Cross-modal interactive lithium battery SOH estimation method

Abstract

The invention discloses a cross-modal interactive lithium battery SOH estimation method. The method comprises the steps of obtaining a lithium battery experiment data set through an experiment; building a cross-modal interactive auto-encoder model, wherein the cross-modal interactive auto-encoder model comprises two encoder modules, two decoder modules and a cross attention mechanism module; a cross-modal interactive auto-encoder model is trained and tested, a loss function is designed, and the loss function comprises modal curve reconstruction loss, inter-modal global similarity constraint loss and inter-modal adjacent non-similarity constraint loss; building a KAN network model; and training and testing the KAN network model. Compared with the prior art, rich multi-modal data is utilized, a cross-modal interactive auto-encoder is established to fuse and reconstruct multi-modal data, health features are extracted without labels, and more comprehensive battery state information is provided compared with single-modal voltage and current data and multi-modal data, so that the method has the advantages of high reliability and high reliability. And the lithium battery SOH estimation precision is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of lithium batteries, and in particular to a cross-modal interactive lithium battery SOH estimation method. Background Art

[0002] Lithium battery energy storage plays a key role in improving new energy systems such as renewable energy and energy transformation. As the battery usage time increases, the battery capacity and performance will gradually decrease, which will lead to battery failures such as overheating and overcharging. Therefore, it is of great significance to fully and accurately estimate the health state (SOH) of lithium batteries.

[0003] In recent years, the method of estimating the health state of lithium batteries using data-driven methods has been widely studied. Most researchers use time domain data such as voltage, current and temperature as a single mode to estimate SOH, or use electrochemical impedance spectroscopy as a single mode for estimation. However, data based on a single mode can only provide partial information about the battery state and cannot accurately capture all aspects of battery performance changes. It is difficult to accurately reflect the SOH state of lithium batteries. For example, voltage and current data cannot reflect the chemical changes inside the battery, and although electrochemical impedance spectroscopy data can reveal electrochemical processes, it may lack information about the performance of the battery under actual working conditions.

[0004] Currently, research on multimodality is still in its infancy. Existing technologies generally rely on simple splicing of labeled data and lack deep fusion, which results in the fused data being unable to fully reflect the complex state of the battery, and is not conducive to improving the accuracy of SOH estimation. Summary of the invention

[0005] In order to overcome the above-mentioned shortcomings of the existing multi-modal based SOH prediction method, the present invention proposes a cross-modal interactive lithium battery SOH estimation method.

[0006] The technical solution of the present invention is as follows:

[0007] S1: Construct a lithium battery data set through experiments, that is, conduct cyclic charge and discharge experiments on T lithium batteries of the same model, collect and record the relaxation voltage curve of the battery during the static period after each charge, collect and record the battery electrochemical impedance spectrum curve in the frequency range of 0.01Hz-100kHz after each discharge, and calculate the battery SOH value of the corresponding charge and discharge cycle according to the total discharge amount. After the experiment, the relaxation voltage curve set U, electrochemical impedance spectrum curve set E and battery SOH set obtained by all batteries in all cycles are collected to construct the data set A = {U, E, SOH};

[0008] S2: Build a cross-modal interactive autoencoder model, which includes two encoder modules, two decoder modules and a cross-attention mechanism module. The specific steps include: building a relaxation voltage modal encoder and an electrochemical impedance spectroscopy modal encoder to extract low-dimensional potential features Z of the relaxation voltage modality and the electrochemical impedance spectroscopy modality, respectively. U and Z E ; Build a cross-attention mechanism module for low-dimensional latent feature fusion and decoupling, which is used to U and Z E Generating fusion modality latent features and Build relaxation voltage mode decoder and electrochemical impedance spectroscopy mode decoder, respectively, to fusion mode potential features and To reconstruct the curve U rf and E rf ;

[0009] S3: training and testing a cross-modal interactive autoencoder model, the specific steps include: designing a model loss function, wherein the model loss function includes a modal curve reconstruction loss, a global similarity constraint loss between modalities, and an adjacent non-similarity constraint loss between modalities; constructing a training set {U Train ,E Train} and test set U Test ,E Test}; The cross-modal interactive autoencoder model is trained and tested using the training set and the test set, during which the model loss function is used to calculate the loss value, and the network parameters in the cross-modal interactive autoencoder model are iteratively updated by the optimizer to continuously reduce the loss value until the loss value of the test set meets the threshold requirement, then the training is terminated and the model is saved;

[0010] S4: Building a KAN network model. The specific steps include: building multiple KAN network layers, each KAN network layer is composed of a set of learnable B-spline functions;

[0011] S5: Training and testing the KAN network model. The specific steps include: first, inputting the dataset A = {U, E, SOH} obtained in step S1 into the cross-modal interactive autoencoder model trained in step S3, and outputting the fusion modality potential features. The feature set HF and the battery SOH set are used to construct a feature data set Z = {HF, SOH}, and the training set {HF Train ,SOH Train} and the test set {HF Test ,SOH Test}; Then, the training set {HF Train ,SOH Train} in the feature set HF Train Sort by feature importance and add them to the optimal feature set one by one Input into the KAN network model built in step S4, train the model, use K-fold cross validation to comprehensively evaluate the contribution of each feature to the model performance, and select the optimal feature set Finally, save the optimal feature set The trained KAN network model is used for the test set HF Test Also select a set with the same features as the training set And the optimal feature set for the test set Perform SOH estimation.

