A lithium battery soh estimation method based on physical information automatic encoder

Abstract

The application discloses a lithium battery SOH estimation method based on a physical information automatic encoder, and the method comprises the following steps: obtaining a lithium battery aging data set; data preprocessing and data set division; training and obtaining a physical information automatic encoder based on relaxation voltage, wherein a training loss function of the physical information automatic encoder comprises a reconstruction loss of the relaxation voltage and a physical information reconstruction loss of the relaxation voltage; training and obtaining a physical information automatic encoder based on electrochemical impedance spectroscopy, wherein a training loss function of the physical information automatic encoder comprises a reconstruction loss of the electrochemical impedance spectroscopy and a physical information reconstruction loss of the electrochemical impedance spectroscopy; constructing a feature training set and a feature test set; training and obtaining a lithium battery SOH estimation model; and online estimating the SOH of the lithium battery. The application introduces a physical information reconstruction loss into an automatic encoder model, can extract health features which have stronger correlation with physical characteristics of the battery, enhances the aging information and generalization ability of the extracted features, and can improve the SOH estimation precision.

Description

Technical Field

[0001] This invention relates to the field of lithium battery technology, and in particular to a lithium battery SOH estimation method based on a physical information autoencoder. Background Technology

[0002] Lithium-ion batteries are widely used in electric vehicles, portable electronic products, and energy storage devices. The State of Health (SOH) of a lithium-ion battery is a key indicator for evaluating its aging status; therefore, accurately estimating the SOH is crucial for assessing its aging condition and performance.

[0003] Data-driven SOH estimation methods for lithium-ion batteries are currently a hot research topic, with the extraction of health features being particularly crucial. Existing researchers have proposed various feature extraction methods for battery health characteristics, broadly categorized into manual feature extraction methods and neural network-based feature extraction methods. Neural network-based methods, in particular, can automatically extract features from measurement data such as current, voltage, temperature, and electrochemical impedance spectroscopy using powerful neural networks, avoiding the complexity of manually constructing features. However, this feature extraction method is a "black box" process, heavily reliant on operational data and lacking understanding of the underlying electrochemical principles and physical constraints of the battery system. On one hand, this makes it difficult to guarantee a correlation between the extracted health features and the battery's physical characteristics, thus affecting the accuracy of SOH estimation. On the other hand, it causes lithium-ion battery SOH estimation models to over-rely on training data, potentially leading to model outputs that violate battery physical characteristics in practical applications, resulting in poor generalization ability and robustness under new operating conditions. Summary of the Invention

[0004] To overcome the problems of existing neural network-based feature extraction methods, this invention proposes a lithium battery SOH estimation method based on a physical information autoencoder.

[0005] The technical solution of this invention is as follows:

[0006] S1: Obtain the lithium battery aging dataset. Specific steps include: conducting cyclic charge-discharge experiments on multiple lithium batteries of the same type using constant current-constant voltage charging and constant current discharging methods; during each charge-discharge cycle, a rest period is set between the full charge and discharge processes, i.e., the battery enters the relaxation stage; relaxation voltage curves are collected during the relaxation stage; electrochemical impedance spectroscopy (EIS) within a certain frequency range is collected at fixed time points during the relaxation stage; and the state of harmonics (SOH) value of the lithium battery in this cycle is recorded after the discharge is completed; the relaxation voltage curves, EIS, and SOH of the lithium battery are combined to form dataset D.

[0007] S2: Data preprocessing and dataset partitioning, which involves cleaning the data in dataset D described in step S1, including outlier removal and missing value imputation, and further partitioning the preprocessed dataset into training set D. Train and test set D Test ;

[0008] S3: Train and obtain a physical information autoencoder based on relaxation voltage. Specific steps include: First, build a physical information autoencoder model AE based on relaxation voltage. rv The AE rv The model consists of an encoder network and a decoder network; through the model AE rv For the training set D described in step S2 Train The relaxation voltage curve in the original data is reconstructed to obtain the reconstructed relaxation voltage curve. Based on the relaxation voltage curve and the reconstructed relaxation voltage curve, the reconstruction loss of the relaxation voltage is constructed. rv Based on a second-order RC equivalent circuit model, the relaxation voltage curve and the reconstructed relaxation voltage curve are fitted to obtain the equivalent circuit physical parameters of the relaxation voltage curve and the reconstructed relaxation voltage curve. Then, the physical information reconstruction loss of the relaxation voltage is further constructed. loss rv and Perform linear weighting to obtain AE rv The model's training loss function; based on AE rv The training loss function of the model uses the training set D described in step S2. Train Training AE on relaxation voltage curve rv Save the model after training is complete.

[0009] S4: Train and obtain a physical information autoencoder based on electrochemical impedance spectroscopy. Specific steps include: first, building a physical information autoencoder model AE based on electrochemical impedance spectroscopy. eis The AE eis The model consists of an encoder network and a decoder network; through the model AE eis For the training set D described in step S2 Train The electrochemical impedance spectroscopy (EIS) in the original image is reconstructed to obtain the reconstructed EIS. Based on the original EIS and the reconstructed EIS, the reconstruction loss of the EIS is constructed. eis Based on a fractional-order equivalent circuit model, the electrochemical impedance spectroscopy (EIS) and the reconstructed EIS are fitted to obtain the equivalent circuit physical parameters of the EIS and the reconstructed EIS. Then, the physical information reconstruction loss of the EIS is further constructed. loss eis and Perform linear weighting to obtain AE eis The model's training loss function; based on AE eis The training loss function of the model uses the training set D described in step S2. Train Electrochemical impedance spectroscopy training AE eis Save the model after training is complete.

[0010] S5: Construct the feature training set and feature test set. Specific steps include: [The text abruptly ends here, likely due to an incomplete sentence or a formatting error.] Train The relaxation voltage curve is input into the model AE described in step S3. rv In the encoder network, the relaxation voltage features of the training set are obtained. The training set D described in step S2 Train The electrochemical impedance spectroscopy is input into the model AE described in step S4. eis In the encoder network, the electrochemical impedance spectral characteristics of the training set are obtained. The above and training set D Train The SOH values ​​in the training set F constitute the features. Train The test set D described in step S2 Test The relaxation voltage curve is input into the model AE described in step S3. rv In the encoder network, the relaxation voltage characteristics of the test set are obtained. The test set D described in step S2 Test The electrochemical impedance spectroscopy is input into the model AE described in step S4. eis In the encoder network, the electrochemical impedance spectral characteristics of the test set are obtained. The above and test set D Test The SOH values ​​in the feature test set F Test ;

[0011] S6: Train and obtain the SOH estimation model for lithium batteries. Specific steps include: building an extreme learning machine model; training the model based on the feature set F described in step S5. Train and feature test set F Test The Extreme Learning Machine model is trained and tested. If the test accuracy does not meet the requirements, it is retrained. If the accuracy meets the requirements, the model is saved.

[0012] S7: Online estimation of SOH in lithium batteries, specifically including: acquiring relaxation voltage curves and electrochemical impedance spectroscopy from actual lithium battery charge and discharge processes; performing data preprocessing using the method described in step S2; and inputting the preprocessed relaxation voltage curves into the physical information autoencoder model AE based on relaxation voltage described in step S3. rvThe relaxation voltage characteristics are obtained in the encoder network; the preprocessed electrochemical impedance spectroscopy is then input into the physical information autoencoder model AE based on electrochemical impedance spectroscopy described in step S4. eis In the encoder network, the electrochemical impedance spectroscopy features are obtained; the relaxation voltage features and electrochemical impedance spectroscopy features are input into the extreme learning machine model described in step S6 to estimate the SOH of the lithium battery.

