A multi-view lithium battery SOH estimation method based on electrochemical impedance spectroscopy data

Abstract

The present application discloses a multi-view lithium battery SOH estimation method based on electrochemical impedance spectroscopy data, the method comprising: selecting multiple lithium batteries of the same type for charge and discharge cycle testing, pausing and collecting impedance data at a specific SOC during each charge and discharge cycle; generating four different electrochemical impedance spectroscopy views based on the impedance data of the training set; using a deep learning feature extraction network to extract features; constructing a low-rank optimization model and introducing a consistency regularization term to fuse the extracted feature matrix; building an artificial neural network model for SOH estimation; using a test set to perform performance verification on the trained model; during actual operation, the lithium battery management system uses the SOH estimation model to perform online SOH estimation. The present invention constructs multi-angle views based on electrochemical impedance spectroscopy data, mines the battery health information contained in the electrochemical impedance spectroscopy data from a global perspective and a local perspective, and improves the accuracy of SOH estimation.

Description

Technical Field

[0001] The present invention relates to the technical field of battery energy storage, and in particular to a multi-view lithium battery SOH estimation method based on electrochemical impedance spectroscopy data. Background Art

[0002] A lot of research has been conducted at home and abroad on how to accurately estimate the SOH of lithium batteries. Among them, electrochemical impedance spectroscopy data can reflect the electrochemical process inside the battery and is widely used. There are two main methods for feature extraction of electrochemical impedance spectroscopy modes, namely, manual feature extraction and automatic feature extraction. Manual feature extraction relies heavily on the researchers' prior knowledge and experience, which makes it difficult to popularize. When using neural networks, artificial intelligence and other methods to automatically extract features, the features in the data can be learned through multi-layer nonlinear transformations of neural networks, as electrochemical impedance spectroscopy data contains complex electrochemical information. The representation of these features does not rely on prior knowledge and feature engineering, and can be optimized and adapted based on large amounts of data.

[0003] However, the current method of using electrochemical impedance spectroscopy to estimate the SOH of lithium batteries generally uses a single angle view or real and imaginary data at a specific frequency as the input of the estimation model. It does not form a multi-angle view by combining different forms of inputs, lacks substantial fusion operations, and is difficult to fully mine the battery health information contained in the electrochemical impedance spectroscopy data. At the same time, most existing studies treat the neural network for automatic feature extraction as a black box, and do not fully consider the consistency between the input composed of different expressions of electrochemical impedance spectroscopy data and the characteristics within the battery, and cannot fully describe battery aging. Summary of the invention

[0004] In order to overcome the problems existing in the above-mentioned existing method for estimating SOH of lithium batteries using electrochemical impedance spectroscopy, the present invention proposes a multi-view lithium battery SOH estimation method based on electrochemical impedance spectroscopy data.

[0005] The primary purpose of the present invention is to solve the above technical problems. The technical solution of the present invention is as follows:

[0006] S1: Select multiple lithium batteries of the same type for charge and discharge cycle tests, pause at a specific SOC during each charge and discharge cycle and collect impedance data, including impedance data before charging, impedance data when charging to 50% SOC, impedance data when charging to 80% SOC, impedance data when charging is completed, impedance data after standing for 15 minutes after charging, impedance data before discharging, impedance data when discharging to 80% SOC, impedance data when discharging to 50% SOC, impedance data when discharging is completed, and impedance data after standing for 15 minutes after discharging; form the above data into a data set, and divide the data set into a training set and a test set at a ratio of 70% and 30%;

[0007] S2: Generate four different electrochemical impedance spectroscopy views based on the impedance data of the training set, namely Nyquist plot, phase-frequency plot, impedance data 2D plot, and phase-frequency data 2D plot; with the goal of showing a global perspective, use the real and imaginary parts of the impedance to form a Nyquist plot, and use the relationship between the phase angle and the frequency to form a phase-frequency plot; with the goal of showing local information, use the Markov transfer field method to form impedance data 2D plots and phase-frequency data 2D plots based on impedance data and phase angle data, respectively, to reveal periodic patterns, trend changes, and anomalies in the data;

[0008] S3: inputting the four electrochemical impedance spectroscopy views in step S2 into a deep learning feature extraction network respectively, and the network outputs feature matrices of the four electrochemical impedance spectroscopy views to capture features at different levels in the image, wherein the deep learning feature extraction network is composed of a residual learning layer, a convolutional neural network, and an attention layer;

