Lithium ion battery SOC (State of Charge) estimation method, system, medium and equipment

Abstract

The invention discloses a lithium ion battery SOC estimation method, system, medium and equipment based on electrochemical impedance spectroscopy and parameter fine tuning migration, and the method comprises the steps: collecting electrochemical impedance spectroscopy data of a lithium ion battery in different states of charge; screening out input features through correlation analysis; constructing a neural network model comprising an encoder for extracting an abstract representation of the input feature, a decoder for reconstructing the input feature, and a predictor for outputting a state of charge estimate; normalizing the input features of the source domain data, inputting the normalized input features into a neural network model, and carrying out unsupervised or self-supervised pre-training to adjust model parameters of an encoder and a decoder; and after the pre-training is completed, freezing at least a part of residual block layers in the encoder, setting the hierarchical learning rate of the encoder and the predictor, inputting the input features of the target domain data into the frozen model, and carrying out supervised fine tuning training to obtain a trained neural network model for SOC estimation.

Description

Technical Field

[0001] This invention relates to the field of lithium-ion battery state of charge estimation technology, and in particular to a lithium-ion battery SOC estimation method, system, medium, and device based on electrochemical impedance spectroscopy and parameter fine-tuning migration. Background Technology

[0002] Lithium-ion batteries possess advantages such as high energy density, wide operating temperature range, and long cycle life, making them a widely used next-generation energy storage medium in various fields. Battery management systems (BMS) ensure the safe and efficient operation of lithium-ion batteries, and accurate estimation of the state of charge (SOC) is a crucial component of BMS. Traditional methods based on monitoring data such as battery voltage and current are susceptible to environmental interference, making it difficult to meet the requirements of high-precision BMS. Electrochemical impedance spectroscopy (EIS) analysis of lithium-ion batteries helps overcome the environmental sensitivity of traditional methods, achieving high-precision and robust SOC estimation under complex environments. However, obtaining EIS data for lithium-ion batteries is difficult and costly, resulting in data scarcity and limiting the development of high-performance SOC estimation models.

[0003] The information disclosed in the background section is only for enhancing the understanding of the background of this invention, and therefore may contain information that does not constitute prior art known to those skilled in the art. Summary of the Invention

[0004] To address the problems of existing technologies, this invention provides a method, system, medium, and device for estimating the state of charge (SOC) of lithium-ion batteries based on electrochemical impedance spectroscopy and parameter fine-tuning migration. It deeply extracts features related to the state of charge of lithium-ion batteries from electrochemical impedance spectroscopy and combines them with parameter fine-tuning migration methods, which significantly reduces the demand for lithium-ion battery electrochemical impedance spectroscopy data and the training cost of the model, and effectively improves the accuracy and robustness of the state of charge estimation.

[0005] A method for estimating the state of charge (SOC) of lithium-ion batteries based on electrochemical impedance spectroscopy and parameter fine-tuning migration includes:

[0006] S100: Collect electrochemical impedance spectroscopy data of lithium-ion batteries under different states of charge. The electrochemical impedance spectroscopy data includes sampling frequency, real part of impedance and imaginary part of impedance. The electrochemical impedance spectroscopy data of one type of lithium-ion battery is used as the target domain data, and its corresponding SOC is known. The electrochemical impedance spectroscopy data of other types of lithium-ion batteries are used as the target domain data, and their corresponding SOC is unknown.

[0007] S200: Based on the electrochemical impedance spectroscopy data, extract the equivalent circuit model parameters and relaxation time distribution characteristics, and use correlation analysis to select features that are more correlated with the state of charge than a preset threshold as input features.

[0008] S300: Construct a neural network model that includes an encoder for extracting abstract representations of input features, a decoder for reconstructing input features, and a predictor for outputting estimates of the state of charge.

[0009] S400: Normalize the input features of the source domain data and input them into the neural network model for unsupervised or self-supervised pre-training to adjust the model parameters of the encoder and decoder;

[0010] S500: After pre-training, freeze at least some residual block layers in the encoder, set the hierarchical learning rates of the encoder and predictor, input the input features of the target domain data into the frozen model, perform supervised fine-tuning training, and obtain the trained neural network model for SOC estimation.

