Lithium ion battery SOC and internal non-uniformity estimation method based on ultrasonic transmission signal

Abstract

The invention relates to a lithium ion battery SOC and internal non-uniformity estimation method based on an ultrasonic transmission signal, and the method comprises the steps: firstly, extracting the amplitude characteristics of the ultrasonic transmission waveform of each measurement point, and forming amplitude matrixes of different SOCs; then, establishing a GPR-AL method for selecting data points and feature sets with relatively low uncertainty; and by taking the feature set as input and constructing a CNN-PELM network structure, the influence of redundant information and noise is reduced, and the accuracy of SOC estimation is improved. And finally, SOC and internal non-uniformity estimation is carried out on the lithium ion battery. According to the method, SOC and internal non-uniformity estimation is more accurately achieved by selectively utilizing data points with low uncertainty and a good relation model structure.

Description

Technical Field

[0001] The present invention relates to the field of health status prediction, in particular A lithium-ion battery SOC based on ultrasonic transmission signal Internal heterogeneity estimation method . Background Art

[0002] Lithium-ion batteries are the most popular energy storage and release carriers in recent years. Accurate estimation of their state of charge (SOC) can more efficiently manage battery usage and maximize battery energy utilization.

[0003] The application of ultrasonic technology in lithium-ion battery state estimation has attracted widespread attention due to its advantages of non-destructive detection, high accuracy and strong real-time performance. Traditional SOC estimation methods are usually based on the assumption that the SOC inside the battery is uniformly distributed and mainly rely on data from a single detection point. However, the SOC inside the battery often exhibits uneven distribution characteristics. This unevenness can cause the performance of local areas of the battery to rapidly decline, which in turn leads to potential safety hazards such as overcharging, over-discharging or thermal runaway. Therefore, in order to achieve higher-precision SOC estimation and accurately map the complex SOC uneven distribution inside the battery, technical means based on multi-point detection are particularly important and urgent. This method can not only improve the accuracy of SOC estimation, but also monitor the internal state of the battery more comprehensively and enhance the safety and reliability of the battery system. When performing ultrasonic multi-point detection on the battery, tens of thousands or even hundreds of thousands of waveform data points can be generated, and due to the influence of process manufacturing, detection environment, measurement error, polarization effect and other factors, some of the collected data will also have problems such as poor quality and high noise. How to process these massive data to extract effective features is a major technical challenge.

[0004] In view of this, it is necessary to construct a lithium-ion battery SOC and internal heterogeneity estimation method based on ultrasonic transmission signals. Summary of the invention

[0005] The purpose of the present invention is to solve the problem of the prior art in realizing the estimation of internal SOC non-uniformity of lithium-ion batteries, and proposes A method for estimating SOC and internal heterogeneity of lithium-ion batteries based on ultrasonic transmission signals This method is aimed at lithium-ion batteries. It uses the Gaussian process regression-active learning (GPR-AL) method to select data points with lower uncertainty from multi-point ultrasonic data to form a feature set, and constructs a convolutional neural network-pooling extreme learning machine (CNN-PELM) network to achieve SOC and internal heterogeneity estimation. It mainly includes the following steps:

[0006] Step 1: Ultrasonic scanning and transmission signal feature extraction;

[0007] Step 2: Selection of excellent feature points based on signal amplitude matrix;

[0008] Step 3: Construction of lithium-ion battery SOC model based on excellent feature point signal data;

[0009] Step 4: Li-ion battery SOC and internal heterogeneity estimation.

[0010] The specific process of step 1 is as follows: fully charge and discharge the lithium-ion battery, record the voltage and current data, calculate the SOC of the battery using the ampere-hour integration method, perform an ultrasonic scan on the battery surface every 10% SOC interval, extract the amplitude of the ultrasonic transmission signal at all scanning points, and construct the signal amplitude matrix MSA t,p,q , forming a data set corresponding to the features and SOC {X, y} = {MSA t,p,q ,SOC t}, t=1,...,T, where X is the feature, y is the SOC, t represents the time corresponding to the SOC, p represents the number of rows of the matrix, and q represents the number of columns of the matrix.