[0013] In this scheme, the specific steps for training and obtaining the physical information autoencoder based on relaxation voltage in step S3 are as follows:

[0014] S301: Building a Physical Information Autoencoder Model (AE) Based on Relaxation Voltage rv The model consists of an encoder network and a decoder network, and the model AE rv Both the encoder and decoder networks consist of fully connected layers;

[0015] S302: Transfer the training set D described in step S2 to... Train The relaxation voltage curve U is input into the model AE. rv The latent feature a is obtained from the encoder network, and the latent feature a is input into the model AE. rv The relaxation voltage curve is reconstructed in the decoder network, and the relevant formula is shown below:

[0016]

[0017] In the formula, The model AE is rv The weight matrix of the encoder network, The model AE is rv The bias vector of the encoder network, The model AE represents rv The activation function of the encoder network, The model AE is rv The weight matrix of the decoder network, The model AE is rv The bias vector of the decoder network. The model AE represents rv The activation function of the decoder network, U pred It is the relaxation voltage curve reconstructed from the training set;

[0018] S303: Reconstruction loss for relaxing voltage rv The calculation formula is as follows:

[0019]

[0020] Among them, Un It is the relaxation voltage curve of the nth sample in the training set. It is the reconstructed relaxation voltage curve of the nth sample in the training set, where N is the number of relaxation voltage curve samples in the training set. Represents the square of the L2 norm;

[0021] S304: Construct a second-order RC equivalent circuit model based on the relaxation stage. The second-order RC equivalent circuit model consists of one internal resistance and two RC parallel parts connected in series. The RC parallel parts consist of one ohmic resistor and one capacitor connected in parallel. The relaxation voltage formula of the second-order RC equivalent circuit model is:

[0022]

[0023] Where OCV represents the open-circuit voltage, I represents the cutoff current during the constant voltage charging stage, t represents the sampling time corresponding to the relaxation voltage data, R1 is the first-order polarization resistor, R2 is the second-order polarization resistor, C1 is the first-order polarization capacitor, C2 is the second-order polarization capacitor, and U(t) represents the relaxation voltage data at sampling time t.

[0024] Furthermore, given the open-circuit voltage OCV, the first-order polarization resistance R1, and the second-order polarization resistance R2, the internal resistance R0 of the second-order RC equivalent circuit model can be calculated. The formula for calculating the internal resistance of the second-order RC equivalent circuit model is as follows:

[0025]

[0026] Where U(0) represents the relaxation voltage at the initial moment of the relaxation phase;

[0027] S305: Calculate the equivalent circuit physical parameters of all N relaxation voltage curves in the training set, that is, based on the relaxation voltage formula of the second-order RC equivalent circuit model described in step S304, use the nonlinear least squares method to fit all N relaxation voltage curves in the training set to obtain the open-circuit voltage, first-order polarization resistance, second-order polarization resistance, first-order polarization capacitance, and second-order polarization capacitance corresponding to all N relaxation voltage curves in the training set. Furthermore, according to the calculation formula of the internal resistance of the second-order RC equivalent circuit model described in step S304, obtain the internal resistance corresponding to all N relaxation voltage curves in the training set.

[0028] The relaxation voltage curve U of the nth sample in the training set n For example, its equivalent circuit physical parameter Phy U,n as follows:

[0029] Phy U,n =[OCV n R 1,n R 2,n C1,n C 2,n R 0,n ]

[0030] Among them, OCV n R 1,n R 2,n C 1,n C 2,n and R 0,n U represents the relaxation voltage curve of the nth sample in the training set. n The corresponding open-circuit voltage, first-order polarization resistance, second-order polarization resistance, first-order polarization capacitance, second-order polarization capacitance, and internal resistance.

[0031] S306: Calculate the equivalent circuit physical parameters of all N reconstructed relaxation voltage curves in the training set, that is, based on the relaxation voltage formula of the second-order RC equivalent circuit model described in step S304, use the nonlinear least squares method to fit all N reconstructed relaxation voltage curves in the training set to obtain the open-circuit voltage, first-order polarization resistance, second-order polarization resistance, first-order polarization capacitance, and second-order polarization capacitance corresponding to all N reconstructed relaxation voltage curves in the training set. Furthermore, according to the calculation formula of the internal resistance of the second-order RC equivalent circuit model described in step S304, obtain the internal resistance corresponding to all N reconstructed relaxation voltage curves in the training set.

[0032] The relaxation voltage curve reconstructed from the nth sample in the training set For example, its equivalent circuit physical parameters as follows:

[0033]

[0034] in, and These represent the relaxation voltage curves of the reconstructed nth sample in the training set. The corresponding open-circuit voltage, first-order polarization resistance, second-order polarization resistance, first-order polarization capacitance, second-order polarization capacitance, and internal resistance;

[0035] S307: Physical information reconstruction loss for constructing relaxation voltage The formula is as follows:

[0036]

[0037] Among them, Phy U,n These are the equivalent circuit physical parameters of the relaxation voltage curve of the nth sample in the training set. represents the equivalent circuit physical parameters of the reconstructed relaxation voltage curve of the nth sample in the training set, where N is the number of relaxation voltage curve samples in the training set. Represents the square of the L2 norm;

[0038] S308: Constructing a Physical Information Autoencoder Model AE Based on Relaxation Voltage rv The training loss function loss1 is obtained by applying the loss function described in step S303. rv and as described in step S307 The linear weighted average is used to obtain loss1, and its calculation formula is as follows:

[0039]

[0040] Where δ is the linear weighting factor of loss1, satisfying δ∈[0,1];

[0041] S309: Using the training set D described in step S2 Train Training the relaxation voltage curve in the autoencoder model AE based on the physical information of the relaxation voltage. rv Based on the loss1 described in step S308, AE is updated using the backpropagation algorithm. rv The model's weight parameters are saved in the AE file after training. rv Model.