[0009] S4: using a low-rank optimization model to fuse the feature matrices of the four electrochemical impedance spectroscopy views described in step S3, the specific steps include constructing a low-rank optimization model, introducing a consistency regularization term to capture the consistency between different electrochemical impedance spectroscopy views, and then solving the minimum value of the objective function through a continuous least squares optimization algorithm, and finally obtaining a fused feature matrix;

[0010] S5: Build a SOH estimation model based on artificial neural network;

[0011] S6: Use the training set to train the model, use the test set to verify the performance of the trained model, use the root mean square error and mean absolute error as indicators to evaluate the model performance. If the performance does not meet the requirements, retrain. If it meets the requirements, save the trained model to the lithium battery management system;

[0012] S7: During actual operation, the lithium battery management system collects the impedance data of the lithium battery, and uses the method described in step S2 to generate four electrochemical impedance spectrum views, namely, the Nyquist diagram, the phase-frequency diagram, the impedance data two-dimensional diagram, and the phase-frequency data two-dimensional diagram. The four electrochemical impedance spectrum views are input into the feature extraction network to obtain a fused feature matrix, and then the SOH estimation model trained in step S6 is used to perform online SOH estimation.

[0013] In this scheme, step S2 generates four different electrochemical impedance spectroscopy views according to the impedance data of the training set, and the specific method is as follows:

[0014] S2-1: Based on all the impedance data collected at each SOC pause point of each battery during each charge and discharge cycle, a one-dimensional impedance array Z = [R1+jX1, R2+jX2, ..., R n +jX n ], specifically including impedance data before charging, impedance data when charging to 50% SOC, impedance data when charging to 80% SOC, impedance data when charging is completed, impedance data after standing for 15 minutes after charging, impedance data before discharging, impedance data when discharging to 80% SOC, impedance data when discharging to 50% SOC, impedance data when discharging is completed, and impedance data after standing for 15 minutes after discharging;

[0015] S2-2: Based on the one-dimensional impedance array in step S2-1, according to the real part of the impedance R=[R1, R2, ..., R n ] and the imaginary part X=[X1,X2…,X n ], draw Nyquist plots on the complex plane;

[0016] S2-3: Based on the one-dimensional impedance array described in step S2-1, calculate the phase angle corresponding to each impedance, and draw a phase-frequency diagram according to the change of the impedance phase angle with frequency;

[0017] S2-4: Based on the one-dimensional impedance array in step S2-1, construct an impedance data interval according to the maximum value and the minimum value of the real part of the impedance data in the one-dimensional impedance array, and divide the interval into four impedance data sub-intervals on average;

[0018] S2-5: According to each SOC pause point described in step S2-1, for each frequency, the number of transfers of the real part value corresponding to the frequency in adjacent SOC pause points between different impedance data sub-intervals is counted, and a Markov transfer field matrix is ​​constructed based on this to form a two-dimensional impedance data graph of the relationship between impedance and frequency. The specific representation of the Markov transfer field matrix S of the impedance data is as follows:

[0019]

[0020] Among them, α 11 Indicates the frequency of internal transfer within the first impedance data subinterval, α 12 Indicates the frequency of the transfer from the first impedance data sub-interval to the second impedance data sub-interval, α 21 It indicates the frequency of the transfer from the second impedance data sub-interval to the first impedance data sub-interval, and the meanings of the remaining elements are similar;

[0021] S2-6: Use the phase angle corresponding to each impedance in step S2-3 to construct a one-dimensional phase angle array, that is, δ = [δ1, δ2, ..., δ n ], construct a phase angle data interval according to the maximum and minimum values ​​of the phase angle data in the one-dimensional phase angle array, and divide the data interval into four phase angle data sub-intervals on average;

[0022] S2-7: According to each SOC pause point described in step S2-1, for each frequency, the number of times the phase angle value corresponding to the frequency in adjacent SOC pause points is transferred between different phase angle data sub-intervals is counted, and a Markov transfer field matrix is ​​constructed based on this to form a two-dimensional graph of phase angle data showing the relationship between phase angle and frequency. The specific representation of the Markov transfer field matrix S' of phase angle data is as follows:

[0023]

[0024] Among them, ε 11 Indicates the frequency of internal transfer of the first phase angle data subinterval, ε 12 Indicates the frequency of the first phase angle data subinterval to the second phase angle data subinterval, ε 21 It indicates the frequency of the second phase angle data sub-interval transferring to the first phase angle data sub-interval, and the meanings of the remaining elements are similar.