[0011] In the aforementioned method for estimating the state of charge (SOC) of a lithium-ion battery based on electrochemical impedance spectroscopy and parameter fine-tuning migration, the electrochemical impedance spectroscopy data collected from the lithium-ion battery at a certain state of charge is used. , For sampling frequency, For the real part of the impedance, Let N be the imaginary part of the impedance and N be the sampling frequency. The source domain data can be represented as:

[0012]

[0013] in, It is the source domain feature vector. It is the number of samples in the source domain.

[0014] The target domain data can be represented as:

[0015]

[0016] in, It is the feature vector of the target domain. It is the target domain tag. It is the number of samples in the target domain.

[0017] In the lithium-ion battery SOC estimation method based on electrochemical impedance spectroscopy and parameter fine-tuning migration, an equivalent circuit model is fitted using a nonlinear least squares optimization algorithm. The objective function to achieve simultaneous fitting of the real and imaginary parts of the impedance with uniform weights during fitting is as follows:

[0018]

[0019] in, Indicates frequency, Represents frequency The real part of the corresponding original impedance spectrum, Represents frequency The real part of the impedance spectrum of the corresponding equivalent circuit model. Represents frequency The imaginary part of the corresponding original impedance spectrum, Represents frequency The imaginary part of the impedance spectrum of the corresponding equivalent circuit model, where N is the frequency number.

[0020] If the optimal equivalent circuit model is known, it is directly selected; if the optimal equivalent circuit model is unknown, it is evaluated and selected from the Randles model, the Randles CPE model, the second-order integer circuit, the second-order fractional circuit model, and the custom equivalent circuit model. The selection is based on the regression evaluation index of the coefficient of determination, root mean square error, and mean absolute percentage error. After selecting the optimal model based on the fitting effect, the parameters of each circuit element of the equivalent circuit model are obtained by reverse engineering.

[0021] In the aforementioned method for estimating the state of charge (SOC) of lithium-ion batteries based on electrochemical impedance spectroscopy and parameter fine-tuning migration, the electrochemical impedance spectroscopy data is represented based on the relaxation time distribution as follows:

[0022]

[0023] in, This is a high-frequency limiting resistor; The relaxation time distribution function represents the contribution weight of different relaxation processes to the total impedance and is an unknown DRT function that needs to be solved.

[0024] The relaxation time distribution analysis is known. Reverse solving The process involves DRT analysis, which yields a continuous distribution function. Further analysis reveals two characteristics: the high-frequency limiting resistance corresponding to the time constant approaching zero in the relaxation time distribution spectrum. Another characteristic is the total polarization resistance. Its value is a function The integral over the entire time constant characterizes the total energy consumption of all polarization processes within a lithium-ion battery, and the calculation formula is as follows:

[0025] .

[0026] In the lithium-ion battery SOC estimation method based on electrochemical impedance spectroscopy and parameter fine-tuning migration, the correlation analysis uses Spearman's rank correlation coefficient to calculate the correlation between each extracted feature and the true state of charge. The preset threshold is 0.8, and features with an absolute correlation coefficient greater than 0.8 are selected as input features.

[0027] In the lithium-ion battery SOC estimation method based on electrochemical impedance spectroscopy and parameter fine-tuning migration, the encoder consists of an initial convolutional layer and multiple one-dimensional residual blocks connected sequentially. The initial convolutional layer includes a one-dimensional convolutional layer, a batch normalization layer, a ReLU activation layer, and a max-pooling layer. Each one-dimensional residual block includes a main path and a shortcut path. The main path contains two one-dimensional convolutional layers and an intermediate batch normalization and activation layer. The shortcut path performs dimension matching through 1×1 convolution when the input and output dimensions are inconsistent. The decoder has a symmetrical structure with the encoder and consists of multiple one-dimensional transposed convolutional layers. The predictor includes a Flatten layer, a fully connected layer, and a ReLU activation layer, and finally outputs a single-value SOC estimate.

[0028] In the lithium-ion battery SOC estimation method based on electrochemical impedance spectroscopy and parameter fine-tuning migration, when freezing some neural network layers, the 2nd and 4th residual block layers in the hierarchical residual learning module are selected to be frozen. The Huber loss function is selected as the loss function for model training, and its calculation formula is as follows:

[0029]

[0030] in, Represents the actual value. Indicates the predicted value. It is a hyperparameter greater than 0.

[0031] An evaluation system for implementing a lithium-ion battery SOC estimation method based on electrochemical impedance spectroscopy and parameter fine-tuning migration includes:

[0032] The acquisition module acquires electrochemical impedance spectroscopy data of lithium-ion batteries under different states of charge. The electrochemical impedance spectroscopy data includes the sampling frequency, the real part of the impedance, and the imaginary part of the impedance.