[0042] In this scheme, the specific steps for training and obtaining the physical information autoencoder based on electrochemical impedance spectroscopy in step S4 are as follows:

[0043] S401: Building a Physical Information Autoencoder Model Based on Electrochemical Impedance Spectroscopy (AE) eis The model consists of an encoder network and a decoder network, and the model AE eis Both the encoder and decoder networks consist of fully connected layers;

[0044] S402: Transfer the training set D described in step S2 to... Train The electrochemical impedance spectroscopy Z in the model AE is input into the model AE. eis The latent feature c is obtained from the encoder network, and the latent feature c is input into the model AE. eis The electrochemical impedance spectroscopy is reconstructed in the decoder network, and the relevant formulas are as follows:

[0045]

[0046] in, The model AE is eis The weight matrix of the encoder network, The model AE is eis The bias vector of the encoder network, The model AE represents eis The activation function of the encoder network, The model AE is eisThe weight matrix of the decoder network, The model AE is eis The bias vector of the decoder network. The model AE represents eis The activation function of the decoder network, Z pred It is the electrochemical impedance spectrum reconstructed from the training set;

[0047] S403: Reconstruction loss in constructing electrochemical impedance spectroscopy eis The formula is as follows:

[0048]

[0049] Among them, Z m It is the electrochemical impedance spectrum of the m-th sample in the training set. It is the reconstructed electrochemical impedance spectrum of the m-th sample in the training set, where M is the number of electrochemical impedance spectrum samples in the training set. Represents the square of the L2 norm;

[0050] S404: Construct a fractional-order equivalent circuit model based on electrochemical impedance spectroscopy. This model describes the electrochemical impedance spectroscopy in four parts: ultra-high frequency impedance, high frequency impedance, mid-frequency impedance, and low-frequency impedance. The ultra-high frequency impedance consists of an inductor and a resistor connected in series. The high frequency impedance consists of a constant-phase element and a resistor connected in parallel. The mid-frequency impedance consists of a constant-phase element and a resistor connected in parallel. The low-frequency impedance consists of a Warburg element. The impedance formulas for the constructed fractional-order equivalent circuit model are as follows:

[0051]

[0052] Where ω is the angular frequency corresponding to the electrochemical impedance data, satisfying ω=2πf, j is the imaginary unit, f is the frequency corresponding to the electrochemical impedance data, Z(ω) is the electrochemical impedance data corresponding to the angular frequency ω, L is the inductance of the ultra-high frequency impedance, r0 is the resistance of the ultra-high frequency impedance, r1 is the resistance of the high frequency impedance, Q1 is the coefficient of the constant phase element of the high frequency impedance, α1 is the order of the constant phase element of the high frequency impedance, r2 is the resistance of the mid-frequency impedance, Q2 is the coefficient of the constant phase element of the mid-frequency impedance, α2 is the order of the constant phase element of the mid-frequency impedance, W is the coefficient of the Warburg element of the low-frequency impedance, β is the order of the Warburg element of the low-frequency impedance, and the values ​​of α1 and α2 are both in the range of (0,1).

[0053] S405: Calculate the equivalent circuit physical parameters of all M electrochemical impedance spectra in the training set, that is, based on the impedance formula of the fractional-order equivalent circuit model described in step S404, use the nonlinear least squares method to fit all M electrochemical impedance spectra in the training set, and obtain the inductance, resistance, resistance, coefficient of constant-phase element, order of constant-phase element, resistance, resistance, coefficient of constant-phase element, order of constant-phase element, coefficient of Warburg element, and order of Warburg element for all M electrochemical impedance spectra in the training set.

[0054] The electrochemical impedance spectroscopy Z of the m-th sample in the training set m For example, its equivalent circuit physical parameter Phy Z,m as follows:

[0055] Phy Z,m =[L m ,r 0,m ,r 1,m Q 1,m ,α 1,n ,r 2,m Q 2,m ,α 2,m W m ,β m ]

[0056] Among them, L m r 0,m r 1,m Q 1,m α 1,m r 2,m Q 2,m α 2,m W m and β m The following are the electrochemical impedance spectra Z of the m-th sample in the training set. m The corresponding inductance of ultra-high frequency impedance, resistance of ultra-high frequency impedance, resistance of high frequency impedance, coefficient of constant phase element of high frequency impedance, order of constant phase element of high frequency impedance, resistance of intermediate frequency impedance, coefficient of constant phase element of intermediate frequency impedance, order of constant phase element of intermediate frequency impedance, coefficient of Warburg element of low frequency impedance and order of Warburg element of low frequency impedance.

[0057] S406: Calculate the equivalent circuit physical parameters of all M reconstructed electrochemical impedance spectra in the training set. Based on the impedance formula of the fractional-order equivalent circuit model described in step S404, use the nonlinear least squares method to fit all M reconstructed electrochemical impedance spectra in the training set to obtain the inductance, resistance, and constant-phase element coefficients of the ultra-high frequency impedance, the order of the constant-phase element of the high frequency impedance, the resistance, resistance, and constant-phase element coefficients of the mid-frequency impedance, the order of the constant-phase element of the mid-frequency impedance, the coefficients and order of the Warburg element of the low-frequency impedance.

[0058] Reconstructed electrochemical impedance spectroscopy of the m-th sample in the training set For example, its equivalent circuit physical parameters as follows:

[0059]

[0060] in, and The following are the reconstructed electrochemical impedance spectra of the m-th sample in the training set. The corresponding inductance of ultra-high frequency impedance, resistance of ultra-high frequency impedance, resistance of high frequency impedance, coefficient of constant phase element of high frequency impedance, order of constant phase element of high frequency impedance, resistance of intermediate frequency impedance, coefficient of constant phase element of intermediate frequency impedance, order of constant phase element of intermediate frequency impedance, coefficient of Warburg element of low frequency impedance and order of Warburg element of low frequency impedance.

[0061] S407: Physical information reconstruction loss in constructing electrochemical impedance spectroscopy The formula is as follows:

[0062]

[0063] Among them, Phy Z,m The equivalent circuit physical parameters represent the electrochemical impedance spectrum of the m-th sample in the training set. Let M represent the equivalent circuit physical parameters of the reconstructed electrochemical impedance spectrum of the m-th sample in the training set, where M is the number of electrochemical impedance spectrum samples in the training set. Represents the square of the L2 norm;

[0064] S408: Constructing a Physical Information Autoencoder Model Based on Electrochemical Impedance Spectroscopy (AE) eis The training loss function loss2 is obtained by applying the loss function described in step S403. eis and as described in step S407 The loss2 is obtained by performing a linear weighting, and the calculation formula is as follows:

[0065]

[0066] Where ε is the linear weighting factor of loss2, satisfying ε∈[0,1];

[0067] S409: Using the training set D described in step S2 Train Electrochemical impedance spectroscopy training in the autoencoder model based on the physical information of electrochemical impedance spectroscopy. eis Based on the training loss function loss2 described in step S408, AE is updated using the backpropagation algorithm. eis The model's weight parameters are saved in the AE file after training. eis Model.

[0068] Compared with the prior art, the beneficial effects of the technical solution of the present invention are:

[0069] This invention proposes a physical information autoencoder (PEA) method for estimating the state of health (SOH) of lithium-ion batteries. PEA autoencoders based on relaxation voltage and electrochemical impedance spectroscopy (EIS) are proposed, respectively. The former introduces the reconstruction loss of the second-order RC equivalent circuit physical parameters of the relaxation voltage as a physical information constraint, while the latter introduces the reconstruction loss of the fractional-order equivalent circuit physical parameters of the EIS as a physical information constraint. On the one hand, the PEA autoencoder can extract health features strongly correlated with the internal physical characteristics of the battery from the relaxation voltage curve and EIS, thereby increasing the interpretability of the features and improving the accuracy of SOH estimation. On the other hand, by introducing physical information constraints, the extracted features are less dependent on data-driven approaches, enhancing their applicability and generalization ability under different operating conditions. Attached Figure Description

[0070] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments recorded in the present invention. For those skilled in the art, other drawings can be obtained based on these drawings.

[0071] Figure 1 This invention proposes a flowchart of the steps for estimating the state of harmonics (SOH) of a lithium battery based on a physical information autoencoder.