[0025] In this solution, the deep learning feature extraction network described in step S3 is composed of a residual learning layer, a convolutional neural network, and an attention layer. The specific extraction steps are as follows:

[0026] S3-1: Input the four views constructed in step S2 into the residual learning layer respectively. The residual learning layer consists of two convolutional layers and a residual block, and finally outputs four feature matrices. The specific calculation formula is as follows:

[0027] Conv1=σ(W1*x+a1)

[0028] Conv2=W2*Conv1+a2

[0029] y adjust =W adjust *x+a adjust

[0030] res=σ(y adjust +Conv2)

[0031] Among them, Conv1 and Conv2 are two convolutional layers, σ is the activation function, RELU function is used here, W1 and W2 are weights, * represents the convolution operation, x is the input electrochemical impedance spectroscopy view, a1 and a2 are biases, y adjust is the adjusted output, W adjust is the adjusted weight, a adjust is the adjusted bias, res is the output of the residual block;

[0032] S3-2: Input the feature matrix output by the residual learning layer described in S3-1 into a convolutional neural network consisting of three convolutional layers and one pooling layer, whose output is defined as p i , i=1,2,3,4;

[0033] S3-3: The convolutional neural network output p described in S3-2 i Input to the attention layer, the expression defining the attention layer is:

[0034] c i =[β(W attent i on *p i +a attent i on )]⊙p i

[0035] Among them, β is the activation function, here we use the sigmoid function, c i represents the output of the attention layer, i=1,2,3,4, W attention represents the weight of the attention layer, a attention represents the attention layer bias;

[0036] S3-4: The output c of the attention layer described in S3-3 i Input into the residual learning layer described in S3-1, update the residual block to obtain a new residual block, and the specific calculation formula is:

[0037] res'=σ(W adjust *c i +a adjust +W2*σ(W1*c i +a1)+a2)

[0038] Among them, res' represents the output of the new residual block;

[0039] S3-5: The output res' of the new residual block described in S3-4 is concatenated with the output ci of the attention layer described in S3-3, and finally the feature matrices of the four electrochemical impedance spectroscopy views are obtained. The specific calculation formula is:

[0040] F i =Concatenate([c i ,res'])

[0041] Among them, F i is the characteristic matrix of the electrochemical impedance spectroscopy view, i=1,2,3,4, and Concatenate represents the concatenation operation.

[0042] In this solution, the low-rank optimization model described in step S4 is used to fuse the feature matrices of the four electrochemical impedance spectroscopy views described in step S3. The specific method is as follows:

[0043] S4-1: Construct a low-rank optimization model, which consists of an objective function and constraints that introduce a consistency regularization term;

[0044] The calculation formula for defining the objective function of introducing the consistency regularization term is as follows:

[0045]

[0046] Among them, λ is the penalty factor, F i The characteristic matrix of the electrochemical impedance spectrum view output in step S3-5, Y i F i The low-rank matrix, H i F i The residual matrix of i || 2,1 Indicates H i The row sparse norm of , Ω is the parameter that controls the weight of the consistency regularization term, ||(1-Ω)Y i +ΩY mean || * Represents (1-Ω)Y i +ΩY mean The nuclear norm of , Y represents the consistency regularization term, Represents Y i -Y mean The Frobenius norm of mean is the average value of all low-rank matrices;

[0047] Define the constraint condition as F i =F i Y i +H i ;

[0048] S4-2: Use the continuous least squares optimization algorithm to solve the minimum value of the objective function, and update the F output in step S4-1 during the iteration process. i , Y i , H i To minimize the value of the objective function and extract the updated matrix Y i With H i ;

[0049] S4-3: All Y updated after the step described in S4-2 i , H i The matrices are concatenated column by column to form a fused low-rank matrix and a fused residual matrix, and the fused low-rank matrix and the fused residual matrix are added together to finally output a fused feature matrix. The calculation formula is as follows:

[0050] Y fuse =concatenate(Y1,…,Y i )

[0051] H fuse =concatenate(Y1,…,Y i )

[0052] F final =Y fuse +H fuse

[0053] Among them, Y fuse represents the fused low-rank matrix, H fuse represents the residual matrix after fusion, F final Represents the fused feature matrix.

[0054] In this solution, during the actual operation described in step S7, the SOH estimation model is used to perform online SOH estimation, and the specific steps are as follows:

[0055] S7-1: Deploy the trained SOH estimation model into the battery management system;

[0056] S7-2: The lithium battery management system collects the impedance data of the lithium battery, and uses the method described in step S2 to generate four electrochemical impedance spectrum views, namely, the Nyquist diagram, the phase-frequency diagram, the impedance data two-dimensional diagram, and the phase-frequency data two-dimensional diagram. The four electrochemical impedance spectrum views are input into the feature extraction network to obtain a fused feature matrix, and then the SOH estimation model is used to perform online SOH estimation.