[0033] The extraction module extracts the equivalent circuit model parameters and relaxation time distribution characteristics based on the electrochemical impedance spectroscopy data, and selects features with a correlation higher than a preset threshold as input features through correlation analysis.

[0034] The building module constructs a neural network model that includes an encoder for extracting abstract representations of input features, a decoder for reconstructing input features, and a predictor for outputting estimates of the state of charge.

[0035] The training module normalizes the input features of the source domain data and inputs them into the neural network model for unsupervised or self-supervised pre-training to adjust the model parameters of the encoder and decoder.

[0036] In the estimation module, after pre-training, at least some residual block layers in the encoder are frozen, the hierarchical learning rates of the encoder and predictor are set, the input features of the target domain data are input into the frozen model, and supervised fine-tuning training is performed to obtain the trained neural network model for SOC estimation.

[0037] A computer storage medium including computer instructions that, when run on a computer, cause the computer to perform the method.

[0038] An electronic device, the electronic device comprising:

[0039] Memory, processor, and computer programs stored in memory and executable on the processor, wherein,

[0040] The processor implements the method when executing the program.

[0041] Compared with the prior art, the present invention has the following advantages: the present invention significantly reduces the demand for lithium-ion battery electrochemical impedance spectroscopy data and the training cost of the model, and effectively improves the generalization ability and estimation accuracy of the lithium-ion battery state of charge estimation model. Attached Figure Description

[0042] Various other advantages and benefits of the present invention will become apparent to those skilled in the art upon reading the detailed description of the preferred embodiments below. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. It is obvious that the drawings described below are merely some embodiments of the invention, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort. Furthermore, the same reference numerals denote the same parts throughout the drawings.

[0043] In the attached diagram:

[0044] Figure 1 This is a flowchart illustrating the steps of a lithium-ion battery SOC estimation method based on electrochemical impedance spectroscopy and parameter fine-tuning migration, according to an embodiment of this disclosure.

[0045] Figure 2 This is a flowchart illustrating the process of solving the state-of-charge characteristics of a lithium-ion battery based on the equivalent circuit model theory and relaxation time distribution analysis method according to an embodiment of this disclosure.

[0046] Figure 3This is a parameter fine-tuning migration model framework diagram of a lithium-ion battery SOC estimation method based on electrochemical impedance spectroscopy and parameter fine-tuning migration, provided according to an embodiment of this disclosure.

[0047] The present invention will be further explained below with reference to the accompanying drawings and embodiments. Detailed Implementation

[0048] The following will refer to the appendix. Figures 1 to 3 Specific embodiments of the invention will be described in more detail below. While specific embodiments of the invention are shown in the accompanying drawings, it should be understood that the invention may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the invention and to fully convey the scope of the invention to those skilled in the art.

[0049] It should be noted that certain terms are used in the specification and claims to refer to specific components. Those skilled in the art will understand that different terms may be used to refer to the same component. This specification and claims do not distinguish components based on differences in terminology, but rather on differences in function. The terms "comprising" or "including" used throughout the specification and claims are open-ended and should be interpreted as "comprising but not limited to." The following descriptions are preferred embodiments for carrying out the invention; however, these descriptions are for the purpose of understanding the general principles of the specification and are not intended to limit the scope of the invention. The scope of protection of this invention is determined by the appended claims.

[0050] To facilitate understanding of the embodiments of the present invention, further explanations and descriptions will be provided below with reference to the accompanying drawings and specific embodiments. The accompanying drawings do not constitute a limitation on the embodiments of the present invention.

[0051] In one embodiment, such as Figure 1 As shown, this disclosure provides a method for estimating the state of charge (SOC) of a lithium-ion battery based on electrochemical impedance spectroscopy and parameter fine-tuning migration, including the following steps:

[0052] S100: Collect electrochemical impedance spectroscopy data of lithium-ion batteries under different states of charge. The electrochemical impedance spectroscopy data includes sampling frequency, real part of impedance and imaginary part of impedance. The electrochemical impedance spectroscopy data of one type of lithium-ion battery is used as the target domain data, and its corresponding SOC is known. The electrochemical impedance spectroscopy data of other types of lithium-ion batteries are used as the target domain data, and their corresponding SOC is unknown.