[0072] Figure 2 This is a schematic diagram of the second-order RC equivalent circuit model based on the relaxation voltage curve used in this invention;

[0073] Figure 3This is a schematic diagram of the fractional-order equivalent circuit model based on electrochemical impedance spectroscopy used in this invention. In the diagram, CPE represents a constant-phase element. Detailed Implementation

[0074] To better understand the above-mentioned objectives, features, and advantages of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be noted that, unless otherwise specified, the embodiments and features described in these embodiments can be combined with each other.

[0075] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and therefore the scope of protection of the invention is not limited to the specific embodiments disclosed below.

[0076] In a specific embodiment, such as Figure 1 As shown, a lithium battery SOH estimation method based on a physical information autoencoder includes the following steps:

[0077] S1: Obtain the lithium battery aging dataset. Specific steps include: conducting cyclic charge-discharge experiments on multiple lithium batteries of the same type using constant current-constant voltage charging and constant current discharging methods; during each charge-discharge cycle, a rest period is set between the full charge and discharge processes, i.e., the battery enters the relaxation stage; relaxation voltage curves are collected during the relaxation stage; electrochemical impedance spectroscopy (EIS) within a certain frequency range is collected at fixed time points during the relaxation stage; and the state of harmonics (SOH) value of the lithium battery in this cycle is recorded after the discharge is completed; the relaxation voltage curves, EIS, and SOH of the lithium battery are combined to form dataset D.

[0078] S2: Data preprocessing and dataset partitioning, which involves cleaning the data in dataset D described in step S1, including outlier removal and missing value imputation, and further partitioning the preprocessed dataset into training set D. Train and test set D Test ;

[0079] S3: Train and obtain a physical information autoencoder based on relaxation voltage. Specific steps include: First, build a physical information autoencoder model AE based on relaxation voltage. rv The AE rv The model consists of an encoder network and a decoder network; through the model AE rv For the training set D described in step S2 Train The relaxation voltage curve in the original data is reconstructed to obtain the reconstructed relaxation voltage curve. Based on the relaxation voltage curve and the reconstructed relaxation voltage curve, the reconstruction loss of the relaxation voltage is constructed. rvBased on a second-order RC equivalent circuit model, the relaxation voltage curve and the reconstructed relaxation voltage curve are fitted to obtain the equivalent circuit physical parameters of the relaxation voltage curve and the reconstructed relaxation voltage curve. Then, the physical information reconstruction loss of the relaxation voltage is further constructed. loss rv and Perform linear weighting to obtain AE rv The model's training loss function; based on AE rv The training loss function of the model uses the training set D described in step S2. Train Training AE on relaxation voltage curve rv Save the model after training is complete.

[0080] S4: Train and obtain a physical information autoencoder based on electrochemical impedance spectroscopy. Specific steps include: first, building a physical information autoencoder model AE based on electrochemical impedance spectroscopy. eis The AE eis The model consists of an encoder network and a decoder network; through the model AE eis For the training set D described in step S2 Train The electrochemical impedance spectroscopy (EIS) in the original image is reconstructed to obtain the reconstructed EIS. Based on the original EIS and the reconstructed EIS, the reconstruction loss of the EIS is constructed. eis Based on a fractional-order equivalent circuit model, the electrochemical impedance spectroscopy (EIS) and the reconstructed EIS are fitted to obtain the equivalent circuit physical parameters of the EIS and the reconstructed EIS. Then, the physical information reconstruction loss of the EIS is further constructed. loss eis and Perform linear weighting to obtain AE eis The model's training loss function; based on AE eis The training loss function of the model uses the training set D described in step S2. Train Electrochemical impedance spectroscopy training AE eis Save the model after training is complete.

[0081] S5: Construct the feature training set and feature test set. Specific steps include: [The text abruptly ends here, likely due to an incomplete sentence or a formatting error.] Train The relaxation voltage curve is input into the model AE described in step S3. rv In the encoder network, the relaxation voltage features of the training set are obtained. The training set D described in step S2 Train The electrochemical impedance spectroscopy is input into the model AE described in step S4. eisIn the encoder network, the electrochemical impedance spectral characteristics of the training set are obtained. The above and training set D Train The SOH values ​​in the training set F constitute the features. Train The test set D described in step S2 Test The relaxation voltage curve is input into the model AE described in step S3. rv In the encoder network, the relaxation voltage characteristics of the test set are obtained. The test set D described in step S2 Test The electrochemical impedance spectroscopy is input into the model AE described in step S4. eis In the encoder network, the electrochemical impedance spectral characteristics of the test set are obtained. The above and test set D Test The SOH values ​​in the feature test set F Test ;

[0082] S6: Train and obtain the SOH estimation model for lithium batteries. Specific steps include: building an extreme learning machine model; training the model based on the feature set F described in step S5. Train and feature test set F Test The Extreme Learning Machine model is trained and tested. If the test accuracy does not meet the requirements, it is retrained. If the accuracy meets the requirements, the model is saved.

[0083] S7: Online estimation of SOH in lithium batteries, specifically including: acquiring relaxation voltage curves and electrochemical impedance spectroscopy from actual lithium battery charge and discharge processes; performing data preprocessing using the method described in step S2; and inputting the preprocessed relaxation voltage curves into the physical information autoencoder model AE based on relaxation voltage described in step S3. rv The relaxation voltage characteristics are obtained in the encoder network; the preprocessed electrochemical impedance spectroscopy is then input into the physical information autoencoder model AE based on electrochemical impedance spectroscopy described in step S4. eis In the encoder network, the electrochemical impedance spectroscopy features are obtained; the relaxation voltage features and electrochemical impedance spectroscopy features are input into the extreme learning machine model described in step S6 to estimate the SOH of the lithium battery.

[0084] Step S3 involves training and obtaining an autoencoder based on relaxation voltage physical information, and the specific steps are as follows:

[0085] S301: Building a Physical Information Autoencoder Model (AE) Based on Relaxation Voltage rv The model consists of an encoder network and a decoder network, and the model AE rv Both the encoder and decoder networks consist of fully connected layers;

[0086] S302: Transfer the training set D described in step S2 to... Train The relaxation voltage curve U is input into the model AE. rv The latent feature a is obtained from the encoder network, and the latent feature a is input into the model AE. rv The relaxation voltage curve is reconstructed in the decoder network, and the relevant formula is shown below:

[0087]

[0088] In the formula, The model AE is rv The weight matrix of the encoder network, The model AE is rv The bias vector of the encoder network, The model AE represents rv The activation function of the encoder network, The model AE is rv The weight matrix of the decoder network, The model AE is rv The bias vector of the decoder network. The model AE represents rv The activation function of the decoder network, U pred It is the relaxation voltage curve reconstructed from the training set;

[0089] S303: Reconstruction loss for relaxing voltage rv The calculation formula is as follows:

[0090]

[0091] Among them, U n It is the relaxation voltage curve of the nth sample in the training set. It is the reconstructed relaxation voltage curve of the nth sample in the training set, where N is the number of relaxation voltage curve samples in the training set. Represents the square of the L2 norm;

[0092] S304: Construct a second-order RC equivalent circuit model based on the relaxation stage, such as... Figure 2 As shown, the second-order RC equivalent circuit model consists of one internal resistance and two RC parallel sections connected in series. The RC parallel sections consist of one ohmic resistor and one capacitor connected in parallel. The relaxation voltage formula for the second-order RC equivalent circuit model is:

[0093]

[0094] Where OCV represents the open-circuit voltage, I represents the cutoff current during the constant voltage charging stage, t represents the sampling time corresponding to the relaxation voltage data, R1 is the first-order polarization resistor, R2 is the second-order polarization resistor, C1 is the first-order polarization capacitor, C2 is the second-order polarization capacitor, and U(t) represents the relaxation voltage data at sampling time t.