[0057] Compared with the prior art, the technical solution of the present invention has the following beneficial effects:

[0058] The present invention proposes a multi-view lithium battery SOH estimation method based on electrochemical impedance spectroscopy data. Based on the electrochemical impedance spectroscopy data, a multi-angle view is constructed, which can fully mine the battery health information contained in the electrochemical impedance spectroscopy data from a global and local perspective. At the same time, by fusing views from different angles for feature extraction, the consistency between views from different angles is captured while maintaining the feature information of each view, thereby improving the accuracy of SOH estimation. BRIEF DESCRIPTION OF THE DRAWINGS

[0059] Figure 1 This is a flow chart of a multi-view lithium battery SOH estimation method based on electrochemical impedance spectroscopy data proposed in the present invention. DETAILED DESCRIPTION

[0060] 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.

[0061] 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.

[0062] In a specific embodiment, Figure 1 As shown, a multi-view lithium battery SOH estimation method based on electrochemical impedance spectroscopy data includes the following steps:

[0063] S1: Select multiple lithium batteries of the same type for charge and discharge cycle tests, pause at a specific SOC during each charge and discharge cycle and collect impedance data, including impedance data before charging, impedance data when charging to 50% SOC, impedance data when charging to 80% SOC, impedance data when charging is completed, impedance data after standing for 15 minutes after charging, impedance data before discharging, impedance data when discharging to 80% SOC, impedance data when discharging to 50% SOC, impedance data when discharging is completed, and impedance data after standing for 15 minutes after discharging; form the above data into a data set, and divide the data set into a training set and a test set at a ratio of 70% and 30%;

[0064] S2: Generate four different electrochemical impedance spectroscopy views based on the impedance data of the training set, namely Nyquist plot, phase-frequency plot, impedance data 2D plot, and phase-frequency data 2D plot; with the goal of showing a global perspective, use the real and imaginary parts of the impedance to form a Nyquist plot, and use the relationship between the phase angle and the frequency to form a phase-frequency plot; with the goal of showing local information, use the Markov transfer field method to form impedance data 2D plots and phase-frequency data 2D plots based on impedance data and phase angle data, respectively, to reveal periodic patterns, trend changes, and anomalies in the data;

[0065] S3: inputting the four electrochemical impedance spectroscopy views in step S2 into a deep learning feature extraction network respectively, and the network outputs feature matrices of the four electrochemical impedance spectroscopy views to capture features at different levels in the image, wherein the deep learning feature extraction network is composed of a residual learning layer, a convolutional neural network, and an attention layer;

[0066] S4: using a low-rank optimization model to fuse the feature matrices of the four electrochemical impedance spectroscopy views described in step S3, the specific steps include constructing a low-rank optimization model, introducing a consistency regularization term to capture the consistency between different electrochemical impedance spectroscopy views, and then solving the minimum value of the objective function through a continuous least squares optimization algorithm, and finally obtaining a fused feature matrix;

[0067] S5: Build a SOH estimation model based on artificial neural network;

[0068] S6: Use the training set to train the model, use the test set to verify the performance of the trained model, use the root mean square error and mean absolute error as indicators to evaluate the model performance. If the performance does not meet the requirements, retrain. If it meets the requirements, save the trained model to the lithium battery management system;

[0069] S7: During actual operation, the lithium battery management system collects the impedance data of the lithium battery, and uses the method described in step S2 to generate four electrochemical impedance spectrum views, namely, the Nyquist diagram, the phase-frequency diagram, the impedance data two-dimensional diagram, and the phase-frequency data two-dimensional diagram. The four electrochemical impedance spectrum views are input into the feature extraction network to obtain a fused feature matrix, and then the SOH estimation model trained in step S6 is used to perform online SOH estimation.

[0070] In this scheme, step S2 generates four different electrochemical impedance spectroscopy views according to the impedance data of the training set, and the specific method is as follows:

[0071] S2-1: Based on all the impedance data collected at each SOC pause point of each battery during each charge and discharge cycle, a one-dimensional impedance array Z = [R1+jX1, R2+jX2, ..., R n +jX n], specifically including impedance data before charging, impedance data when charging to 50% SOC, impedance data when charging to 80% SOC, impedance data when charging is completed, impedance data after standing for 15 minutes after charging, impedance data before discharging, impedance data when discharging to 80% SOC, impedance data when discharging to 50% SOC, impedance data when discharging is completed, and impedance data after standing for 15 minutes after discharging;