[0011] The specific process of step 2 is as follows: for the signal amplitude data set extracted in step 1, the GPR-AL method is used to screen points with smaller uncertainty in each amplitude matrix as excellent feature points, wherein the GPR-AL method is a Gaussian process regression-active learning method, and the method flow is as follows:

[0012] (1) Initialize the active learner for the mapping relationship between signal amplitude features and SOC: Use the Gaussian process regression model as the learner. At the tth moment, randomly select n_initial labeled data sets D = {X n_initial ,y n_initial}={(x i ,y i ),i=1,...,n_initial} are sent to the active learner for training, and considering that the actual measured data has a certain degree of noise, that is: y n_initial =f(X n_initial )+ε, where f represents Gaussian process regression; assuming that the noise ε satisfies the standard Gaussian distribution: Since the Gaussian process is described by the mean function and the covariance function, based on Bayesian theory, we know that y n_initial The prior distribution of is: where μ=[m(x1),...,m(x n_initial )] is the mean function, K ij = k(x i ,x j ) is the covariance function, also called the kernel function;

[0013] (2) Update the active learner: Use the initialized active learner to predict the standard deviation of the remaining points, sort and select the points with large standard deviations for annotation, and obtain a data set with large uncertainty {X n_instances ,y n_instances}, merge with the initial data set D and input into the active learner for retraining, and update the learner. Repeat this step until the set number of training times is reached. The specific steps are as follows:

[0014] a.Set D new ={X n_initial ,y n_initial}, remove {X from {X,y} n_initial ,y n_initial}Get the remaining data set {X rest ,y rest}, define the number of tags n_instances required in each iteration;

[0015] b. Replace {X rest ,y rest}The data set is fed into the active learner, so for the new input data X rest , let its prediction be y * , then y n_initial and * The joint distribution of can be expressed as:

[0016]

[0017] Where K = K(X n_initial ,X n_initial ) is the kernel function matrix between the training input data, K * =K(X n_initial ,X rest ) is the kernel function matrix between the training input data and the test input data, K ** =K(X rest ,X rest ) is the kernel function matrix between the test input data, m(X n_initial )=0,m(X rest )=0; Then, the predicted distribution under the new input data can be obtained through conditional distribution: The predicted mean is: The predicted covariance matrix is: Through the above calculation, we can get each X rest The predicted standard deviation is sorted and the first n_instances points with large standard deviations, that is, large uncertainties, are selected for marking, and {X n_instances ,y n_instances};

[0018] c. Next, D new With large uncertainty data {X n_instances ,y n_instances} to merge and get D new ={X new ,y new =D new ∪{X n_instances ,y n_instances}, and then feed it as a new training set to the active learner to retrain and update its performance: Then, the trained data {X n_instances ,y n_instances}From {X rest ,y rest} remove {X rest ,y rest}={X rest \X n_instances ,y rest \y n_instances}, where \ indicates deletion; repeat b to c until the set number of training times is reached;

[0019] (3) Obtaining excellent points: For the test set X 。 , let its prediction be y 。 , X 。 Input into the trained active learner to get the standard deviation corresponding to each point position:

[0020]

[0021] Sort the standard deviations of the prediction results from small to large, and select the first z points with small standard deviations as excellent points. After the above steps, we can get the excellent features in each ultrasound scan. Let t = t + 1, and repeat steps (2) to (3) until t reaches T. Stacking the excellent features at each moment in order can obtain the input-output combination corresponding to the state of charge: {SFC t ,SOC t}.