[0053] In this step, in a preferred embodiment of the lithium-ion battery SOC estimation method based on electrochemical impedance spectroscopy and parameter fine-tuning migration, the electrochemical impedance spectroscopy data collected by the lithium-ion battery at a certain state of charge... , For sampling frequency, For the real part of the impedance, Let N be the imaginary part of the impedance and N be the sampling frequency. The source domain data can be represented as:

[0054]

[0055] in, It is the source domain feature vector. It is the number of samples in the source domain.

[0056] The target domain data can be represented as:

[0057]

[0058] in, It is the feature vector of the target domain. It is the target domain tag. It is the number of samples in the target domain.

[0059] Furthermore, for those skilled in the art, the source domain and target domain are two fundamental concepts in transfer learning. Existing knowledge is typically referred to as the source domain, and the new knowledge to be learned is called the target domain. Transfer learning allows knowledge from the source domain to be transferred to the target domain. Specifically, in this embodiment, the source domain data can be electrochemical impedance spectroscopy (EIS) data of several types of lithium-ion batteries without SOC tags, while the target domain can be EIS data of a specific type of lithium-ion battery with a true SOC tag. Using the method provided in this embodiment, battery degradation trend information contained in the source domain data can be transferred to the target domain data, thereby completing the SOC estimation for different types of lithium-ion batteries.

[0060] S200: Based on the electrochemical impedance spectroscopy data, extract the equivalent circuit model parameters and relaxation time distribution characteristics, and use correlation analysis to select features that are more correlated with the state of charge than a preset threshold as input features.

[0061] In this step, in a preferred embodiment of the lithium-ion battery SOC estimation method based on electrochemical impedance spectroscopy and parameter fine-tuning migration, an equivalent circuit model is fitted using a nonlinear least squares optimization algorithm. The objective function that achieves simultaneous fitting of the real and imaginary parts of the impedance with uniform weights during fitting is:

[0062]

[0063] in, Indicates frequency, Represents frequency The real part of the corresponding original impedance spectrum, Represents frequency The real part of the impedance spectrum of the corresponding equivalent circuit model. Represents frequency The imaginary part of the corresponding original impedance spectrum, Represents frequency The imaginary part of the impedance spectrum of the corresponding equivalent circuit model, where N is the frequency number.

[0064] If the optimal equivalent circuit model is known, it is directly selected; if the optimal equivalent circuit model is unknown, it is evaluated and selected from the Randles model, the Randles CPE model, the second-order integer circuit, the second-order fractional circuit model, and the custom equivalent circuit model. The selection is based on the regression evaluation index of the coefficient of determination, root mean square error, and mean absolute percentage error. After selecting the optimal model based on the fitting effect, the parameters of each circuit element of the equivalent circuit model are obtained by reverse engineering.

[0065] In a preferred embodiment of the lithium-ion battery SOC estimation method based on electrochemical impedance spectroscopy and parameter fine-tuning migration, the electrochemical impedance spectroscopy data is represented based on the relaxation time distribution as follows:

[0066]

[0067] in, This is a high-frequency limiting resistor; Let be the relaxation time distribution function, representing the contribution weight of different relaxation processes to the total impedance. It is the unknown DRT function that needs to be solved.

[0068] The relaxation time distribution analysis is known. Reverse solving The process involves DRT analysis, which yields a continuous distribution function. Further analysis reveals two characteristics: the high-frequency limiting resistance corresponding to the time constant approaching zero in the relaxation time distribution spectrum. Another characteristic is the total polarization resistance. Its value is a function The integral over the entire time constant characterizes the total energy consumption of all polarization processes within a lithium-ion battery, and the calculation formula is as follows:

[0069] .

[0070] In a preferred embodiment of the lithium-ion battery SOC estimation method based on electrochemical impedance spectroscopy and parameter fine-tuning migration, the correlation analysis uses Spearman's rank correlation coefficient to calculate the correlation between each extracted feature and the true state of charge. The preset threshold is 0.8, and features with an absolute value of Spearman's rank correlation coefficient greater than 0.8 are selected as input features.

[0071] S300: Construct a neural network model that includes an encoder for extracting abstract representations of input features, a decoder for reconstructing input features, and a predictor for outputting estimates of the state of charge.