[0095] Furthermore, given the open-circuit voltage OCV, the first-order polarization resistance R1, and the second-order polarization resistance R2, the internal resistance R0 of the second-order RC equivalent circuit model can be calculated. The formula for calculating the internal resistance of the second-order RC equivalent circuit model is as follows:

[0096]

[0097] Where U(0) represents the relaxation voltage at the initial moment of the relaxation phase;

[0098] S305: Calculate the equivalent circuit physical parameters of all N relaxation voltage curves in the training set, that is, based on the relaxation voltage formula of the second-order RC equivalent circuit model described in step S304, use the nonlinear least squares method to fit all N relaxation voltage curves in the training set to obtain the open-circuit voltage, first-order polarization resistance, second-order polarization resistance, first-order polarization capacitance, and second-order polarization capacitance corresponding to all N relaxation voltage curves in the training set. Furthermore, according to the calculation formula of the internal resistance of the second-order RC equivalent circuit model described in step S304, obtain the internal resistance corresponding to all N relaxation voltage curves in the training set.

[0099] The relaxation voltage curve U of the nth sample in the training set n For example, its equivalent circuit physical parameter Phy U,n as follows:

[0100] Phy U,n =[OCV n R 1,n R 2,n C 1,n C 2,n R 0,n ]

[0101] Among them, OCV n R 1,n R 2,n C 1,n C 2,n and R 0,n U represents the relaxation voltage curve of the nth sample in the training set. n The corresponding open-circuit voltage, first-order polarization resistance, second-order polarization resistance, first-order polarization capacitance, second-order polarization capacitance, and internal resistance.

[0102] S306: Calculate the equivalent circuit physical parameters of all N reconstructed relaxation voltage curves in the training set, that is, based on the relaxation voltage formula of the second-order RC equivalent circuit model described in step S304, use the nonlinear least squares method to fit all N reconstructed relaxation voltage curves in the training set to obtain the open-circuit voltage, first-order polarization resistance, second-order polarization resistance, first-order polarization capacitance, and second-order polarization capacitance corresponding to all N reconstructed relaxation voltage curves in the training set. Furthermore, according to the calculation formula of the internal resistance of the second-order RC equivalent circuit model described in step S304, obtain the internal resistance corresponding to all N reconstructed relaxation voltage curves in the training set.

[0103] The relaxation voltage curve reconstructed from the nth sample in the training set For example, its equivalent circuit physical parameters as follows:

[0104]

[0105] in, and These represent the relaxation voltage curves of the reconstructed nth sample in the training set. The corresponding open-circuit voltage, first-order polarization resistance, second-order polarization resistance, first-order polarization capacitance, second-order polarization capacitance, and internal resistance;

[0106] S307: Physical information reconstruction loss for constructing relaxation voltage The formula is as follows:

[0107]

[0108] Among them, Phy U,n These are the equivalent circuit physical parameters of the relaxation voltage curve of the nth sample in the training set. represents the equivalent circuit physical parameters of the reconstructed relaxation voltage curve of the nth sample in the training set, where N is the number of relaxation voltage curve samples in the training set. Represents the square of the L2 norm;

[0109] S308: Constructing a Physical Information Autoencoder Model AE Based on Relaxation Voltage rv The training loss function loss1 is obtained by applying the loss function described in step S303. rv and as described in step S307 The linear weighted average is used to obtain loss1, and its calculation formula is as follows:

[0110]

[0111] Where δ is the linear weighting factor of loss1, satisfying δ∈[0,1];

[0112] S309: Using the training set D described in step S2 Train Training the relaxation voltage curve in the autoencoder model AE based on the physical information of the relaxation voltage. rv Based on the loss1 described in step S308, AE is updated using the backpropagation algorithm. rv The model's weight parameters are saved in the AE file after training. rv Model.

[0113] The specific steps for training and obtaining the physical information autoencoder based on electrochemical impedance spectroscopy, as described in step S4, are as follows:

[0114] S401: Building a Physical Information Autoencoder Model Based on Electrochemical Impedance Spectroscopy (AE) eis The model consists of an encoder network and a decoder network, and the model AE eis Both the encoder and decoder networks consist of fully connected layers;

[0115] S402: Transfer the training set D described in step S2 to... Train The electrochemical impedance spectroscopy Z in the model AE is input into the model AE. eis The latent feature c is obtained from the encoder network, and the latent feature c is input into the model AE. eis The electrochemical impedance spectroscopy is reconstructed in the decoder network, and the relevant formulas are as follows:

[0116]

[0117] in, The model AE is eis The weight matrix of the encoder network, The model AE is eis The bias vector of the encoder network, The model AE represents eis The activation function of the encoder network, The model AE is eis The weight matrix of the decoder network, The model AE is eis The bias vector of the decoder network. The model AE represents eis The activation function of the decoder network, Z pred It is the electrochemical impedance spectrum reconstructed from the training set;

[0118] S403: Reconstruction loss in constructing electrochemical impedance spectroscopy eis The formula is as follows:

[0119]

[0120] Among them, Z m It is the electrochemical impedance spectrum of the m-th sample in the training set. It is the reconstructed electrochemical impedance spectrum of the m-th sample in the training set, where M is the number of electrochemical impedance spectrum samples in the training set. Represents the square of the L2 norm;

[0121] S404: Construct a fractional-order equivalent circuit model based on electrochemical impedance spectroscopy, such as Figure 3 As shown, the fractional-order equivalent circuit model describes the electrochemical impedance spectrum in four parts: ultra-high frequency impedance, high frequency impedance, mid-frequency impedance, and low-frequency impedance. The ultra-high frequency impedance consists of an inductor and a resistor connected in series; the high frequency impedance consists of a constant-phase element and a resistor connected in parallel; the mid-frequency impedance consists of a constant-phase element and a resistor connected in parallel; and the low-frequency impedance consists of a Warburg element. The impedance formula of the constructed fractional-order equivalent circuit model is as follows:

[0122]

[0123] Where ω is the angular frequency corresponding to the electrochemical impedance data, satisfying ω=2πf, j is the imaginary unit, f is the frequency corresponding to the electrochemical impedance data, Z(ω) is the electrochemical impedance data corresponding to the angular frequency ω, L is the inductance of the ultra-high frequency impedance, r0 is the resistance of the ultra-high frequency impedance, r1 is the resistance of the high frequency impedance, Q1 is the coefficient of the constant phase element of the high frequency impedance, α1 is the order of the constant phase element of the high frequency impedance, r2 is the resistance of the mid-frequency impedance, Q2 is the coefficient of the constant phase element of the mid-frequency impedance, α2 is the order of the constant phase element of the mid-frequency impedance, W is the coefficient of the Warburg element of the low-frequency impedance, β is the order of the Warburg element of the low-frequency impedance, and the values ​​of α1 and α2 are both in the range of (0,1).