[0022] The specific process of step 3 is as follows: the excellent feature combination SFC is used as input, SOC is used as output, and the CNN-PELM neural network is used for training and testing. The CNN-PELM neural network model is a convolutional neural network-pooling extreme learning machine network model. After the input data passes through the CNN feature extraction layer, the extracted features are fitted using the PELM regression fitting layer. The network model is constructed as follows:

[0023] CNN-PELM is composed of a feature extraction layer composed of a convolutional neural network and a regression fitting layer composed of a pooling extreme learning machine: (1) First, the CNN feature extraction layer: for the input excellent feature combination SFC, it is extracted through a convolutional neural network constructed in the form of "residual connection" to obtain SFC conv : SFC conv =Resnet18(SFC), where the feature extraction layer consists of 5 parts. The first part is a convolutional layer, a normalization layer, and a pooling layer. The convolutional layer has a kernel size of 2, a stride of 2, a padding of 3, and an output channel of 64. The convolutional kernel size of the pooling layer is 3, a stride of 2, and a padding of 1. Then, the second part is two residual blocks, which have the same structure, and each residual block has 2 convolutional layers with the same structure. The convolutional kernel size of each convolutional layer is 3, the stride is 1, the padding is 1, and the output channel is 64. Secondly, the third part is two residual blocks. The input and output dimensions of the first residual block are inconsistent. Therefore, in order to match the dimensions, a convolution layer is added on top of the original two convolution layers. The convolution kernel size of the original two convolution layers is 3, the stride is 2 and 1, and the padding is 1. The convolution kernel size of the newly added third convolution layer is 1, the stride is 2, padding=0, and the output channel is 128; the input and output dimensions of the second residual block are consistent. It has two convolution layers with the same structure, the convolution kernel size is 3, the stride is 1, the padding is 1, and the output channel is 128. Subsequently, the fourth part is similar to the third part in structure. The only difference is that the output channel of the first residual block is 256, and the output channel of the second residual block is 256. Similarly, the fifth part is also similar to the third part in structure. The only difference is that the output channel of the first residual block is 512, and the output channel of the second residual block is 512. Among them, when the input and output dimensions are consistent, the residual block is represented as: y (1) =ReLu(conv(ReLu(conv(x)))+x), when the input and output dimensions are inconsistent, the residual block is represented as: (2) =ReLu(conv(ReLu(conv(x)))+conv(x)); (2) Then comes the PELM regression fitting layer: After the input data passes through the feature extraction layer, the PELM regression fitting layer is used to extract the feature SFC conv Perform fitting. T=SOC t Represents the true value of the regression fit. The calculation process of PELM is as follows:

[0024] a. Calculate the output matrix H of the hidden layer: H = mean_pool (SFC conv )

[0025] b. Next, calculate the output weight matrix β: M = (H′H) -1 , β=MH′T.

[0026] c. Therefore, the output of PELM is: Among them, ' represents the transpose of the matrix, and the connection weight β between the hidden layer and the output layer is obtained by solving the equation The least squares solution of is obtained.

[0027] The specific process of step 4 is as follows: the corresponding SOC change is derived by comparing the difference in SA values. Based on step 2, the point with smaller uncertainty in each ultrasonic scan image can be obtained. Find the mean of the feature data at these points The mean value and the estimated SOC value obtained in step 3 are As the reference characteristic value and reference SOC value respectively, according to the characteristic value SA at each location point = [SA1, ..., SA T ]=[MSA 1,p,q ,...,MSA T,p,q ] and the formula abs(SOC each_site -SOC est )=abs(SA-SA ave ) Calculate the SOC value SOC at each point each_site , and obtain the estimated SOC and internal heterogeneity of the lithium-ion battery.

[0028] Through the above steps, a method for estimating SOC and internal heterogeneity of lithium-ion batteries based on ultrasonic transmission signals is given.

[0029] The excellent effect of the present invention is that by using GPR-AL to select excellent point positions, high-quality information can be screened out from a large amount of information, while avoiding the problems of information redundancy, accidental results, and unstable model training caused by random selection. The learner that completes SOC estimation makes comprehensive use of the powerful feature representation ability of CNN and the efficient regression fitting ability and robustness of PELM when designing the structure, which can significantly improve the overfitting problem that may be caused by the fully connected layer. The difference in SOC inside the battery is derived by the difference in SA at each position point, realizing the visualization of the SOC heterogeneity inside the battery. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] Figure 1 The figure shows a flow chart of a method for estimating SOC and internal heterogeneity of lithium-ion batteries based on ultrasonic transmission signals.