[0012] In this solution, step S2 is to build a cross-modal interactive autoencoder model, specifically:

[0013] S201: Build a relaxation voltage modal encoder with L fully connected layers, and convert the relaxation voltage curve training set U Train The input is sent to the first layer encoder for encoding. The output of each layer of encoding is used as the input of the next layer of encoder. The last layer of encoder outputs the potential feature data of the relaxation voltage, as follows:

[0014]

[0015] Among them, U Train represents the normalized relaxation voltage curve, represents the characteristics of the first fully connected layer network output in the relaxation voltage modality encoder, represents the output feature of the lth fully connected layer network in the relaxation voltage modality encoder, Z U represents the potential characteristics of the relaxation voltage output after passing through the L-layer encoder network, is the activation function of the lth fully connected layer in the relaxed voltage modality encoder, and are the weight and bias matrices of the lth fully connected layer in the relaxed voltage modality encoder;

[0016] S202: Build an electrochemical impedance spectroscopy modal encoder containing L fully connected layers, and convert the electrochemical impedance spectroscopy curve training set E Train The input is sent to the first layer of encoder for encoding. The output of each layer of encoding is used as the input of the next layer of encoder. The last layer of encoder outputs the potential characteristic data of the electrochemical impedance spectroscopy, as follows:

[0017]

[0018] Among them, E Train represents the normalized electrochemical impedance spectroscopy curve, Represents the characteristics of the output of the first fully connected layer network in the electrochemical impedance spectroscopy modal encoder network, represents the output feature of the lth fully connected layer network in the electrochemical impedance spectroscopy modal encoder network, Z E represents the potential features output after passing through the L-layer encoder network, is the activation function of the lth fully connected layer in the electrochemical impedance spectroscopy modal encoder network, and is the weight and bias matrix in the lth fully connected layer in the electrochemical impedance spectroscopy modal encoder network layer;

[0019] S203: Construct a cross-attention mechanism module to convert the relaxation voltage potential feature Z output in step 201 into U and the electrochemical impedance spectroscopy potential feature Z outputted in step 202 E , which is converted into query, key, and value vectors corresponding to relaxation voltage and electrochemical impedance spectroscopy potential features through learnable parameters, as shown in the following formula:

[0020]

[0021] in, is the weight parameter of the potential feature, Q U , K U 、V U is the query,key,value corresponding to the relaxation voltage potential feature, Q E , K E 、V E is the query, key, and value corresponding to the potential features of electrochemical impedance spectroscopy;

[0022] S204: Input the query, key, and value of the relaxation voltage and electrochemical impedance spectrum obtained in step 203 into the attention mechanism layer, and output two potential features decoupled after fusing the two modalities, as shown in the following formula:

[0023]

[0024] in, for The potential feature representation after integrating electrochemical impedance spectroscopy modalities, for The latent feature representation after fusion of relaxation voltage modes, softmax(·) represents the softmax function;

[0025] S205: Build an electrochemical impedance spectroscopy modal decoder of the P-layer fully connected layer, and convert the decoupled fusion modal potential features obtained in step 204 into As the input of the first layer decoder, it is input into the electrochemical impedance spectroscopy modal decoder. The output of each layer of decoding is used as the input of the next layer of decoder. The output of the last layer of the decoder is the reconstructed electrochemical impedance spectroscopy curve, which is as follows:

[0026]

[0027] Among them, E rf For E Train The curve reconstructed by the electrochemical impedance spectroscopy decoder, is the feature output by the pth fully connected layer network in the electrochemical impedance spectroscopy decoder network, is the activation function of the pth fully connected layer in the electrochemical impedance spectroscopy decoding layer, and is the weight and bias matrix in the pth fully connected layer in the electrochemical impedance spectroscopy decoder;

[0028] S206: Build a relaxed voltage modal decoder of the P-layer fully connected layer, and transform the decoupled fusion modal potential features obtained in step 204 into As the input of the first layer decoder, and input to the relaxation voltage modal decoder, the output of each layer of decoding is used as the input of the next layer of decoder, and the output of the last layer of the decoder is the reconstructed relaxation voltage curve, which is as follows:

[0029]

[0030] Among them, U rf For U Train The curve reconstructed by the relaxation voltage decoder, is the feature output of the pth fully connected layer network in the relaxed voltage decoder network, is the activation function of the pth fully connected layer in the relaxed voltage decoder, and are the weight and bias matrices in the pth fully connected layer in the relaxed voltage decoder.

[0031] In this solution, the training and testing of the cross-modal interactive autoencoder model described in step S3 specifically includes:

[0032] S301: Setting and initializing training parameters of a cross-modal interactive autoencoder model;

[0033] S302: Data preprocessing: normalizing the relaxation voltage curve set U and the electrochemical impedance spectrum curve set E in the data set A obtained in step S1 to obtain a normalized set and construct a training set {U Train ,E Train} and the test set {U Test ,E Test};

[0034] S303: Design a model loss function, which includes a modal curve reconstruction loss, an inter-modal global similarity constraint loss, and an inter-modal adjacent non-similarity constraint loss, and is specifically as follows:

[0035]

[0036] in, is the sum of all loss functions of the entire cross-modal interactive autoencoder model, is the sum of the loss functions for the relaxation voltage curve and the electrochemical impedance spectroscopy curve reconstruction, The characteristic of the relaxation voltage curve is Characteristics of electrochemical impedance spectroscopy curves The similarity constraint loss function between The characteristic of the relaxation voltage curve is Characteristics of EIS curves The adjacent non-similarity constraint loss function between , β and δ are The weight parameter of

[0037] The relaxation voltage curve and electrochemical impedance spectroscopy curve reconstruct the loss function It consists of two parts, including measuring the relaxation voltage curve U after model reconstruction rf Compared with the original relaxation voltage curve U Train The loss function of the similarity between them and the electrochemical impedance spectroscopy curve E after the model reconstruction is measured rf Compared with the original electrochemical impedance spectroscopy curve E Train The loss function of the similarity between is as follows:

[0038]

[0039] in, Represents the U Train The voltage value of the jth sampling point of the relaxation voltage curve of the i-th sample in is, Represents E Train The impedance value of the jth sampling point of the electrochemical impedance spectroscopy curve of the i-th sample, N U is the length of the relaxation voltage curve, N E is the length of the electrochemical impedance spectroscopy curve, N is the length of the electrochemical impedance spectroscopy curve;