[0124] S405: Calculate the equivalent circuit physical parameters of all M electrochemical impedance spectra in the training set, that is, based on the impedance formula of the fractional-order equivalent circuit model described in step S404, use the nonlinear least squares method to fit all M electrochemical impedance spectra in the training set, and obtain the inductance, resistance, resistance, coefficient of constant-phase element, order of constant-phase element, resistance, resistance, coefficient of constant-phase element, order of constant-phase element, coefficient of Warburg element, and order of Warburg element for all M electrochemical impedance spectra in the training set.

[0125] The electrochemical impedance spectroscopy Z of the m-th sample in the training set m For example, its equivalent circuit physical parameter Phy Z,mas follows:

[0126] Phy Z,m =[L m ,r 0,m ,r 1,m Q 1,m ,α 1,m ,r 2,m Q 2,m ,α 2,m W m ,β m ]

[0127] Among them, L m r 0,m r 1,m Q 1,m α 1,m r 2,m Q 2,m α 2,m W m and β m The following are the electrochemical impedance spectra Z of the m-th sample in the training set. m The corresponding inductance of ultra-high frequency impedance, resistance of ultra-high frequency impedance, resistance of high frequency impedance, coefficient of constant phase element of high frequency impedance, order of constant phase element of high frequency impedance, resistance of intermediate frequency impedance, coefficient of constant phase element of intermediate frequency impedance, order of constant phase element of intermediate frequency impedance, coefficient of Warburg element of low frequency impedance and order of Warburg element of low frequency impedance.

[0128] S406: Calculate the equivalent circuit physical parameters of all M reconstructed electrochemical impedance spectra in the training set. Based on the impedance formula of the fractional-order equivalent circuit model described in step S404, use the nonlinear least squares method to fit all M reconstructed electrochemical impedance spectra in the training set to obtain the inductance, resistance, and constant-phase element coefficients of the ultra-high frequency impedance, the order of the constant-phase element of the high frequency impedance, the resistance, resistance, and constant-phase element coefficients of the mid-frequency impedance, the order of the constant-phase element of the mid-frequency impedance, the coefficients and order of the Warburg element of the low-frequency impedance.

[0129] Reconstructed electrochemical impedance spectroscopy of the m-th sample in the training set For example, its equivalent circuit physical parameters as follows:

[0130]

[0131] in, and The following are the reconstructed electrochemical impedance spectra of the m-th sample in the training set. The corresponding inductance of ultra-high frequency impedance, resistance of ultra-high frequency impedance, resistance of high frequency impedance, coefficient of constant phase element of high frequency impedance, order of constant phase element of high frequency impedance, resistance of intermediate frequency impedance, coefficient of constant phase element of intermediate frequency impedance, order of constant phase element of intermediate frequency impedance, coefficient of Warburg element of low frequency impedance and order of Warburg element of low frequency impedance.

[0132] S407: Physical information reconstruction loss in constructing electrochemical impedance spectroscopy The formula is as follows:

[0133]

[0134] Among them, Phy Z,m The equivalent circuit physical parameters represent the electrochemical impedance spectrum of the m-th sample in the training set. Let M represent the equivalent circuit physical parameters of the reconstructed electrochemical impedance spectrum of the m-th sample in the training set, where M is the number of electrochemical impedance spectrum samples in the training set. Represents the square of the L2 norm;

[0135] S408: Constructing a Physical Information Autoencoder Model Based on Electrochemical Impedance Spectroscopy (AE) eis The training loss function loss2 is obtained by applying the loss function described in step S403. eis and as described in step S407 The loss2 is obtained by performing a linear weighting, and the calculation formula is as follows:

[0136]

[0137] Where ε is the linear weighting factor of loss2, satisfying ε∈[0,1];

[0138] S409: Using the training set D described in step S2 Train Electrochemical impedance spectroscopy training in the autoencoder model based on the physical information of electrochemical impedance spectroscopy. eis Based on the training loss function loss2 described in step S408, AE is updated using the backpropagation algorithm. eis The model's weight parameters are saved in the AE file after training. eis Model.

Claims

1. A lithium battery SOH estimation method based on physical information autoencoder, characterized in that, The specific steps are as follows: S1: Obtain a lithium battery aging data set, the specific steps comprising: performing cyclic charging and discharging experiments on a plurality of same type lithium batteries in a constant current-constant voltage charging and constant current discharging manner; during each charging and discharging cycle, setting a static time between full charging and discharging, i.e., the battery enters a relaxation stage, collecting a relaxation voltage curve in the relaxation stage, collecting an electrochemical impedance spectrum in a certain frequency range at a fixed time point in the relaxation stage, and recording the SOH value of the lithium battery in the current cycle after discharging; the relaxation voltage curve, the electrochemical impedance spectrum and the SOH of the lithium battery together constitute a data set D; S2: data preprocessing and dataset division, that is, data cleaning is performed on the data in the dataset D described in step S1, including outlier removal and missing value filling, and the dataset after data preprocessing is further divided into a training set D Train and a test set D Test ; S3: Train and obtain a physical information autoencoder based on relaxation voltage. Specific steps include: First, build a physical information autoencoder model AE based on relaxation voltage. rv The AE rv The model consists of an encoder network and a decoder network; through the model AE rv For the training set D described in step S2 Train The relaxation voltage curve in the original data is reconstructed to obtain the reconstructed relaxation voltage curve. Based on the relaxation voltage curve and the reconstructed relaxation voltage curve, the reconstruction loss of the relaxation voltage is constructed. rv Based on a second-order RC equivalent circuit model, the relaxation voltage curve and the reconstructed relaxation voltage curve are fitted to obtain the equivalent circuit physical parameters of the relaxation voltage curve and the reconstructed relaxation voltage curve. Then, the physical information reconstruction loss of the relaxation voltage is further constructed. loss rv and Perform linear weighting to obtain AE rv The model's training loss function; based on AE rv The training loss function of the model uses the training set D described in step S2. Train Training AE on relaxation voltage curve rv Save the model after training is complete. S4: training and obtaining an automatic encoder of physical information based on electrochemical impedance spectroscopy, the specific steps comprising: first, building an automatic encoder of physical information based on electrochemical impedance spectroscopy model AE eis , the AE eis model consisting of an encoder network and a decoder network; reconstructing the electrochemical impedance spectroscopy in the training set D eis described in step S2 through the model AE Train , to obtain the reconstructed electrochemical impedance spectroscopy; based on the electrochemical impedance spectroscopy and the reconstructed electrochemical impedance spectroscopy, constructing the reconstruction loss loss eis of electrochemical impedance spectroscopy; based on the fractional order equivalent circuit model, fitting the electrochemical impedance spectroscopy and the reconstructed electrochemical impedance spectroscopy respectively, to obtain the equivalent circuit physical parameters of the electrochemical impedance spectroscopy and the equivalent circuit physical parameters of the reconstructed electrochemical impedance spectroscopy, and then further constructing the physical information reconstruction loss of electrochemical impedance spectroscopy eis and ; linearly weighting loss eis to obtain the training loss function of the AE eis model; based on the training loss function of the AE Train model, training the AE eis model using the electrochemical impedance spectroscopy in the training set D described in step S2, and saving the model after training. S5: Construct the feature training set and feature test set. Specific steps include: [The text abruptly ends here, likely due to an incomplete sentence or a formatting error.] Train The relaxation voltage curve is input into the model AE described in step S3. rv In the encoder network, the relaxation voltage features of the training set are obtained. The training set D described in step S2 Train The electrochemical impedance spectroscopy is input into the model AE described in step S4. eis In the encoder network, the electrochemical impedance spectral characteristics of the training set are obtained. The above and training set D Train The SOH values ​​in the training set F constitute the features. Train The test set D described in step S2 Test The relaxation voltage curve is input into the model AE described in step S3. rv In the encoder network, the relaxation voltage characteristics of the test set are obtained. The test set D described in step S2 Test The electrochemical impedance spectroscopy is input into the model AE described in step S4. eis In the encoder network, the electrochemical impedance spectral characteristics of the test set are obtained. The above and test set D Test The SOH values ​​in the feature test set F Test ; S6: Train and obtain the lithium battery SOH estimation model, the specific steps comprising: building an extreme learning machine model; based on the feature training set F Train and feature test set F Test , described in step S5, train and test the extreme learning machine model, retrain if the test accuracy does not meet the requirements, and save the model if the requirements are met; S7: online estimation of lithium battery SOH, the specific steps comprising: collecting relaxation voltage curve and electrochemical impedance spectrum from actual lithium battery charging and discharging process, and adopting the method described in step S2 to perform data preprocessing; inputting the relaxation voltage curve after data preprocessing into the encoder network of the physical information autoencoder model AE rv described in step S3 to obtain relaxation voltage characteristics; inputting the electrochemical impedance spectrum after data preprocessing into the encoder network of the physical information autoencoder model AE eis described in step S4 to obtain electrochemical impedance spectrum characteristics; and inputting the relaxation voltage characteristics and the electrochemical impedance spectrum characteristics into the extreme learning machine model described in step S6 to estimate the lithium battery SOH.