[0031] Figure 2 GPR-AL method shown

[0032] Figure 3 Shown is the CNN-PELM network architecture

[0033] Figure 4 The following is a comparison of the root mean square error (RMSE) and mean absolute error (MAE) results for different numbers of excellent points z.

[0034] Figure 5a Shown is a comparison of the results of different methods for Cell1

[0035] Figure 5b Shown is a comparison of the results of different methods of Cell2

[0036] Figure 6a Shown are the absolute errors (AE) of different methods for Cell1

[0037] Figure 6b Shown are the absolute errors (AE) of different methods for Cell2

[0038] Figure 7a The internal SOC non-uniformity histogram of Cell1 is shown.

[0039] Figure 7b The figure shows the internal SOC non-uniformity histogram of Cell2.

[0040] Figure 8 The figure shows the comparison between the estimated local SOC value and the tested local SOC value. DETAILED DESCRIPTION

[0041] In order to make the features and advantages of the present invention more clearly understood, the following is a detailed description in conjunction with the accompanying drawings and embodiments:

[0042] like Figure 1 As shown, the present invention provides a method for estimating SOC and internal heterogeneity of a lithium-ion battery based on ultrasonic transmission signals, and the specific process is as follows:

[0043] Step 1: Ultrasonic scanning and transmission signal feature extraction. Specifically: fully charge and discharge the lithium-ion battery, record the voltage and current data, calculate the battery's SOC using the ampere-hour integration method, conduct an ultrasonic scan on the battery surface every 10% SOC interval, extract the amplitude of the ultrasonic transmission signal at all scanning points, and construct the signal amplitude matrix MSA t,p,q , forming a data set corresponding to the features and SOC {X, y} = {MSA t,p,q ,SOC t}, t=1,...,T, where X is the feature, y is the SOC, t represents the time corresponding to the SOC, p represents the number of rows of the matrix, and q represents the number of columns of the matrix.

[0044] Step 2: Selection of excellent feature points based on the signal amplitude matrix. Specifically, for the signal amplitude data set extracted in step 1, the GPR-AL method is used to select points with small uncertainty in each amplitude matrix as excellent feature points. The GPR-AL method is a Gaussian process regression-active learning method. The method flow is as follows:

[0045] (1) Initialize the active learner for the mapping relationship between signal amplitude features and SOC: Use the Gaussian process regression model as the learner. At the tth moment, randomly select n_initial labeled data sets D = {X n_initial ,y n_initial}={(x i ,y i ),i=1,...,n_initial} are sent to the active learner for training, and considering that the actual measured data has a certain degree of noise, that is: y n_initial =f(X n_initial )+ε, where f represents Gaussian process regression; assuming that the noise ε satisfies the standard Gaussian distribution: Since the Gaussian process is described by the mean function and the covariance function, based on Bayesian theory, we know that y n_initial The prior distribution of is: where μ=[m(x1),...,m(x n_initial )] is the mean function, K ij = k(x i ,x j ) is the covariance function, also called the kernel function;

[0046] (2) Update the active learner: Use the initialized active learner to predict the standard deviation of the remaining points, sort and select the points with large standard deviations for annotation, and obtain a data set with large uncertainty {X n_instances ,y n_instances}, merge with the initial data set D and input into the active learner for retraining, and update the learner. Repeat this step until the set number of training times is reached. The specific steps are as follows:

[0047] a.Set D new ={X n_initial ,y n_initial}, remove {X from {X,y} n_initial ,y n_initial}Get the remaining data set {X rest ,y rest}, define the number of tags n_instances required in each iteration;