[0040] The similarity constraint loss function between the relaxation voltage curve and the electrochemical impedance spectroscopy curve Includes the features of each output layer of the encoder of both modes With features The maximum multi-core mean difference between them is calculated by using a combination of Gaussian kernel functions with different bandwidth parameters, as shown in the following formula:

[0041]

[0042] Where K is the number of kernel functions, λ k is the weight of the kth kernel function, which needs to satisfy Refers to using the kth kernel function to calculate the feature and The maximum mean difference between two feature distributions is expanded using the Gaussian kernel function as follows:

[0043]

[0044] Among them, ||·|| is the straight-line distance between two points in Euclidean space, δ k is the bandwidth parameter of the kth kernel function, which is used to control the width of the Gaussian kernel. Output feature of the first layer encoder of the impedance spectrum The data of the i-th dimension;

[0045] The adjacent non-similarity constraint loss function between the relaxation voltage curve and the electrochemical impedance spectroscopy curve modes Includes features of the fully connected output layer of the relaxed voltage modulus encoder The output features of adjacent layers of the corresponding electrochemical impedance spectroscopy modal encoder The inverse of the cosine similarity loss function between the two, the output characteristics of the electrochemical impedance spectroscopy modal encoder The features of the adjacent output layer of the corresponding relaxed voltage modality encoder The opposite number of the cosine similarity loss function between is as follows:

[0046]

[0047] S304: Model training and testing, i.e., using the training set {U Train ,E Train}Train the cross-modal interactive autoencoder model described in step S2, using the test set {U Test ,E Test}Test the trained cross-modal interactive autoencoder model to verify the reconstruction accuracy of the model. If it does not meet the requirements, retrain until the obtained model meets the accuracy requirements. Save the model and use it as the cross-modal interactive autoencoder model.

[0048] In this solution, step S4 of building a KAN network model specifically includes:

[0049] S401: Construct a regression algorithm model based on the KAN network, define a KAN network structure with q layers, and use a learnable B-spline function in the activation function, as shown in the following formula:

[0050] f(HF)=Φ q ·(Φ q-1 ·(…(Φ 2 (Φ 1 ·HF)))) (20)

[0051] Among them, Φ q is the learnable activation function matrix of the qth hidden layer unit, HF is the fusion modality potential feature extracted by the cross-modal interactive autoencoder model;

[0052] The Φ q The activation function matrix can be learned. This matrix is ​​a one-dimensional function representation matrix of the hidden layer nodes of the previous KAN network layer. The specific formula is as follows:

[0053]

[0054] in, Indicates the nth q+1 The input neurons and the nth neurons in the qth layer q The transformation function between the output neurons;

[0055] Said The transformation function is parameterized by the B-spline function, which has learnable parameters and can be optimized through training data. In the KAN network, the B-spline function is used to approximate the learnable activation function that constitutes the edge of the network, as shown in the following formula:

[0056]

[0057] Among them, B i (x) is the B-spline basis function, c i is the B-spline basis function B i The coefficients of (·) can be obtained through network training.

[0058] In this solution, the training and testing of the KAN network model in step S5 specifically includes:

[0059] S501: Input the dataset A = {U, E, SOH} obtained in step S1 into the cross-modal interactive autoencoder model trained in step S3, and output the fusion modality potential feature set The feature set HF and the battery SOH set are used to construct a feature data set Z = {HF, SOH}, and the training set HF is divided into Train ,SOH Train} and the test set {HF Test ,SOH Test};

[0060] S502: The training data set with n potential features of fused relaxation voltage mode and electrochemical impedance spectroscopy mode obtained in step S501 is Each individual feature Combined with the SOH value, n single feature data sets are constructed, as shown in the following formula:

[0061]

[0062] S503: The KAN network model described in the training step S4 is constructed, and the root mean square error of the cross-validation prediction results based on the model training is used as a single feature Evaluation index for HF Train n single features in Rank the order of importance;

[0063] S504: Construct an empty optimal feature set According to the feature data set HF obtained in step S501 Train Get the remaining feature data set

[0064] S505: Feature data set According to the feature importance ranking obtained in S503, feature x is selected one by one from large to small and added to the feature data set Building a dataset Train the regression prediction model based on the KAN network, and use the estimated root mean square error obtained by K-fold cross validation of the model on sample S as the set The evaluation indicators are as follows:

[0065]

[0066] Among them, x best is the feature dataset in the current iteration The best features to choose; is a feature set SOH estimates in the KAN regression model;

[0067] S506: The optimal feature x obtained in step S505 is best Add to current feature set and remove it from the feature set

[0068] S507: In order to balance the complexity of the model and improve the performance of the model to achieve better prediction results, the maximum number of features and performance improvement are set as the stop conditions of the feature subset search process. Whether the number of features and cross-validation performance obtained in steps S505 and S506 reach the improvement target is used as the condition for stopping iteration. If both conditions are met at the same time, the selection of feature subsets is stopped. If one of them is not met, the optimal subset search of steps S505 and S506 is continued. The specific condition formula is as follows:

[0069]

[0070] S508: Save the optimal feature set obtained in step S507 The trained KAN network model is used for the test set HF Test Also select a set with the same features as the training set And the optimal feature set for the test set Perform SOH estimation.

[0071] Compared with the prior art, the technical solution of the present invention has the following beneficial effects:

[0072] The present invention uses the battery relaxation voltage curve mode and the electrochemical impedance spectrum curve mode as input modes, making full use of the complementarity of the two modes to provide a comprehensive evaluation for the battery's health status and aging mechanism analysis. By establishing a cross-modal interactive autoencoder model, the model can deeply fuse, align and reconstruct multimodal data without battery health status labels, thereby extracting multimodal features related to the battery health status. The above two advantages are both conducive to improving the accuracy of lithium battery SOH estimation. BRIEF DESCRIPTION OF THE DRAWINGS

[0073] In order to more clearly illustrate the technical solutions of the exemplary embodiments of the present invention, the drawings required for use in the embodiments are briefly introduced below. It should be understood that the following drawings only show certain embodiments of the present invention and therefore should not be regarded as limiting the scope. For ordinary technicians in this field, other related drawings can be obtained based on these drawings without creative work.