2. The lithium battery SOH estimation method based on physical information automatic encoder according to claim 1, wherein, The specific steps of training and obtaining the physical information automatic encoder based on the relaxation voltage in step S3 are as follows: S301: build a physical information automatic encoder model AE based on relaxation voltage rv The model AE is composed of an encoder network and a decoder network rv Both the encoder network and the decoder network of the model AE are composed of fully connected layers S302: Transfer the training set D described in step S2 to... Train The relaxation voltage curve U is input into the model AE. rv The latent feature a is obtained from the encoder network, and the latent feature a is input into the model AE. rv The relaxation voltage curve is reconstructed in the decoder network, and the relevant formula is shown below: wherein, is a weight matrix of the encoder network of the model AE rv , is a bias vector of the encoder network of the model AE rv , denotes an activation function of the encoder network of the model AE rv , is a weight matrix of the decoder network of the model AE rv , is a bias vector of the decoder network of the model AE rv , denotes an activation function of the decoder network of the model AE rv , pred is a reconstructed relaxation voltage curve of the training set; S303: Construct the reconstruction loss loss of the relaxation voltage rv The calculation formula is as follows: wherein U n is the relaxation voltage curve of the n-th sample in the training set, is the reconstructed relaxation voltage curve of the n-th sample in the training set, N is the number of samples of the relaxation voltage curve in the training set, denotes the square of the L2 norm; S304: Construct a second-order RC equivalent circuit model based on the relaxation stage, the second-order RC equivalent circuit model is composed of one internal resistance and two RC parallel parts in series, the RC parallel part is composed of one ohmic resistance and one capacitor in parallel, and the relaxation voltage formula of the second-order RC equivalent circuit model is: Wherein, OCV represents the open circuit voltage, I represents the cutoff current in the constant voltage charging stage, t represents the sampling time corresponding to the relaxation voltage data, R1 is the first-order polarization resistance, R2 is the second-order polarization resistance, C1 is the first-order polarization capacitance, C2 is the second-order polarization capacitance, and U(t) represents the relaxation voltage data at time t; Further, after the open circuit voltage OCV, the first-order polarization resistance R1 and the second-order polarization resistance R2 are known, the internal resistance R0 of the second-order RC equivalent circuit model is calculated, and the calculation formula of the internal resistance of the second-order RC equivalent circuit model is: Wherein, U(0) represents the relaxation voltage at the initial time of the relaxation stage; S305: Calculate the equivalent circuit physical parameters of all N relaxation voltage curves in the training set, i.e., the relaxation voltage formula of the second-order RC equivalent circuit model described in step S304, use the nonlinear least squares method to fit all N relaxation voltage curves in the training set, obtain the open circuit voltage, first-order polarization resistance, second-order polarization resistance, first-order polarization capacitance, and second-order polarization capacitance corresponding to all N relaxation voltage curves in the training set, and further obtain the internal resistance corresponding to all N relaxation voltage curves in the training set according to the calculation formula of the internal resistance of the second-order RC equivalent circuit model described in step S304; U(n) = U(n) + (U(n) - U(n-1)) * (1 - exp(-t / T)) n For example, the equivalent circuit physical parameter Phy U,n As follows: Phy U,n = [OCV n , R 1,n , R 2,n , C 1,n , C 2,n , R 0,n ] wherein OCV n , R 1,n , R 2,n , C 1,n , C 2,n and R 0,n respectively represent the open circuit voltage, the first order polarization resistance, the second order polarization resistance, the first order polarization capacitance, the second order polarization capacitance and the internal resistance of the relaxation voltage curve U n of the nth sample in the training set. S306: Calculate the equivalent circuit physical parameters of all N reconstructed relaxation voltage curves in the training set, i.e., the relaxation voltage formula of the second-order RC equivalent circuit model described in step S304, use the nonlinear least squares method to fit all N reconstructed relaxation voltage curves in the training set, obtain the open circuit voltage, first-order polarization resistance, second-order polarization resistance, first-order polarization capacitance, and second-order polarization capacitance corresponding to all N reconstructed relaxation voltage curves in the training set, and further obtain the internal resistance corresponding to all N reconstructed relaxation voltage curves in the training set according to the calculation formula of the internal resistance of the second-order RC equivalent circuit model described in step S304; the reconstructed relaxation voltage curve of the n-th sample in the training set for example, the equivalent circuit physical parameters as follows: wherein, and respectively represent the reconstructed relaxation voltage curve of the n-th sample in the training set the corresponding open circuit voltage, first order polarization resistance, second order polarization resistance, first order polarization capacitance, second order polarization capacitance and internal resistance; S307: Construct the physical information reconstruction loss of the relaxation voltage The formula is as follows: wherein Phy U,n is the equivalent circuit physical parameter of the n-th sample in the training set, is the equivalent circuit physical parameter of the reconstructed relaxation voltage curve of the n-th sample in the training set, N is the number of samples of the relaxation voltage curve in the training set, denotes the square of the L2 norm; S308: Constructing a physical information automatic encoder model AE based on the relaxation voltage rv loss function loss1 of the training, i.e. by loss rv described in step S303 and linearly weighted loss1, whose calculation formula is: Wherein, δ is a linear weighting factor of loss1, and satisfies δ∈[0,1]. S309: training the relaxation voltage curve in the training set D described in step S2 using the physical information automatic encoder model AE based on the relaxation voltage Train rv , updating the weight parameters of the AE model using the back propagation algorithm based on the loss1 described in step S308, and saving the AE model after the training is completed. rv rv model.