on feature screening technology, characterized in that: The following steps are involved: Acquisition of data sets: selecting lithium battery data in multiple different states to obtain data sets; Extraction and preliminary processing of initial features: extract a set of initial features from the data set, and perform preliminary processing on the initial features to remove redundant and irrelevant features to generate a feature set; Screening of feature sets: constructing an effective feature screening technology based on Fisher information matrix to screen feature sets and divide them into training sets and test sets; Initial training of the machine learning model: input the training set into the machine learning model to predict the health status of lithium batteries, and select the best model and the best hyperparameter configuration through tuning and cross-validation of the machine learning model parameters to output preliminary prediction results; Evaluation and optimization of machine learning models, using root mean square error and mean absolute error to evaluate preliminary prediction results, and optimizing hyperparameters and feature combinations in the training process based on the evaluation error to improve the prediction performance of machine learning models; Prediction and verification of lithium battery health status: The optimized machine learning model is used to predict the health status of lithium batteries on the test set, and the prediction results are compared with those obtained by other feature screening methods to verify the accuracy and generalization ability of the lithium battery health status prediction results.

2. The lithium battery health status prediction method based on feature screening technology according to claim 1 is characterized in that: The lithium battery data includes charging current, voltage, temperature and discharging current, voltage and temperature. The initial characteristics include the battery voltage, temperature and current waveform and number of cycles during charging and discharging.

3. The lithium battery health status prediction method based on feature screening technology according to claim 1 is characterized in that: The methods for preliminary processing of initial features include statistical analysis methods, frequency domain feature extraction methods and time domain feature extraction methods. The statistical analysis method processes the initial features to form statistical features, the frequency domain feature extraction method extracts the battery frequency characteristics, and the time domain feature extraction method extracts the battery time characteristics.

4. The method for predicting the health status of a lithium battery based on feature screening technology according to claim 1, characterized in that: The feature screening technology based on Fisher information matrix includes Fisher information matrix construction, convex optimization problem solving and importance ranking.

5. The method for predicting the health status of a lithium battery based on feature screening technology according to claim 4, characterized in that: The Fisher information matrix is ​​constructed by using the partial derivative matrix calculated by back propagation in the neural network model to construct an approximate Fisher information matrix, converting the feature selection process into an optimization problem. The specific steps are: Input all the extracted features into the neural network for multiple trainings and save the training results; The back propagation function is used to perform reverse transmission and pass it to the first layer of the neural network to obtain the sensitivity matrix, which is expressed as: ， Among them, L is the output of the first layer, is the corresponding parameter feature, is the symbol of partial derivative; The average second-order statistics related to the influence of parameters on the output are obtained through the sensitivity matrix, and its expression is: ， Where n is the number of samples, Q is the measurement error covariance matrix, T is the transpose of the matrix, S is the sensitivity matrix of the output to the input features, and the average second-order statistics related to the impact of the features on the output are obtained by calculating the product of S and the transpose of S. The inverse matrix of F is used as an approximation of the error covariance matrix. Convex optimization problem solving includes: Consider feature selection as the optimal experimental design problem, select p features from n targets to retain, where n ≥ p, and the optimal experimental design problem expression is: ， in, represents the characteristic parameter, is the optimal feature selection set, Represents the characteristics The information matrix, is a row vector consisting of 1s in all entries, , is a non-negative weight factor that chooses 1 or 0, By all The collection composed of represents the scalar Fisher information matrix inverse function. argmin is used to represent the parameter value of a function that achieves the minimum value in the domain. For each data set, S is a positive definite matrix. For designs involving a large number of weighting factors, the continuous optimization problem is converted into a convex optimization problem, expressed as: ， Among them, the weighted term It is defined as a continuous variable ranging from 0 to 1; Applying the E-optimal design criterion to transform into an equivalent semidefinite programming problem to determine the key features is expressed as: ， in, is the minimum eigenvalue, I is the unit matrix, Indicates that for all All conditions must be met; The importance ranking is obtained based on the equivalent semidefinite programming problem The value of Represents the importance of each feature obtained. Arrange in order of size to get the order of relevance. The features are retained for subsequent network training.

6. The lithium battery health status prediction method based on feature screening technology according to claim 1 is characterized in that: Machine learning models include multi-layer perceptron, convolutional neural network, and graph neural network.

7. The method for predicting the health status of a lithium battery based on feature screening technology according to claim 6, characterized in that: In the network configuration, the multilayer perceptron consists of three layers, the convolutional neural network contains two convolutional layers, two pooling layers and one fully connected layer, and the graph neural network consists of two graph convolutional layers and one fully connected layer. The multilayer perceptron, convolutional neural network, and graph neural network all apply batch normalization and linear rectification function activation functions, and set a random inactivation rate of 0.1 to prevent overfitting.

8. The method for predicting the health status of a lithium battery based on feature screening technology according to claim 1, characterized in that: 80% of the filtered feature set is used as the training set, and 20% of the filtered feature set is used as the test set.

9. The lithium battery health status prediction method based on feature screening technology according to claim 1, characterized in that: Other feature screening methods include Spearman method, Pearson method, and Kendall method.

10. The lithium battery health status prediction method based on feature screening technology according to claim 1, characterized in that: Also includes: Model update and adaptive adjustment: Through regular model updates, online learning or incremental learning methods, the model parameters are gradually adjusted to ensure that the prediction accuracy is continuously improved as the battery aging process progresses.

Images