[0072] In a preferred embodiment of the lithium-ion battery SOC estimation method based on electrochemical impedance spectroscopy and parameter fine-tuning migration, the encoder consists of an initial convolutional layer and multiple one-dimensional residual blocks connected sequentially. The initial convolutional layer includes a one-dimensional convolutional layer, a batch normalization layer, a ReLU activation layer, and a max-pooling layer. Each one-dimensional residual block includes a main path and a shortcut path. The main path contains two one-dimensional convolutional layers and an intermediate batch normalization and activation layer. The shortcut path performs dimension matching through 1×1 convolution when the input and output dimensions are inconsistent. The decoder has a symmetrical structure with the encoder and consists of multiple one-dimensional transposed convolutional layers. The predictor includes a Flatten layer, a fully connected layer, and a ReLU activation layer, and finally outputs a single-value SOC estimate.

[0073] The encoder consists of an initial convolutional layer and four one-dimensional residual blocks linked sequentially; the decoder has a symmetrical structure with the encoder; the predictor consists of a Flatten layer, a fully connected layer, and a ReLU layer.

[0074] The specific composition and connection order of the initial convolutional layer in the encoder are as follows: one-dimensional convolutional layer, batch normalization layer, ReLU layer and max pooling layer. The one-dimensional convolutional layer has 1 input channel, 32 output channels, a kernel size of 3, a stride of 1, and 1 padding. The max pooling layer has a kernel size of 2.

[0075] In the four-layer residual block link of the encoder, the first layer residual block has 32 input channels and 32 output channels; the second layer residual block has 32 input channels and 64 output channels, with a step size of 2; the first layer residual block has 64 input channels and 64 output channels; the second layer residual block has 64 input channels and 32 output channels, with a step size of 2.

[0076] The one-dimensional residual block has two pathways: a main pathway and a shortcut pathway. The outputs of the main pathway and the shortcut pathway are added element-wise in an adder to achieve residual learning. This operation combines the residual features learned by the main pathway with the original features or their projections from the shortcut pathway. The sum is then passed through a ReLU layer to generate the final output of the residual block. This structure effectively alleviates the vanishing gradient problem prevalent in deep neural networks, allowing signals and gradients to propagate directly in extremely deep networks.

[0077] The main pathway primarily functions to perform two convolutional transformations and nonlinear activations on the input data. It consists of a first convolutional layer, a batch normalization layer, a ReLU layer, a second convolutional layer, and a batch normalization layer linked sequentially. The first convolutional layer has the same number of input channels as the residual block layer, the same number of output channels as the residual block layer, a kernel size of 3, a stride of 1, and a padding of 1. The first convolutional layer also has the same number of input channels as the residual block layer, the same number of output channels as the residual block layer, a kernel size of 3, a stride of 1, and a padding of 1.

[0078] The shortcut path is used to linearly transform the input data to match the output of the main path in terms of spatiotemporal dimension and number of channels, enabling element-wise addition. This path is active when the stride of the main path is not equal to 1 or the number of input channels is not equal to the number of output channels; otherwise, the path is an empty sequence composed of identity mappings, directly jumping the input data to the adder. The shortcut path consists of a one-dimensional convolutional layer and a batch normalization layer linked sequentially. The kernel size of the one-dimensional convolutional layer is 1, and its stride is consistent with that of the first convolutional layer of the main path, responsible for adjusting the dimension and number of channels of the input data.

[0079] The decoder, used to convert the abstract data extracted by the encoder into raw data, is composed of a first transposed convolutional layer, a batch normalization layer, a ReLU layer, a second transposed convolutional layer, a batch normalization layer, a ReLU layer, a third transposed convolutional layer, a batch normalization layer, a ReLU layer, a fourth transposed convolutional layer, and an adaptive one-dimensional average pooling layer, sequentially linked together. The first transposed convolutional layer has 32 input channels, 64 output channels, a kernel size of 3, a stride of 2, and padding of [missing information]. The first transposed convolutional layer has 64 input channels, 64 output channels, a kernel size of 3, a stride of 1, and 1 padding. The second transposed convolutional layer has 64 input channels, 32 output channels, a kernel size of 3, a stride of 2, 1 padding, and 1 output padding. The fourth transposed convolutional layer has 32 input channels, 1 output channel, a kernel size of 3, a stride of 2, 1 padding, and 1 output padding.

[0080] The predictor, used to further extract the abstract data to obtain the predicted value SOH of the lithium-ion battery, is composed of a Flatten layer, a first fully connected layer, a ReLU layer, and a second fully connected layer linked sequentially. The first fully connected layer has 32 input channels and 16 output channels, and the second fully connected layer has 16 input channels and 1 output channel.