​​ 3.The lithium battery SOH estimation method based on physical information auto-encoder of claim 1, wherein, The training and obtaining of the physical information automatic encoder based on electrochemical impedance spectroscopy in step S4 are specifically as follows: S401: build an automatic encoder model AE based on electrochemical impedance spectroscopy eis The model AE is composed of an encoder network and a decoder network, and the model AE eis The encoder network and the decoder network of the model AE are both composed of fully connected layers. S402: Transfer the training set D described in step S2 to... Train The electrochemical impedance spectroscopy Z in the model AE is input into the model AE. eis The latent feature c is obtained from the encoder network, and the latent feature c is input into the model AE. eis The electrochemical impedance spectroscopy is reconstructed in the decoder network, and the relevant formulas are as follows: wherein, is a weight matrix of an encoder network of the model AE eis , is a bias vector of an encoder network of the model AE eis , denotes an activation function of an encoder network of the model AE eis , is a weight matrix of a decoder network of the model AE eis , is a bias vector of a decoder network of the model AE eis , denotes an activation function of a decoder network of the model AE eis , pred is a reconstructed electrochemical impedance spectrum of the training set; S403: Constructing a reconstruction loss loss of electrochemical impedance spectroscopy eis The formula is as follows: wherein Z m is the electrochemical impedance spectrum of the mth sample in the training set, is the reconstructed electrochemical impedance spectrum of the mth sample in the training set, M is the number of samples of the electrochemical impedance spectrum in the training set, denotes the square of the L2 norm; S404: A fractional order equivalent circuit model based on electrochemical impedance spectroscopy is constructed, which describes the electrochemical impedance spectroscopy into four parts, namely, ultra-high frequency impedance, high frequency impedance, medium frequency impedance and low frequency impedance. The ultra-high frequency impedance is composed of one inductor and one resistor in series, the high frequency impedance is composed of one constant phase element and one resistor in parallel, the medium frequency impedance is composed of one constant phase element and one resistor in parallel, and the low frequency impedance is composed of one Warburg element. The impedance formula of the constructed fractional order equivalent circuit model is specifically as follows: Wherein, ω is the angular frequency corresponding to the electrochemical impedance data, satisfying ω = 2πf, j is the imaginary unit, f is the frequency corresponding to the electrochemical impedance data, Z(ω) is the electrochemical impedance data corresponding to the angular frequency ω, L is the inductance of the ultra-high frequency impedance, r0 is the resistance of the ultra-high frequency impedance, r1 is the resistance of the high frequency impedance, Q1 is the coefficient of the constant phase element of the high frequency impedance, α1 is the order of the constant phase element of the high frequency impedance, r2 is the resistance of the medium frequency impedance, Q2 is the coefficient of the constant phase element of the medium frequency impedance, α2 is the order of the constant phase element of the medium frequency impedance, W is the coefficient of the Warburg element of the low frequency impedance, and β is the order of the Warburg element of the low frequency impedance. The value range of α1 and α2 is (0, 1); S405: The equivalent circuit physical parameters of all M electrochemical impedance spectroscopies in the training set are calculated, that is, the impedance formula of the fractional order equivalent circuit model described in step S404 is used to fit all M electrochemical impedance spectroscopies in the training set, so as to obtain the inductance of the ultra-high frequency impedance, the resistance of the ultra-high frequency impedance, the resistance of the high frequency impedance, the coefficient of the constant phase element of the high frequency impedance, the order of the constant phase element of the high frequency impedance, the resistance of the medium frequency impedance, the coefficient of the constant phase element of the medium frequency impedance, the order of the constant phase element of the medium frequency impedance, the coefficient of the Warburg element of the low frequency impedance and the order of the Warburg element of the low frequency impedance corresponding to all M electrochemical impedance spectroscopies in the training set; EIS Z of the mth sample in the training set m For example, the equivalent circuit physical parameters Phy Z,m As follows: Phy Z,m = [L m ,r 0,m ,r 1,m ,Q 1,m ,α 1,m ,r 2,m ,Q 2,m ,α 2,m ,W m ,β m ] Among them, L m r 0,m r 1,m W 1,m α 1,m r 2,m Q 2,m α 2,m W m and β m The following are the electrochemical impedance spectra Z of the m-th sample in the training set. m The corresponding inductance of ultra-high frequency impedance, resistance of ultra-high frequency impedance, resistance of high frequency impedance, coefficient of constant phase element of high frequency impedance, order of constant phase element of high frequency impedance, resistance of intermediate frequency impedance, coefficient of constant phase element of intermediate frequency impedance, order of constant phase element of intermediate frequency impedance, coefficient of Warburg element of low frequency impedance and order of Warburg element of low frequency impedance. S406: The equivalent circuit physical parameters of all M reconstructed electrochemical impedance spectroscopies in the training set are calculated, that is, the impedance formula of the fractional order equivalent circuit model described in step S404 is used to fit all M reconstructed electrochemical impedance spectroscopies in the training set, so as to obtain the inductance of the ultra-high frequency impedance, the resistance of the ultra-high frequency impedance, the resistance of the high frequency impedance, the coefficient of the constant phase element of the high frequency impedance, the order of the constant phase element of the high frequency impedance, the resistance of the medium frequency impedance, the coefficient of the constant phase element of the medium frequency impedance, the order of the constant phase element of the medium frequency impedance, the coefficient of the Warburg element of the low frequency impedance and the order of the Warburg element of the low frequency impedance corresponding to all M reconstructed electrochemical impedance spectroscopies in the training set; reconstructed electrochemical impedance spectrum of the mth sample in the training set for example, the equivalent circuit physical parameters as follows: wherein, and EISm represents the reconstructed electrochemical impedance spectroscopy of the mth sample in the training set the inductance of the corresponding ultra-high frequency impedance, the resistance of the ultra-high frequency impedance, the resistance of the high frequency impedance, the coefficient of the constant phase element of the high frequency impedance, the order of the constant phase element of the high frequency impedance, the resistance of the medium frequency impedance, the coefficient of the constant phase element of the medium frequency impedance, the order of the constant phase element of the medium frequency impedance, the coefficient of the Warburg element of the low frequency impedance and the order of the Warburg element of the low frequency impedance; S407: Constructing a physical information reconstruction loss of the electrochemical impedance spectrum The formula is as follows: wherein PhyZ, m denotes the equivalent circuit physical parameters of the electrochemical impedance spectroscopy of the mth sample in the training set, denotes the equivalent circuit physical parameters of the reconstructed electrochemical impedance spectroscopy of the mth sample in the training set, M is the number of samples of the electrochemical impedance spectroscopy in the training set, denotes the square of the L2 norm; S408: Constructing a physical information automatic encoder model AE based on electrochemical impedance spectroscopy eis loss2, i.e. loss eis described in step S403 and loss2 described in step S407 loss2 is obtained by linear weighting, and the calculation formula is: Wherein, ε is a linear weighting factor of loss2, satisfying ε∈[0,1]; S409: using the training set D described in step S2 to train the electrochemical impedance spectrum automatic encoder model AE based on the physical information of the electrochemical impedance spectrum Train eis , based on the training loss function loss2 described in step S408, updating the weight parameters of the AE model using the back propagation algorithm, and saving the AE model after training is completed. eis eis ​​​

Images