[0074] Figure 1 A schematic diagram of a cross-modal interactive lithium battery SOH estimation method provided in an embodiment of the present invention.

[0075] Figure 2 This is the relaxation voltage curve of the lithium battery after charging.

[0076] Figure 3 This is the electrochemical impedance spectroscopy curve collected after the lithium battery is discharged. DETAILED DESCRIPTION

[0077] In order to more clearly understand the above-mentioned objectives, features and advantages of the present invention, the present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be noted that the embodiments of the present application and the features in the embodiments can be combined with each other without conflict.

[0078] In the following description, many specific details are set forth to facilitate a full understanding of the present invention. However, the present invention may also be implemented in other ways different from those described herein. Therefore, the protection scope of the present invention is not limited to the specific embodiments disclosed below.

[0079] For easier understanding, see Figure 1 The present invention provides an embodiment of a cross-modal interactive lithium battery SOH estimation method, comprising:

[0080] S1: Construct a lithium battery data set through experiments, that is, conduct cyclic charge and discharge experiments on T lithium batteries of the same model, collect and record the relaxation voltage curve of the battery during the static period after each charge, collect and record the battery electrochemical impedance spectrum curve in the frequency range of 0.01Hz-100kHz after each discharge, and calculate the battery SOH value of the corresponding charge and discharge cycle according to the total discharge amount. After the experiment, the relaxation voltage curve set U, electrochemical impedance spectrum curve set E and battery SOH set obtained by all batteries in all cycles are constructed to construct a data set A = {U, E, SOH}, where the relaxation voltage curve is as follows Figure 2 The electrochemical impedance spectroscopy curve is shown in Figure 3 As shown;

[0081] S2: Build a cross-modal interactive autoencoder model, which includes two encoder modules, two decoder modules and a cross-attention mechanism module. The specific steps include: building a relaxation voltage modal encoder and an electrochemical impedance spectroscopy modal encoder to extract low-dimensional potential features Z of the relaxation voltage modality and the electrochemical impedance spectroscopy modality, respectively. U and Z E ; Build a cross-attention mechanism module for low-dimensional latent feature fusion and decoupling, which is used to U and Z E Generating fusion modality latent features and Build relaxation voltage mode decoder and electrochemical impedance spectroscopy mode decoder, respectively, to fusion mode potential features and To reconstruct the curve U rf and E rf ;

[0082] S3: training and testing a cross-modal interactive autoencoder model, the specific steps include: designing a model loss function, wherein the model loss function includes a modal curve reconstruction loss, a global similarity constraint loss between modalities, and an adjacent non-similarity constraint loss between modalities; constructing a training set {U Train ,E Train} and test set U Test ,E Test}; The cross-modal interactive autoencoder model is trained and tested using the training set and the test set, during which the model loss function is used to calculate the loss value, and the network parameters in the cross-modal interactive autoencoder model are iteratively updated by the optimizer to continuously reduce the loss value until the loss value of the test set meets the threshold requirement, then the training is terminated and the model is saved;

[0083] S4: Building a KAN network model. The specific steps include: building multiple KAN network layers, each KAN network layer is composed of a set of learnable B-spline functions;

[0084] S5: Training and testing the KAN network model. The specific steps include: first, inputting the dataset A = {U, E, SOH} obtained in step S1 into the cross-modal interactive autoencoder model trained in step S3, and outputting the fusion modality potential features. The feature set HF and the battery SOH set are used to construct a feature data set Z = {HF, SOH}, and the training set {HF Train ,SOH Train} and the test set {HF Test ,SOH Test}; Then, the training set {HF Train ,SOH Train} in the feature set HF Train Sort by feature importance and add them to the optimal feature set one by one Input into the KAN network model built in step S4, train the model, use K-fold cross validation to comprehensively evaluate the contribution of each feature to the model performance, and select the optimal feature set Finally, save the optimal feature set The trained KAN network model is used for the test set HF Test Also select a set with the same features as the training set And the optimal feature set for the test set Perform SOH estimation.

[0085] In this solution, step S2 is to build a cross-modal interactive autoencoder model, specifically:

[0086] S201: Build a relaxation voltage modal encoder with L fully connected layers, and convert the relaxation voltage curve training set U Train The input is sent to the first layer encoder for encoding. The output of each layer of encoding is used as the input of the next layer of encoder. The last layer of encoder outputs the potential feature data of the relaxation voltage, as follows:

[0087]

[0088] Among them, U Train represents the normalized relaxation voltage curve, represents the characteristics of the first fully connected layer network output in the relaxation voltage modality encoder, represents the output feature of the lth fully connected layer network in the relaxation voltage modality encoder, Z U represents the potential characteristics of the relaxation voltage output after passing through the L-layer encoder network, is the activation function of the lth fully connected layer in the relaxed voltage modality encoder, and are the weight and bias matrices of the lth fully connected layer in the relaxed voltage modality encoder;

[0089] S202: Build an electrochemical impedance spectroscopy modal encoder containing L fully connected layers, and convert the electrochemical impedance spectroscopy curve training set E Train The input is sent to the first layer of encoder for encoding. The output of each layer of encoding is used as the input of the next layer of encoder. The last layer of encoder outputs the potential characteristic data of the electrochemical impedance spectroscopy, as follows:

[0090]

[0091] Among them, E Train represents the normalized electrochemical impedance spectroscopy curve, Represents the characteristics of the output of the first fully connected layer network in the electrochemical impedance spectroscopy modal encoder network, represents the output feature of the lth fully connected layer network in the electrochemical impedance spectroscopy modal encoder network, Z E represents the potential features output after passing through the L-layer encoder network, is the activation function of the lth fully connected layer in the electrochemical impedance spectroscopy modal encoder network, and is the weight and bias matrix in the lth fully connected layer in the electrochemical impedance spectroscopy modal encoder network layer;