[0048] b. Replace {X rest ,y rest}The data set is fed into the active learner, so for the new input data X rest , let its prediction be y * , then y n_initial and * The joint distribution of can be expressed as:

[0049]

[0050] Where K = K(X n_initial ,X n_initial ) is the kernel function matrix between the training input data, K * =K(X n_initial ,X rest ) is the kernel function matrix between the training input data and the test input data, K ** =K(X rest ,X rest ) is the kernel function matrix between the test input data, m(X n_initial )=0,m(X rest )=0; Then, the predicted distribution under the new input data can be obtained through conditional distribution: The predicted mean is: The predicted covariance matrix is: Through the above calculation, we can get each X rest The predicted standard deviation is sorted and the first n_instances points with large standard deviations, that is, large uncertainties, are selected for marking, and {X n_instances ,y n_instances};

[0051] c. Next, D new With large uncertainty data {X n_instances ,y n_instances} to merge and get D new ={X new ,y new =D new ∪{X n_instances ,y n_instances}, and then feed it as a new training set to the active learner to retrain and update its performance: Then, the trained data {X n_instances ,y n_instances}From {X rest ,y rest} remove {X rest ,y rest}={X rest \X n_instances ,y rest \y n_instances}, where \ indicates deletion; repeat b to c until the set number of training times is reached;

[0052] (3) Obtaining excellent points: For the test set X 。 , let its prediction be y 。 , X 。 Input into the trained active learner to get the standard deviation corresponding to each point position:

[0053]

[0054] Sort the standard deviations of the prediction results from small to large, and select the first z points with small standard deviations as excellent points. After the above steps, we can get the excellent features in each ultrasound scan. Let t = t + 1, and repeat steps (2) to (3) until t reaches T. Stacking the excellent features at each moment in order can obtain the input-output combination corresponding to the state of charge: {SFC t ,SOC t}.

[0055] Step 3: Construct a lithium-ion battery SOC model based on excellent feature point signal data. Specifically, take the excellent feature combination SFC as input and SOC as output, and use the CNN-PELM neural network for training and testing. The CNN-PELM neural network model is a convolutional neural network-pooling extreme learning machine network model. After the input data passes through the CNN feature extraction layer, the PELM regression fitting layer is used to fit the extracted features. The network model is constructed as follows:

[0056] CNN-PELM is composed of a feature extraction layer composed of a convolutional neural network and a regression fitting layer composed of a pooling extreme learning machine: (1) First, the CNN feature extraction layer: for the input excellent feature combination SFC, it is extracted through a convolutional neural network constructed in the form of "residual connection" to obtain SFC conv : SFC conv=Resnet18(SFC), where the feature extraction layer consists of 5 parts. The first part is a convolutional layer, a normalization layer, and a pooling layer. The convolutional layer has a kernel size of 2, a stride of 2, a padding of 3, and an output channel of 64. The convolutional kernel size of the pooling layer is 3, a stride of 2, and a padding of 1. Then, the second part is two residual blocks, which have the same structure, and each residual block has 2 convolutional layers with the same structure. The convolutional kernel size of each convolutional layer is 3, the stride is 1, the padding is 1, and the output channel is 64. Secondly, the third part is two residual blocks. The input and output dimensions of the first residual block are inconsistent. Therefore, in order to match the dimensions, a convolution layer is added on top of the original two convolution layers. The convolution kernel size of the original two convolution layers is 3, the stride is 2 and 1, and the padding is 1. The convolution kernel size of the newly added third convolution layer is 1, the stride is 2, padding=0, and the output channel is 128; the input and output dimensions of the second residual block are consistent. It has two convolution layers with the same structure, the convolution kernel size is 3, the stride is 1, the padding is 1, and the output channel is 128. Subsequently, the fourth part is similar to the third part in structure. The only difference is that the output channel of the first residual block is 256, and the output channel of the second residual block is 256. Similarly, the fifth part is also similar to the third part in structure. The only difference is that the output channel of the first residual block is 512, and the output channel of the second residual block is 512. Among them, when the input and output dimensions are consistent, the residual block is represented as: y (1) =ReLu(conv(ReLu(conv(x)))+x), when the input and output dimensions are inconsistent, the residual block is represented as: (2) =ReLu(conv(ReLu(conv(x)))+conv(x)); (2) Then comes the PELM regression fitting layer: After the input data passes through the feature extraction layer, the PELM regression fitting layer is used to extract the feature SFC conv Perform fitting. T=SOC t Represents the true value of the regression fit. The calculation process of PELM is as follows: a. Calculate the output matrix H of the hidden layer: H = mean_pool (SFC conv ); b. Next, calculate the output weight matrix β: M = (H′H) -1 , β=MH′T; c. Therefore, the output of PELM is: Among them, ' represents the transpose of the matrix, and the connection weight β between the hidden layer and the output layer is obtained by solving the equation The least squares solution of is obtained.