[0072] S2-2: Based on the one-dimensional impedance array in step S2-1, according to the real part of the impedance R=[R1, R2, ..., R n ] and the imaginary part X=[X1,X2…,X n ], draw Nyquist plots on the complex plane;

[0073] S2-3: Based on the one-dimensional impedance array described in step S2-1, calculate the phase angle corresponding to each impedance, and draw a phase-frequency diagram according to the change of the impedance phase angle with frequency;

[0074] S2-4: Based on the one-dimensional impedance array in step S2-1, construct an impedance data interval according to the maximum value and the minimum value of the real part of the impedance data in the one-dimensional impedance array, and divide the interval into four impedance data sub-intervals on average;

[0075] S2-5: According to each SOC pause point described in step S2-1, for each frequency, the number of transfers of the real part value corresponding to the frequency in adjacent SOC pause points between different impedance data sub-intervals is counted, and a Markov transfer field matrix is ​​constructed based on this to form a two-dimensional impedance data graph of the relationship between impedance and frequency. The specific representation of the Markov transfer field matrix S of the impedance data is as follows:

[0076]

[0077] Among them, α 11 Indicates the frequency of internal transfer within the first impedance data subinterval, α 12 Indicates the frequency of the transfer from the first impedance data sub-interval to the second impedance data sub-interval, α 21 It indicates the frequency of the transfer from the second impedance data sub-interval to the first impedance data sub-interval, and the meanings of the remaining elements are similar;

[0078] S2-6: Use the phase angle corresponding to each impedance in step S2-3 to construct a one-dimensional phase angle array, that is, δ = [δ1, δ2, ..., δ n ], construct a phase angle data interval according to the maximum and minimum values ​​of the phase angle data in the one-dimensional phase angle array, and divide the data interval into four phase angle data sub-intervals on average;

[0079] S2-7: According to each SOC pause point described in step S2-1, for each frequency, the number of times the phase angle value corresponding to the frequency in adjacent SOC pause points is transferred between different phase angle data sub-intervals is counted, and a Markov transfer field matrix is ​​constructed based on this to form a two-dimensional graph of phase angle data of the relationship between phase angle and frequency. The specific representation of the Markov transfer field matrix S' is as follows:

[0080]

[0081] Among them, ε 11 Indicates the frequency of internal transfer of the first phase angle data subinterval, ε 12 Indicates the frequency of the first phase angle data subinterval to the second phase angle data subinterval, ε 21 It indicates the frequency of the second phase angle data sub-interval transferring to the first phase angle data sub-interval, and the meanings of the remaining elements are similar.

[0082] In this solution, the deep learning feature extraction network described in step S3 is composed of a residual learning layer, a convolutional neural network, and an attention layer. The specific extraction steps are as follows:

[0083] S3-1: Input the four views constructed in step S2 into the residual learning layer respectively. The residual learning layer consists of two convolutional layers and a residual block, and finally outputs four feature matrices. The specific calculation formula is as follows:

[0084] Conv1=σ(W1*x+a1)

[0085] Conv2=W2*Conv1+a2

[0086] y adjust =W adjust *x+a adjust

[0087] res=σ(y adjust +Conv2)

[0088] Among them, Conv1 and Conv2 are two convolutional layers, σ is the activation function, RELU function is used here, W1 and W2 are weights, * represents the convolution operation, x is the input electrochemical impedance spectroscopy view, a1 and a2 are biases, y adjust is the adjusted output, W adjust is the adjusted weight, a adjust is the adjusted bias, res is the output of the residual block;

[0089] S3-2: Input the feature matrix output by the residual learning layer described in S3-1 into a convolutional neural network consisting of three convolutional layers and one pooling layer, whose output is defined as p i , i=1,2,3,4;

[0090] S3-3: The convolutional neural network output p described in S3-2 i Input to the attention layer, the expression defining the attention layer is:

[0091] c i =[β(W attention *p i +a attent i on )]⊙p i

[0092] Among them, β is the activation function, here we use the sigmoid function, c i represents the output of the attention layer, i=1,2,3,4, W attention represents the weight of the attention layer, a attention represents the attention layer bias;

[0093] S3-4: The output c of the attention layer described in S3-3 i Input into the residual learning layer described in S3-1, update the residual block to obtain a new residual block, and the specific calculation formula is:

[0094] res'=σ(W adjust *c i +a adjust +W2*σ(W1*c i +a1)+a2)

[0095] Among them, res' represents the output of the new residual block;

[0096] S3-5: Combine the output res' of the new residual block described in S3-4 with the output c of the attention layer described in S3-3 i After splicing, the characteristic matrix of four electrochemical impedance spectroscopy views is finally obtained. The specific calculation formula is:

[0097] F i =Concatenate([c i ,res'])

[0098] Among them, F i is the characteristic matrix of the electrochemical impedance spectroscopy view, i=1,2,3,4, and Concatenate represents the concatenation operation.