[0081] S400: Normalize the input features of the source domain data and input them into the neural network model for unsupervised or self-supervised pre-training to adjust the model parameters of the encoder and decoder;

[0082] During training, the model normalizes the maximum absolute value of the source domain data, scaling the numerical range to [-1, 1]. The scaled data is then input into the encoder to generate high-dimensional data features. These high-dimensional features are then input into the decoder for training, reconstructing the original electrochemical impedance spectroscopy (EIS) features. This process is then reverse-propagated, and the parameters of the encoder and decoder are optimized based on the loss between the original EIS features and the features generated by the decoder. During model training, the Huber loss function is used. A simple early stopping mechanism is also implemented: the model is considered to have completed training and converged after 100 consecutive rounds of no decrease in loss.

[0083] S500: After pre-training, freeze at least some residual block layers in the encoder, set the hierarchical learning rates of the encoder and predictor, input the input features of the target domain data into the frozen model, perform supervised fine-tuning training, and obtain the trained neural network model for SOC estimation.

[0084] In a preferred embodiment of the lithium-ion battery SOC estimation method based on electrochemical impedance spectroscopy and parameter fine-tuning migration, when freezing some neural network layers, the 2nd and 4th residual block layers in the layered residual learning module are selected to be frozen. The Huber loss function is selected as the loss function for model training, and its calculation formula is as follows:

[0085]

[0086] in, Represents the actual value. Indicates the predicted value. It is a hyperparameter greater than 0.

[0087] In the supervised fine-tuning training of the model, the encoder and predictor of the frozen partial layers are spliced ​​together, and the target domain data is input into the encoder and predictor for training. Backpropagation is used to adjust the parameters of the encoder and predictor. During the fine-tuning training process, the present invention sets a simple early stopping mechanism. The model is considered to have completed training and converged after the loss has not decreased for 50 consecutive rounds.

[0088] When the model is used for SOC estimation, the features of the lithium-ion battery electrochemical impedance spectroscopy data are input into the trained model to obtain the corresponding SOC estimate.

[0089] An evaluation system for implementing a lithium-ion battery SOC estimation method based on electrochemical impedance spectroscopy and parameter fine-tuning migration includes:

[0090] The acquisition module acquires electrochemical impedance spectroscopy data of lithium-ion batteries under different states of charge. The electrochemical impedance spectroscopy data includes the sampling frequency, the real part of the impedance, and the imaginary part of the impedance.

[0091] The extraction module extracts the equivalent circuit model parameters and relaxation time distribution characteristics based on the electrochemical impedance spectroscopy data, and selects features with a correlation higher than a preset threshold as input features through correlation analysis.

[0092] The building module constructs a neural network model that includes an encoder for extracting abstract representations of input features, a decoder for reconstructing input features, and a predictor for outputting estimates of the state of charge.

[0093] The training module normalizes the input features of the source domain data and inputs them into the neural network model for unsupervised or self-supervised pre-training to adjust the model parameters of the encoder and decoder.

[0094] In the estimation module, after pre-training, at least some residual block layers in the encoder are frozen, the hierarchical learning rates of the encoder and predictor are set, the input features of the target domain data are input into the frozen model, and supervised fine-tuning training is performed to obtain the trained neural network model for SOC estimation.

[0095] A computer storage medium including computer instructions that, when run on a computer, cause the computer to perform the method.

[0096] An electronic device, the electronic device comprising:

[0097] Memory, processor, and computer programs stored in memory and executable on the processor, wherein,

[0098] The processor implements the method when executing the program.

[0099] Figure 2 This is a flowchart for solving the state-of-charge characteristics of lithium-ion batteries based on equivalent circuit model theory and relaxation time distribution analysis. In this process, we initially possess the real and imaginary parts of the lithium-ion battery electrochemical impedance spectroscopy data.

[0100] Table 1

[0101]

[0102] In the process of extracting the parameters of the equivalent circuit model of a lithium-ion battery, Python is used to model common equivalent circuit model components, including but not limited to resistors R, capacitors C, inductors L, constant phase angle components CPE, and Warburg resistors. Initially, we do not know the optimal equivalent circuit model of a lithium-ion battery, so we list some common equivalent circuit models as shown in Table 1.