[0057] Step 4: Estimation of SOC and internal heterogeneity of lithium-ion batteries. Specifically, the corresponding SOC changes are derived by comparing the differences in SA values. Based on step 2, the points with smaller uncertainty in each ultrasonic scan can be obtained. Find the mean of the feature data at these points The mean value and the estimated SOC value obtained in step 3 are As the reference characteristic value and reference SOC value respectively, according to the characteristic value SA at each location point = [SA1, ..., SA T ]=[MSA 1,p,q ,...,MSA T,p,q ] and the formula abs(SOC each_site -SOC est )=abs(SA-SA ave ) Calculate the SOC value SOC at each point each_site , and obtain the estimated SOC and internal heterogeneity of the lithium-ion battery.

[0058] This experiment uses the TOPS-LD50A ultrasonic battery scanning system to conduct research. A 1MHz broadband focused ultrasonic beam is used to test three lithium iron phosphate soft-pack batteries (51×100mm) of the same model and rated capacity of 1Ah at different charge and discharge rates. A total of 120×220 points of data are tested. The following example uses Cell3 as the training set and Cell1 as the test set.

[0059]

[0060]

[0061] The accuracy of the proposed method in realizing the internal heterogeneity of SOC was objectively verified by inductively coupled plasma optical emission spectroscopy (ICP-OES).

[0062] Figure 4 The comparison chart of RMSE and MAE results for different numbers of excellent points z is given. Figure 5a The comparison chart of the results of different methods of Cell1 is given. Figure 5b The results of different methods of Cell2 are compared. Figure 6a The AE diagrams of different methods for Cell1 are given. Figure 6b The AE diagrams of different methods of Cell2 are given. Figure 7a The internal SOC non-uniformity histogram of Cell1 is given. Figure 7b The internal SOC non-uniformity histogram of Cell2 is given. Figure 8 A comparison chart of the estimated local SOC value and the tested local SOC value is given.

[0063] exist Figure 4 In the example, when the number of excellent points z is set to 100, the estimation accuracy of SOC by the method of the present invention is high and the error is small, which meets the basic requirements.

[0064] for Figure 5a and Figure 5b , whether it is Cell1 or Cell2, under the same input and output conditions, the SOC density distribution map obtained by CNN-PELM is closer to the real SOC density distribution map. And for Figure 6a and Figure 6b ,Through objective evaluation, it can be found that CNN-PELM shows smaller errors in the ,SOC estimation process, and AE does not exceed 0.1, which shows that the ,estimation performance of the method of the present ,invention is significant.

[0065] Depend on Figure 7a and Figure 7b It can be seen that the internal SOC non-uniformity histogram is basically similar to a Gaussian distribution. For lithium iron phosphate batteries, due to their unique material system and stable voltage platform, when the battery is in the middle SOC range, the electrochemical reaction inside the battery is relatively mild, resulting in little change in the local SOC distribution. Figure 8 It can be seen that the MAE between the estimated local SOC value and the tested local SOC value is 0.0037, and the determination coefficient (R2) is 0.9696, which not only shows that the error between the estimated value and the tested value is small, but also shows that there is a very strong correlation between the two. Therefore, it is feasible and accurate to use ultrasonic signals to estimate the internal SOC inhomogeneity.