[0099] In this solution, the low-rank optimization model described in step S4 is used to fuse the feature matrices of the four electrochemical impedance spectroscopy views described in step S3. The specific method is as follows:

[0100] S4-1: Construct a low-rank optimization model, which consists of an objective function and constraints that introduce a consistency regularization term;

[0101] The calculation formula for defining the objective function of introducing the consistency regularization term is as follows:

[0102]

[0103] Among them, λ is the penalty factor, F i The characteristic matrix of the electrochemical impedance spectrum view output in step S3-5, Y i F i The low-rank matrix, H i F i The residual matrix of i || 2,1 Indicates H i The row sparse norm of , Ω is the parameter that controls the weight of the consistency regularization term, ||(1-Ω)Y i +ΩY mean || * Represents (1-Ω)Y i +ΩY mean The nuclear norm of , Y represents the consistency regularization term, Represents Y i -Y mean The Frobenius norm of mean is the average value of all low-rank matrices;

[0104] Define the constraint condition as F i =F i Y i +H i ;

[0105] S4-2: Use the continuous least squares optimization algorithm to solve the minimum value of the objective function, and update the F output in step S4-1 during the iteration process. i , Y i , H i To minimize the value of the objective function and extract the updated matrix Y i With H i ;

[0106] S4-3: All Y updated after the step described in S4-2 i , H i The matrices are concatenated column by column to form a fused low-rank matrix and a fused residual matrix, and the fused low-rank matrix and the fused residual matrix are added together to finally output a fused feature matrix. The calculation formula is as follows:

[0107] Y fuse =concatenate(Y1,…,Y i )

[0108] H fuse =concatenate(Y1,…,Y i )

[0109] F final =Y fuse +H fuse

[0110] Among them, Y fuse represents the fused low-rank matrix, H fuse represents the residual matrix after fusion, F final Represents the fused feature matrix.

[0111] In this solution, during the actual operation described in step S7, the SOH estimation model is used to perform online SOH estimation, and the specific steps are as follows:

[0112] S7-1: Deploy the trained SOH estimation model into the battery management system;

[0113] S7-2: The lithium battery management system collects the impedance data of the lithium battery, and uses the method described in step S2 to generate four electrochemical impedance spectrum views, namely, the Nyquist diagram, the phase-frequency diagram, the impedance data two-dimensional diagram, and the phase-frequency data two-dimensional diagram. The four electrochemical impedance spectrum views are input into the feature extraction network to obtain a fused feature matrix, and then the SOH estimation model is used to perform online SOH estimation.

Claims

1. A multi-view lithium battery SOH estimation method based on electrochemical impedance spectroscopy data, characterized in that: The specific steps are as follows: S1: Select multiple lithium batteries of the same type for charge and discharge cycle tests, pause at a specific SOC during each charge and discharge cycle and collect impedance data, including impedance data before charging, impedance data when charging to 50% SOC, impedance data when charging to 80% SOC, impedance data when charging is completed, impedance data after standing for 15 minutes after charging, impedance data before discharging, impedance data when discharging to 80% SOC, impedance data when discharging to 50% SOC, impedance data when discharging is completed, and impedance data after standing for 15 minutes after discharging; form the above data into a data set, and divide the data set into a training set and a test set at a ratio of 70% and 30%; S2: Generate four different electrochemical impedance spectroscopy views based on the impedance data of the training set, namely Nyquist plot, phase-frequency plot, impedance data 2D plot, and phase-frequency data 2D plot; with the goal of showing a global perspective, use the real and imaginary parts of the impedance to form a Nyquist plot, and use the relationship between the phase angle and the frequency to form a phase-frequency plot; with the goal of showing local information, use the Markov transfer field method to form impedance data 2D plots and phase-frequency data 2D plots based on impedance data and phase angle data, respectively, to reveal periodic patterns, trend changes, and anomalies in the data; S3: inputting the four electrochemical impedance spectroscopy views in step S2 into a deep learning feature extraction network respectively, and the network outputs feature matrices of the four electrochemical impedance spectroscopy views to capture features at different levels in the image, wherein the deep learning feature extraction network is composed of a residual learning layer, a convolutional neural network, and an attention layer; S4: using a low-rank optimization model to fuse the feature matrices of the four electrochemical impedance spectroscopy views described in step S3, the specific steps include constructing a low-rank optimization model, introducing a consistency regularization term to capture the consistency between different electrochemical impedance spectroscopy views, and then solving the minimum value of the objective function through a continuous least squares optimization algorithm, and finally obtaining a fused feature matrix; S5: Build a SOH estimation model based on artificial neural network; S6: Use the training set to train the model, use the test set to verify the performance of the trained model, use the root mean square error and mean absolute error as indicators to evaluate the model performance. If the performance does not meet the requirements, retrain. If it meets the requirements, save the trained model to the lithium battery management system; S7: During actual operation, the lithium battery management system collects the impedance data of the lithium battery, and uses the method described in step S2 to generate four electrochemical impedance spectrum views, namely, the Nyquist diagram, the phase-frequency diagram, the impedance data two-dimensional diagram, and the phase-frequency data two-dimensional diagram. The four electrochemical impedance spectrum views are input into the feature extraction network to obtain a fused feature matrix, and then the SOH estimation model trained in step S6 is used to perform online SOH estimation.