[0092] S203: Construct a cross-attention mechanism module to convert the relaxation voltage potential feature Z output in step 201 into U and the electrochemical impedance spectroscopy potential feature Z outputted in step 202 E, which is converted into query, key, and value vectors corresponding to relaxation voltage and electrochemical impedance spectroscopy potential features through learnable parameters, as shown in the following formula:

[0093]

[0094] in, is the weight parameter of the potential feature, Q U , K U 、V U is the query,key,value corresponding to the relaxation voltage potential feature, Q E , K E 、V E is the query, key, and value corresponding to the potential features of electrochemical impedance spectroscopy;

[0095] S204: Input the query, key, and value of the relaxation voltage and electrochemical impedance spectrum obtained in step 203 into the attention mechanism layer, and output two potential features decoupled after fusing the two modalities, as shown in the following formula:

[0096]

[0097] in, for The potential feature representation after integrating electrochemical impedance spectroscopy modalities, for The latent feature representation after fusion of relaxation voltage modes, softmax(·) represents the softmax function;

[0098] S205: Build an electrochemical impedance spectroscopy modal decoder of the P-layer fully connected layer, and convert the decoupled fusion modal potential features obtained in step 204 into As the input of the first layer decoder, it is input into the electrochemical impedance spectroscopy modal decoder. The output of each layer of decoding is used as the input of the next layer of decoder. The output of the last layer of the decoder is the reconstructed electrochemical impedance spectroscopy curve, which is as follows:

[0099]

[0100] Among them, E rf For E Train The curve reconstructed by the electrochemical impedance spectroscopy decoder, is the feature output by the pth fully connected layer network in the electrochemical impedance spectroscopy decoder network, is the activation function of the pth fully connected layer in the electrochemical impedance spectroscopy decoding layer, and is the weight and bias matrix in the pth fully connected layer in the electrochemical impedance spectroscopy decoder;

[0101] S206: Build a relaxed voltage modal decoder of the P-layer fully connected layer, and transform the decoupled fusion modal potential features obtained in step 204 into As the input of the first layer decoder, and input to the relaxation voltage modal decoder, the output of each layer of decoding is used as the input of the next layer of decoder, and the output of the last layer of the decoder is the reconstructed relaxation voltage curve, which is as follows:

[0102]

[0103] Among them, U rf For U Train The curve reconstructed by the relaxation voltage decoder, is the feature output of the pth fully connected layer network in the relaxed voltage decoder network, is the activation function of the pth fully connected layer in the relaxed voltage decoder, and is the weight and bias matrix in the pth fully connected layer in the relaxed voltage decoder;

[0104] In this solution, the training and testing of the cross-modal interactive autoencoder model in step S3 is specifically as follows:

[0105] S301: Setting and initializing training parameters of a cross-modal interactive autoencoder model;

[0106] S302: Data preprocessing: normalizing the relaxation voltage curve set U and the electrochemical impedance spectrum curve set E in the data set A obtained in step S1 to obtain a normalized set and construct a training set {U Train ,E Train} and the test set {U Test ,E Test};

[0107] S303: Design a model loss function, which includes a modal curve reconstruction loss, an inter-modal global similarity constraint loss, and an inter-modal adjacent non-similarity constraint loss, and is specifically as follows:

[0108]

[0109] in, is the sum of all loss functions of the entire cross-modal interactive autoencoder model, is the sum of the loss functions for the relaxation voltage curve and the electrochemical impedance spectroscopy curve reconstruction, The characteristic of the relaxation voltage curve is Characteristics of electrochemical impedance spectroscopy curves The similarity constraint loss function between The characteristic of the relaxation voltage curve is Characteristics of EIS curves The adjacent non-similarity constraint loss function between , β and δ are The weight parameter of

[0110] The relaxation voltage curve and electrochemical impedance spectroscopy curve reconstruct the loss function It consists of two parts, including measuring the relaxation voltage curve U after model reconstruction rf Compared with the original relaxation voltage curve U Train The loss function of the similarity between them and the electrochemical impedance spectroscopy curve E after the model reconstruction is measured rf Compared with the original electrochemical impedance spectroscopy curve E Train The loss function of the similarity between is as follows:

[0111]

[0112] in, Represents the U Train The voltage value of the jth sampling point of the relaxation voltage curve of the i-th sample in is, Represents E Train The impedance value of the jth sampling point of the electrochemical impedance spectroscopy curve of the i-th sample, N U is the length of the relaxation voltage curve, N E is the length of the electrochemical impedance spectroscopy curve, N is the length of the electrochemical impedance spectroscopy curve;

[0113] The similarity constraint loss function between the relaxation voltage curve and the electrochemical impedance spectroscopy curve Includes the features of each output layer of the encoder of both modes With features The maximum multi-core mean difference between them is calculated by using a combination of Gaussian kernel functions with different bandwidth parameters, as shown in the following formula:

[0114]

[0115] Where K is the number of kernel functions, λ k is the weight of the kth kernel function, which needs to satisfy Refers to using the kth kernel function to calculate the feature and The maximum mean difference between two feature distributions is expanded using the Gaussian kernel function as follows:

[0116]

[0117] Among them, ||·|| is the straight-line distance between two points in Euclidean space, δ kis the bandwidth parameter of the kth kernel function, which is used to control the width of the Gaussian kernel. Output feature of the first layer encoder of the impedance spectrum The data of the i-th dimension;

[0118] The adjacent non-similarity constraint loss function between the relaxation voltage curve and the electrochemical impedance spectroscopy curve modes Includes features of the fully connected output layer of the relaxed voltage modulus encoder The output features of adjacent layers of the corresponding electrochemical impedance spectroscopy modal encoder The inverse of the cosine similarity loss function between the two, the output characteristics of the electrochemical impedance spectroscopy modal encoder The features of the adjacent output layer of the corresponding relaxed voltage modality encoder The opposite number of the cosine similarity loss function between is as follows:

[0119]

[0120] S304: Model training and testing, i.e., using the training set {U Train ,E Train}Train the cross-modal interactive autoencoder model described in step S2, using the test set {U Test ,E Test}Test the trained cross-modal interactive autoencoder model to verify the reconstruction accuracy of the model. If it does not meet the requirements, retrain until the obtained model meets the accuracy requirements. Save the model and use it as the cross-modal interactive autoencoder model.