[0103] After establishing the equivalent circuit model, given initial equivalent circuit model parameters, the optimal equivalent circuit model parameters are obtained by continuously minimizing the objective function using the Levenberg-Marquardt algorithm. The objective function formula is as follows:

[0104]

[0105] in, Indicates frequency, Represents frequency The real part of the corresponding original impedance spectrum, Represents frequency The real part of the impedance spectrum of the corresponding equivalent circuit model. Represents frequency The imaginary part of the corresponding original impedance spectrum, Represents frequency The imaginary part of the impedance spectrum of the corresponding equivalent circuit model, where N is the frequency number.

[0106] After determining the optimal model parameters for each equivalent circuit, the optimal equivalent circuit model is selected by evaluating regression indicators such as the coefficient of determination, root mean square error, and mean absolute percentage error between the collected raw data points and the electrochemical impedance spectroscopy data points of the fitted equivalent circuit model. In this example, the second-order integer circuit, the second-order fractional circuit, and the custom circuit all perform well, with coefficients of determination all above 0.96 and root mean square errors all below 3. 10 -6 In the following section, considering computational complexity, the equivalent circuit model parameters of the solved second-order integer-order circuit are finally selected.

[0107] Relaxation time distribution analysis is a time-domain analytical tool for electrochemical impedance spectroscopy (EIS). It does not rely on prior knowledge of the object of study and requires no equivalent circuit model, thus allowing the separation and analysis of highly overlapping physicochemical processes in EIS data. The core idea of ​​relaxation time distribution is to decompose a complex multi-timescale relaxation process into a series of processes with different relaxation times. The impedance of a lithium-ion battery is a superposition of simple processes. Based on the relaxation time distribution analysis theory, the internal impedance of a lithium-ion battery can be described as:

[0108]

[0109] in, This is a high-frequency limiting resistor; Let be the relaxation time distribution function, representing the contribution weight of different relaxation processes to the total impedance. It is the unknown DRT function that needs to be solved.

[0110] In the process of performing relaxation time distribution analysis, our goal is to solve for... However, our data consists of discrete data points ranging from 0.1Hz to 10kHz, making the solution process inherently ill-conditioned and unable to be directly reversed. Therefore, we employ an extended basis, discretization, deconvolution, and regularization, followed by a grid search algorithm for solution. function

[0111] Solving for the results Subsequently, based on this, we derived two characteristics: one is the high-frequency limiting resistance. This corresponds to the resistance when the time constant in the relaxation time distribution spectrum approaches zero; another characteristic is the total polarization resistance. Its value is a function The integral over the entire time constant can characterize the total energy consumption of all polarization processes inside a lithium-ion battery, and its calculation formula is as follows:

[0112]

[0113] After obtaining the equivalent circuit model parameters and relaxation time distribution characteristics in the electrochemical impedance spectroscopy, the features with strong correlations are selected as the inputs to the model based on the correlation.

[0114] Figure 3 This is a parameter fine-tuning migration model framework diagram for a lithium-ion battery SOC estimation method based on electrochemical impedance spectroscopy and parameter fine-tuning migration. (This is understandable.) Figure 3 The encoder, decoder, and predictor in this document are merely descriptions of the functionality of network modules, which may consist of convolutional layers, pooling layers, batch normalization layers, and fully connected layers, etc., rather than representing a specific structure. Those skilled in the art may use different terms to refer to the same component. This specification and claims do not distinguish components based on differences in terminology, but rather on differences in their functionality.

[0115] By training the encoder and decoder with a large amount of unlabeled source domain data, the encoder can learn the basic common features of these electrochemical impedance spectroscopy data. Then, some layer parameters of the encoder are frozen, and the target domain data is input into the decoder and predictor with the frozen parameters for fine-tuning. This achieves precise adjustments for the task, reducing training costs and improving model robustness. Further details on the training process are provided below.

[0116] The technical solution of this disclosure will be further illustrated below by comparing the prediction results before and after using this method.

[0117] Specifically, the main evaluation indicators used in this comparative experiment are root mean square error, mean absolute error, and coefficient of determination.

[0118] In this experiment, the data used were electrochemical impedance spectroscopy (EIS) data from 20 lithium-ion batteries of four different types collected in the laboratory. The frequency measurement range for the lithium-ion battery EIS data acquisition was 0.1 Hz to 10 kHz. The SOC range was 0-100%, with each of the first 25 discharge cycles releasing 2% SOC and each of the subsequent 50 discharge cycles releasing 1% SOC. The models and parameters of the four types of lithium-ion batteries used for data acquisition are shown in Table 2. All of them had a lithium-ion battery health of 100%, and the ambient temperature was controlled at room temperature (25℃).