2. A multi-view lithium battery SOH estimation method based on electrochemical impedance spectroscopy data according to claim 1, characterized in that: In step S2, four different electrochemical impedance spectroscopy views are generated according to the impedance data of the training set. The specific method is as follows: S2-1: Based on all the impedance data collected at each SOC pause point of each battery during each charge and discharge cycle, a one-dimensional impedance array Z = [R1+jX1, R2+jX2, ..., R n +jX n ], specifically including impedance data before charging, impedance data when charging to 50% SOC, impedance data when charging to 80% SOC, impedance data when charging is completed, impedance data after standing for 15 minutes after charging, impedance data before discharging, impedance data when discharging to 80% SOC, impedance data when discharging to 50% SOC, impedance data when discharging is completed, and impedance data after standing for 15 minutes after discharging; S2-2: Based on the one-dimensional impedance array in step S2-1, according to the real part of the impedance R=[R1, R2, ..., R n ] and the imaginary part X=[X1,X2…,X n ], draw Nyquist plots on the complex plane; S2-3: Based on the one-dimensional impedance array described in step S2-1, calculate the phase angle corresponding to each impedance, and draw a phase-frequency diagram according to the change of the impedance phase angle with frequency; S2-4: Based on the one-dimensional impedance array in step S2-1, construct an impedance data interval according to the maximum value and the minimum value of the real part of the impedance data in the one-dimensional impedance array, and divide the interval into four impedance data sub-intervals on average; S2-5: According to each SOC pause point described in step S2-1, for each frequency, the number of transfers of the real part value corresponding to the frequency in adjacent SOC pause points between different impedance data sub-intervals is counted, and a Markov transfer field matrix is ​​constructed based on this to form a two-dimensional impedance data graph of the relationship between impedance and frequency. The specific representation of the Markov transfer field matrix S of the impedance data is as follows: Among them, α 11 Indicates the frequency of internal transfer within the first impedance data subinterval, α 12 Indicates the frequency of the transfer from the first impedance data sub-interval to the second impedance data sub-interval, α 21 It indicates the frequency of the transfer from the second impedance data sub-interval to the first impedance data sub-interval, and the meanings of the remaining elements are similar; S2-6: Use the phase angle corresponding to each impedance in step S2-3 to construct a one-dimensional phase angle array, that is, δ = [δ1, δ2, ..., δ n ], construct a phase angle data interval according to the maximum and minimum values ​​of the phase angle data in the one-dimensional phase angle array, and divide the data interval into four phase angle data sub-intervals on average; S2-7: According to each SOC pause point described in step S2-1, for each frequency, the number of times the phase angle value corresponding to the frequency in adjacent SOC pause points is transferred between different phase angle data sub-intervals is counted, and a Markov transfer field matrix is ​​constructed based on this to form a two-dimensional graph of phase angle data showing the relationship between phase angle and frequency. The specific representation of the Markov transfer field matrix S' of phase angle data is as follows: Among them, ε 11 Indicates the frequency of internal transfer of the first phase angle data subinterval, ε 12 Indicates the frequency of the first phase angle data subinterval to the second phase angle data subinterval, ε 21 It indicates the frequency of the second phase angle data sub-interval transferring to the first phase angle data sub-interval, and the meanings of the remaining elements are similar.