[0121] In this solution, step S4 of building a KAN network model is as follows:

[0122] S401: Construct a regression algorithm model based on the KAN network, define a KAN network structure with q layers, and use a learnable B-spline function in the activation function, as shown in the following formula:

[0123] f(HF)=Φ q ·(Φ q-1 ·(…(Φ 2 (Φ 1 ·HF)))) (20)

[0124] Among them, Φ q is the learnable activation function matrix of the qth hidden layer unit, HF is the fusion modality potential feature extracted by the cross-modal interactive autoencoder model;

[0125] The Φ qThe activation function matrix can be learned. This matrix is ​​a one-dimensional function representation matrix of the hidden layer nodes of the previous KAN network layer. The specific formula is as follows:

[0126]

[0127] in, Indicates the nth q+1 The input neurons and the nth neurons in the qth layer q The transformation function between the output neurons;

[0128] Said The transformation function is parameterized by the B-spline function, which has learnable parameters and can be optimized through training data. In the KAN network, the B-spline function is used to approximate the learnable activation function that constitutes the edge of the network, as shown in the following formula:

[0129]

[0130] Among them, B i (x) is the B-spline basis function, c i is the B-spline basis function B i The coefficients of (·) can be obtained through network training.

[0131] In this solution, the training and testing of the KAN network model in step S5 is specifically as follows:

[0132] S501: Input the dataset A = {U, E, SOH} obtained in step S1 into the cross-modal interactive autoencoder model trained in step S3, and output the fusion modality potential feature set The feature set HF and the battery SOH set are used to construct a feature data set Z = {HF, SOH}, and the training set HF is divided into Train ,SOH Train} and the test set {HF Test ,SOH Test};

[0133] S502: The training data set with n potential features of fused relaxation voltage mode and electrochemical impedance spectroscopy mode obtained in step S501 is Each individual feature Combined with the SOH value, n single feature data sets are constructed, as shown in the following formula:

[0134]

[0135] S503: The KAN network model described in the training step S4 is constructed, and the root mean square error of the cross-validation prediction results based on the model training is used as a single feature Evaluation index for HFTrain n single features in Rank the order of importance;

[0136] S504: Construct an empty optimal feature set According to the feature data set HF obtained in step S501 Train Get the remaining feature data set

[0137] S505: Feature data set According to the feature importance ranking obtained in S503, feature x is selected one by one from large to small and added to the feature data set Building a dataset Train the regression prediction model based on the KAN network, and use the estimated root mean square error obtained by K-fold cross validation of the model on sample S as the set The evaluation indicators are as follows:

[0138]

[0139] Among them, x best is the feature dataset in the current iteration The best features to choose; is a feature set SOH estimates in the KAN regression model;

[0140] S506: The optimal feature x obtained in step S505 is best Add to current feature set and remove it from the feature set

[0141] S507: In order to balance the complexity of the model and improve the performance of the model to achieve better prediction results, the maximum number of features and performance improvement are set as the stop conditions of the feature subset search process. Whether the number of features and cross-validation performance obtained in steps S505 and S506 reach the improvement target is used as the condition for stopping iteration. If both conditions are met at the same time, the selection of feature subsets is stopped. If one of them is not met, the optimal subset search of steps S505 and S506 is continued. The specific condition formula is as follows:

[0142]

[0143] S508: Save the optimal feature set obtained in step S507 The trained KAN network model is used for the test set HF Test Also select a set with the same features as the training set And the optimal feature set for the test set Perform SOH estimation.

Claims

1. A cross-modal interactive lithium battery SOH estimation method, comprising a cross-modal interactive autoencoder model and a KAN network model, characterized in that: The specific steps are as follows: S1: Construct a lithium battery data set through experiments, that is, conduct cyclic charge and discharge experiments on T lithium batteries of the same model, collect and record the relaxation voltage curve of the battery during the static period after each charge, collect and record the battery electrochemical impedance spectrum curve in the frequency range of 0.01Hz-100kHz after each discharge, and calculate the battery SOH value of the corresponding charge and discharge cycle according to the total discharge amount. After the experiment, the relaxation voltage curve set U, electrochemical impedance spectrum curve set E and battery SOH set obtained by all batteries in all cycles are collected to construct the data set A = {U, E, SOH}; S2: Build a cross-modal interactive autoencoder model, which includes two encoder modules, two decoder modules and a cross-attention mechanism module. The specific steps include: building a relaxation voltage modal encoder and an electrochemical impedance spectroscopy modal encoder to extract low-dimensional potential features Z of the relaxation voltage modality and the electrochemical impedance spectroscopy modality, respectively. U and Z E ; Build a cross-attention mechanism module for low-dimensional latent feature fusion and decoupling, which is used to U and Z E Generating fusion modality latent features and Build relaxation voltage mode decoder and electrochemical impedance spectroscopy mode decoder, respectively, to fusion mode potential features and To reconstruct the curve U rf and E rf ; S3: training and testing a cross-modal interactive autoencoder model, the specific steps include: designing a model loss function, wherein the model loss function includes a modal curve reconstruction loss, a global similarity constraint loss between modalities, and an adjacent non-similarity constraint loss between modalities; constructing a training set {U Train ,E Train } and the test set {U Test ,E Test }; The cross-modal interactive autoencoder model is trained and tested using the training set and the test set, during which the model loss function is used to calculate the loss value, and the network parameters in the cross-modal interactive autoencoder model are iteratively updated by the optimizer to continuously reduce the loss value until the loss value of the test set meets the threshold requirement, then the training is terminated and the model is saved; S4: Building a KAN network model. The specific steps include: building multiple KAN network layers, each KAN network layer is composed of a set of learnable B-spline functions; S5: Training and testing the KAN network model. The specific steps include: first, inputting the dataset A = {U, E, SOH} obtained in step S1 into the cross-modal interactive autoencoder model trained in step S3, and outputting the fusion modality potential features. The feature set HF and the battery SOH set are used to construct a feature data set Z = {HF, SOH}, and the training set {HF Train ,SOH Train } and the test set {HF Test ,SOH Test }; Then, the training set {HF Train ,SOH Train } in the feature set HF Train Sort by feature importance and add them to the optimal feature set one by one Input into the KAN network model built in step S4, train the model, use K-fold cross validation to comprehensively evaluate the contribution of each feature to the model performance, and select the optimal feature set Finally, save the optimal feature set The trained KAN network model is used for the test set HF Test Also select a set with the same features as the training set And the optimal feature set for the test set Perform SOH estimation.