[0066] The above is a preferred embodiment of the present invention. For those skilled in the art, the specific embodiments of the present invention may be modified or replaced by equivalents without departing from the overall concept of the present invention, and these should also be regarded as the protection scope of the present invention.

Claims

1. A method for estimating SOC and internal heterogeneity of lithium-ion batteries based on ultrasonic transmission signals, characterized in that The following steps are involved: Step 1: Ultrasonic scanning and transmission signal feature extraction; The lithium-ion battery is fully charged and discharged, the voltage and current data are recorded, the SOC of the battery is calculated using the ampere-hour integration method, an ultrasonic scan is performed on the battery surface every 10% SOC interval, the amplitude of the ultrasonic transmission signal at all scanning points is extracted, and the signal amplitude matrix MSA is constructed. t,p,q , forming a data set corresponding to the features and SOC {X, y} = {MSA t,p,q ,SOC t }, t = 1, ..., T, where X is the feature, y is the SOC, t is the time corresponding to the SOC, p is the number of rows in the matrix, and q is the number of columns in the matrix; Step 2: Selection of excellent feature points based on signal amplitude matrix; For the signal amplitude data set extracted in step 1, the GPR-AL method is used to screen points with smaller uncertainty in each amplitude matrix as excellent feature points. The GPR-AL method is a Gaussian process regression-active learning method. The method flow is as follows: First, initialize the active learner for the mapping relationship between signal amplitude features and SOC: use the Gaussian process regression model as an active learner, randomly select n_initial labeled data sets D = {X n_initial ,y n_initial }={(x i ,y i ),i=1,...,n_initial} are sent to the learner for training, and y n_initial =f(X n_initial )+ε, where f is Gaussian process regression and ε is noise; Secondly, update the learner: use the initialized active learner to predict the standard deviation of the remaining points, sort and select the points with large standard deviation for annotation, and obtain a data set with large uncertainty {X n_instances ,y n_instances }, merge with the initial data set D and input into the active learner for retraining, and update the learner, repeating this step until the set number of training times is reached; Then, get the excellent points: for the test set X 。 , let its prediction be y 。 , X 。 Input it into the trained active learner to get the standard deviation corresponding to each point position, and select the point with small standard deviation as the excellent feature point SFC; Step 3: Construction of lithium-ion battery SOC model based on excellent feature point signal data; The excellent feature point SFC is used as input and SOC is used as output. The CNN-PELM neural network model is used for training and testing. The CNN-PELM neural network model is a convolutional neural network-pooling extreme learning machine network model. After the input data passes through the CNN feature extraction layer, the extracted feature SFC is extracted using the PELM regression fitting layer. conv Fitting; T = SOC t Represents the true value of the regression fit. The calculation process of PELM is as follows: First, calculate the output matrix H of the hidden layer: H = mean_pool (SFC conv ), and then calculate the output weight matrix β: M = (H′H) -1 , β=MH′T, and the final output is: Step 4: Li-ion battery SOC and internal heterogeneity estimation; The corresponding SOC changes are deduced by comparing the differences in amplitude values; based on step 2, the SFC of the points with smaller uncertainty in each ultrasonic scan image can be obtained, and the mean SA of the characteristic data at these point positions can be calculated. ave , and the mean value and the estimated SOC value SOC obtained in step 3 est As reference characteristic value and reference SOC value respectively, according to the amplitude value at each position point and the formula abs(SOC each_site -SOC est )=abs(SA-SA ave ) Calculate the SOC value SOC at each point each_site , get the estimated value of SOC and internal heterogeneity of lithium-ion battery; Through the above steps, a method for estimating SOC and internal heterogeneity of lithium-ion batteries based on ultrasonic transmission signals is given.

Images