3. The multi-view lithium battery SOH estimation method based on electrochemical impedance spectroscopy data according to claim 1, characterized in that: The deep learning feature extraction network in step S3 is composed of a residual learning layer, a convolutional neural network, and an attention layer. The specific extraction steps are as follows: S3-1: Input the four views constructed in step S2 into the residual learning layer respectively. The residual learning layer consists of two convolutional layers and a residual block, and finally outputs four feature matrices. The specific calculation formula is as follows: Conv1=σ(W1*x+a1) Conv2=W2*Conv1+a2 y adjust =W adjust *x+a adjust res=σ(y adjust +Conv2) Among them, Conv1 and Conv2 are two convolutional layers, σ is the activation function, RELU function is used here, W1 and W2 are weights, * represents the convolution operation, x is the input electrochemical impedance spectroscopy view, a1 and a2 are biases, y adjust is the adjusted output, W adjust is the adjusted weight, a adjust is the adjusted bias, res is the output of the residual block; S3-2: Input the feature matrix output by the residual learning layer described in S3-1 into a convolutional neural network consisting of three convolutional layers and one pooling layer, whose output is defined as p i , i=1,2,3,4; S3-3: The convolutional neural network output p described in S3-2 i Input to the attention layer, the expression defining the attention layer is: c i =[β(W attention *p i +a attention )]⊙p i Among them, β is the activation function, here we use the sigmoid function, c i represents the output of the attention layer, i=1,2,3,4, W attention represents the weight of the attention layer, a attention represents the attention layer bias; S3-4: The output c of the attention layer described in S3-3 i Input into the residual learning layer described in S3-1, update the residual block to obtain a new residual block, and the specific calculation formula is: res'=σ(W adjust *c i +a adjust +W2*σ(W1*c i +a1)+a2) Among them, res' represents the output of the new residual block; S3-5: Combine the output res' of the new residual block described in S3-4 with the output c of the attention layer described in S3-3 i After splicing, the characteristic matrix of four electrochemical impedance spectroscopy views is finally obtained. The specific calculation formula is: F i =Concatenate([c i ,res']) Among them, F i is the characteristic matrix of the electrochemical impedance spectroscopy view, i=1,2,3,4, and Concatenate represents the concatenation operation.

4. The multi-view lithium battery SOH estimation method based on electrochemical impedance spectroscopy data according to claim 1, characterized in that: In step S4, the low-rank optimization model is used to fuse the feature matrices of the four electrochemical impedance spectroscopy views in step S3. The specific method is as follows: S4-1: Construct a low-rank optimization model, which consists of an objective function and constraints that introduce a consistency regularization term; The calculation formula for defining the objective function of introducing the consistency regularization term is as follows: Wherein, λ is the penalty factor, Fi is the characteristic matrix of the electrochemical impedance spectrum view output in step S3-5, and Y i F i The low-rank matrix, H i F i The residual matrix of i || 2,1 Indicates H i The row sparse norm of , Ω is the parameter that controls the weight of the consistency regularization term, ||(1-Ω)Y i +ΩY mean || * Represents (1-Ω)Y i +ΩY mean The nuclear norm of , Y represents the consistency regularization term, Represents Y i -Y mean The Frobenius norm of mean is the average value of all low-rank matrices; Define the constraint condition as F i =F i Y i +H i ; S4-2: Use the continuous least squares optimization algorithm to solve the minimum value of the objective function, and update the F output in step S4-1 during the iteration process. i , Y i , H i To minimize the value of the objective function and extract the updated matrix Y i With H i ; S4-3: All Y updated after the step described in S4-2 i , H i The matrices are concatenated column by column to form a fused low-rank matrix and a fused residual matrix, and the fused low-rank matrix and the fused residual matrix are added together to finally output a fused feature matrix. The calculation formula is as follows: AND fuse =concatenate(Y1,…,Y i ) H fuse =concatenate(Y1,…,Y i ) F final =Y fuse +H fuse Among them, Y fuse represents the fused low-rank matrix, H fuse represents the residual matrix after fusion, F final Represents the fused feature matrix.

5. The multi-view lithium battery SOH estimation method based on electrochemical impedance spectroscopy data according to claim 1, characterized in that: During the actual operation described in step S7, the SOH estimation model is used to perform online SOH estimation, and the specific steps are as follows: S7-1: Deploy the trained SOH estimation model into the battery management system; S7-2: The lithium battery management system collects the impedance data of the lithium battery, and uses the method described in step S2 to generate four electrochemical impedance spectrum views, namely, the Nyquist diagram, the phase-frequency diagram, the impedance data two-dimensional diagram, and the phase-frequency data two-dimensional diagram. The four electrochemical impedance spectrum views are input into the feature extraction network to obtain a fused feature matrix, and then the SOH estimation model is used to perform online SOH estimation.

Images