2. A cross-modal interactive lithium battery SOH estimation method according to claim 1, characterized in that: The specific steps of building a cross-modal interactive autoencoder model described in step S2 are as follows: S201: Build a relaxation voltage modal encoder with L fully connected layers, and convert the relaxation voltage curve training set U Train The input is sent to the first layer encoder for encoding. The output of each layer of encoding is used as the input of the next layer of encoder. The last layer of encoder outputs the potential feature data of the relaxation voltage, as follows: Among them, U Train represents the normalized relaxation voltage curve, represents the characteristics of the first fully connected layer network output in the relaxation voltage modality encoder, represents the output feature of the lth fully connected layer network in the relaxation voltage modality encoder, Z U represents the potential characteristics of the relaxation voltage output after passing through the L-layer encoder network, is the activation function of the lth fully connected layer in the relaxed voltage modality encoder, and are the weight and bias matrices of the lth fully connected layer in the relaxed voltage modality encoder; S202: Build an electrochemical impedance spectroscopy modal encoder containing L fully connected layers, and convert the electrochemical impedance spectroscopy curve training set E Train The input is sent to the first layer of encoder for encoding. The output of each layer of encoding is used as the input of the next layer of encoder. The last layer of encoder outputs the potential characteristic data of the electrochemical impedance spectroscopy, as follows: Among them, E Train represents the normalized electrochemical impedance spectroscopy curve, Represents the characteristics of the output of the first fully connected layer network in the electrochemical impedance spectroscopy modal encoder network, represents the output feature of the lth fully connected layer network in the electrochemical impedance spectroscopy modal encoder network, Z E represents the potential features output after passing through the L-layer encoder network, is the activation function of the lth fully connected layer in the electrochemical impedance spectroscopy modal encoder network, and is the weight and bias matrix in the lth fully connected layer in the electrochemical impedance spectroscopy modal encoder network layer; S203: Construct a cross-attention mechanism module to convert the relaxation voltage potential feature Z output in step 201 into U and the electrochemical impedance spectroscopy potential feature Z outputted in step 202 E , which is converted into query, key, and value vectors corresponding to relaxation voltage and electrochemical impedance spectroscopy potential features through learnable parameters, as shown in the following formula: in, is the weight parameter of the potential feature, Q U , K U 、V U is the query,key,value corresponding to the relaxation voltage potential feature, Q E , K E 、V E is the query, key, and value corresponding to the potential features of electrochemical impedance spectroscopy; S204: Input the query, key, and value of the relaxation voltage and electrochemical impedance spectrum obtained in step 203 into the attention mechanism layer, and output two potential features decoupled after fusing the two modalities, as shown in the following formula: in, for The potential feature representation after integrating electrochemical impedance spectroscopy modalities, for The latent feature representation after fusion of relaxation voltage modes, softmax(·) represents the softmax function; S205: Build an electrochemical impedance spectroscopy modal decoder of the P-layer fully connected layer, and convert the decoupled fusion modal potential features obtained in step 204 into As the input of the first layer decoder, it is input into the electrochemical impedance spectroscopy modal decoder. The output of each layer of decoding is used as the input of the next layer of decoder. The output of the last layer of the decoder is the reconstructed electrochemical impedance spectroscopy curve, which is as follows: Among them, E rf For E Train The curve reconstructed by the electrochemical impedance spectroscopy decoder, is the feature output by the pth fully connected layer network in the electrochemical impedance spectroscopy decoder network, is the activation function of the pth fully connected layer in the electrochemical impedance spectroscopy decoding layer, and is the weight and bias matrix in the pth fully connected layer in the electrochemical impedance spectroscopy decoder; S206: Build a relaxed voltage modal decoder of the P-layer fully connected layer, and transform the decoupled fusion modal potential features obtained in step 204 into As the input of the first layer decoder, and input to the relaxation voltage modal decoder, the output of each layer of decoding is used as the input of the next layer of decoder, and the output of the last layer of the decoder is the reconstructed relaxation voltage curve, which is as follows: Among them, U rf For U Train The curve reconstructed by the relaxation voltage decoder, is the feature output of the pth fully connected layer network in the relaxed voltage decoder network, is the activation function of the pth fully connected layer in the relaxed voltage decoder, and are the weight and bias matrices in the pth fully connected layer in the relaxed voltage decoder.

3. A cross-modal interactive lithium battery SOH estimation method according to claim 1, characterized in that: The specific steps of training and testing the cross-modal interactive autoencoder model described in step S3 are as follows: S301: Setting and initializing training parameters of a cross-modal interactive autoencoder model; S302: Data preprocessing: normalizing the relaxation voltage curve set U and the electrochemical impedance spectrum curve set E in the data set A obtained in step S1 to obtain a normalized set and construct a training set {U Train ,E Train } and the test set {U Test ,E Test }; S303: Design a model loss function, which includes a modal curve reconstruction loss, an inter-modal global similarity constraint loss, and an inter-modal adjacent non-similarity constraint loss, and is specifically as follows: in, is the sum of all loss functions of the entire cross-modal interactive autoencoder model, is the sum of the loss functions for the relaxation voltage curve and the electrochemical impedance spectroscopy curve reconstruction, The characteristic of the relaxation voltage curve is Characteristics of electrochemical impedance spectroscopy curves The similarity constraint loss function between The characteristic of the relaxation voltage curve is Characteristics of EIS curves The adjacent non-similarity constraint loss function between , β and δ are The weight parameter of The relaxation voltage curve and electrochemical impedance spectroscopy curve reconstruct the loss function It consists of two parts, including measuring the relaxation voltage curve U after model reconstruction rf Compared with the original relaxation voltage curve U Train The loss function of the similarity between them and the electrochemical impedance spectroscopy curve E after the model reconstruction is measured rf Compared with the original electrochemical impedance spectroscopy curve E Train The loss function of the similarity between is as follows: in, Represents the U Train The voltage value of the jth sampling point of the relaxation voltage curve of the i-th sample in is, Represents E Train The impedance value of the jth sampling point of the electrochemical impedance spectroscopy curve of the i-th sample, N U is the length of the relaxation voltage curve, N E is the length of the electrochemical impedance spectroscopy curve, N is the length of the electrochemical impedance spectroscopy curve; The similarity constraint loss function between the relaxation voltage curve and the electrochemical impedance spectroscopy curve Includes the features of each output layer of the encoder of both modes With features The maximum multi-core mean difference between them is calculated by using a combination of Gaussian kernel functions with different bandwidth parameters, as shown in the following formula: Where K is the number of kernel functions, λ k is the weight of the kth kernel function, which needs to satisfy Refers to using the kth kernel function to calculate the feature and The maximum mean difference between two feature distributions is expanded using the Gaussian kernel function as follows: Among them, ||·|| is the straight-line distance between two points in Euclidean space, δ k is the bandwidth parameter of the kth kernel function, which is used to control the width of the Gaussian kernel. Output feature of the first layer encoder of the impedance spectrum The data of the i-th dimension; The adjacent non-similarity constraint loss function between the relaxation voltage curve and the electrochemical impedance spectroscopy curve modes Includes features of the fully connected output layer of the relaxed voltage modulus encoder The output features of adjacent layers of the corresponding electrochemical impedance spectroscopy modal encoder The inverse of the cosine similarity loss function between the two, the output characteristics of the electrochemical impedance spectroscopy modal encoder The features of the adjacent output layer of the corresponding relaxed voltage modality encoder The opposite number of the cosine similarity loss function between is as follows: S304: Model training and testing, i.e., using the training set {U Train ,E Train }Train the cross-modal interactive autoencoder model described in step S2, using the test set {U Test ,E Test }Test the trained cross-modal interactive autoencoder model to verify the reconstruction accuracy of the model. If it does not meet the requirements, retrain until the obtained model meets the accuracy requirements. Save the model and use it as the cross-modal interactive autoencoder model.

4. A cross-modal interactive lithium battery SOH estimation method according to claim 1, characterized in that: The specific steps of building the KAN network model described in step S4 are as follows: S401: Construct a regression algorithm model based on the KAN network, define a KAN network structure with q layers, and use a learnable B-spline function in the activation function, as shown in the following formula: f(HF)=Φ q ·(Φ q-1 ·(…(Φ2(Φ1·HF)))) Among them, Φ q is the learnable activation function matrix of the qth hidden layer unit, HF is the fusion modality potential feature extracted by the cross-modal interactive autoencoder model; The Φ q The activation function matrix can be learned. This matrix is ​​a one-dimensional function representation matrix of the hidden layer nodes of the previous KAN network layer. The specific formula is as follows: in, Indicates the nth q+1 The input neurons and the nth neurons in the qth layer q The transformation function between the output neurons; Said The transformation function is parameterized by the B-spline function, which has learnable parameters and can be optimized through training data. In the KAN network, the B-spline function is used to approximate the learnable activation function that constitutes the edge of the network, as shown in the following formula: Among them, B i (x) is the B-spline basis function, c i is the B-spline basis function B i The coefficients of (·) can be obtained through network training.

5. A cross-modal interactive lithium battery SOH estimation method according to claim 1, characterized in that: The specific steps of training and testing the KAN network model described in step S5 are as follows: S501: Input the dataset A = {U, E, SOH} obtained in step S1 into the cross-modal interactive autoencoder model trained in step S3, and output the fusion modality potential feature set The feature set HF and the battery SOH set are used to construct a feature data set Z = {HF, SOH}, and the training set {HF Train ,SOH Train } and the test set {HF Test ,SOH Test }; S502: The training data set with n potential features of fused relaxation voltage mode and electrochemical impedance spectroscopy mode obtained in step S501 is Each individual feature Combined with the SOH value, n single feature data sets are constructed, as shown in the following formula: S503: The KAN network model described in the training step S4 is constructed, and the root mean square error of the cross-validation prediction results based on the model training is used as a single feature Evaluation index for HF Train n single features in Rank the importance; S504: Construct an empty optimal feature set According to the feature data set HF obtained in step S501 Train Get the remaining feature data set S505: Feature data set According to the feature importance ranking obtained in S503, feature x is selected one by one from large to small and added to the feature data set Building a dataset Train the regression prediction model based on the KAN network, and use the estimated root mean square error obtained by K-fold cross validation of the model on sample S as the set The evaluation indicators are as follows: Among them, x best is the feature dataset in the current iteration The best features to choose; is a feature set SOH estimates in the KAN regression model; S506: The optimal feature x obtained in step S505 is best Add to current feature set and remove it from the feature set S507: In order to balance the complexity of the model and improve the performance of the model to achieve better prediction results, the maximum number of features and performance improvement are set as the stop conditions of the feature subset search process. Whether the number of features and cross-validation performance obtained in steps S505 and S506 reach the improvement target is used as the condition for stopping iteration. If both conditions are met at the same time, the selection of feature subsets is stopped. If one of them is not met, the optimal subset search of steps S505 and S506 is continued. The specific condition formula is as follows: S508: Save the optimal feature set obtained in step S507 The trained KAN network model is used for the test set HF Test Also select a set with the same features as the training set And the optimal feature set for the test set Perform SOH estimation.

Images