[0119] Table 2

[0120]

[0121] After feature extraction, to verify the effectiveness of the invention, four tasks (Task-1 to Task-4) were set, each with different source and target domain data. For Task-1, NCA battery data was used as the target domain, and data from other lithium-ion batteries (after removing labels) was used as the source domain. For Task-2, NCM battery data was used as the target domain, and data from other lithium-ion batteries (after removing labels) was used as the source domain. For Task-3, LFP battery data was used as the target domain, and data from other lithium-ion batteries (after removing labels) was used as the source domain. For Task-4, LCO battery data was used as the target domain, and data from other lithium-ion batteries (after removing labels) was used as the source domain. To ensure fairness in the experiment, all methods used the same hyperparameter settings, namely: 300 pre-training epochs and 50 fine-tuning epochs; batch size of 128 (the number of samples input to the model each time); AdamW as the optimizer; a uniform learning rate of 0.001 during pre-training, and a hierarchical learning rate strategy during fine-tuning, with a learning rate of 0.001 for the decoder and 0.0001 for the predictor; a regularization coefficient of 0.01; and cosine annealing as the learning rate scheduling strategy. The final comparison results between experiments using and not using this method are shown in Table 3.

[0122] Table 3

[0123]

[0124] ① This method is not used: This means that only the encoder and predictor are used for training, and no parameter fine-tuning or transfer is performed. Therefore, the source domain data is not used, and only the target domain data is used.

[0125] The experimental results show that the method of the present invention is significantly better than that of the unused method, and can effectively improve the accuracy of lithium-ion battery charge estimation.

[0126] Furthermore, this invention collects electrochemical impedance spectroscopy (EIS) data of various lithium-ion batteries under different states of charge (SOC) as source and target domain data. Traditional SOC estimation methods rely on a large amount of labeled data, while EIS data acquisition is costly and samples are scarce, making it difficult to support deep model training. This invention provides diverse EIS datasets across models and batches, constructing the "source-target domain" data foundation required for transfer learning, providing data support for subsequent feature extraction and model generalization, and alleviating the problem of insufficient labeled data in the target domain. EIS data is transformed into component parameters with clear electrochemical physical meaning, reflecting key processes such as ohmic impedance, charge transfer impedance, and double-layer capacitance within the battery, improving the physical interpretability of features and their sensitivity to SOC changes. An ECM model is fitted based on the nonlinear least squares method, and the optimal model is selected through indicators such as the coefficient of determination and RMSE. The optimal ECM structure differs for different battery types or aging states, and a fixed model may lead to large fitting biases. Adaptive model selection is implemented to ensure maximum fitting accuracy, avoid parameter distortion caused by model mismatch, and improve the robustness and accuracy of feature extraction. Extract relaxation time distribution (DRT) features, including high-frequency limiting resistance R0 and total polarization resistance R. p Multiple overlapping electrochemical processes in EIS are difficult to separate, and traditional methods cannot resolve multi-timescale kinetic behavior. The DRT method, without requiring prior circuit assumptions, decomposes the complex impedance response into relaxation processes on continuous timescales, extracting R0 and R... pThis approach can characterize battery ohmic loss and overall polarization, enhancing the ability to depict dynamic processes related to SOC. Spearman's rank correlation coefficient is used to screen features strongly correlated with SOC. The extracted ECM and DRT features may contain redundant or weakly correlated variables, affecting model convergence speed and generalization performance. A subset of features with strong monotonic correlation to SOC is automatically selected, reducing input dimensionality, model complexity, and improving training efficiency and prediction stability. A joint encoder-decoder-predictor model structure is constructed. The encoder uses one-dimensional residual blocks. Traditional fully connected networks struggle to effectively capture local dependencies between EIS features, and deep networks suffer from gradient vanishing. One-dimensional convolution and residual structures effectively extract local patterns between features, and residual connections alleviate gradient vanishing in deep networks, allowing the model to delve deeper into abstract feature representations. The decoder is used to reconstruct the input, assisting the encoder in learning general feature expressions. Unsupervised pre-training (encoder + decoder) is performed on source domain data. Due to limited labeled data in the target domain, direct training of deep models is prone to overfitting and poor generalization ability. By leveraging a large amount of unlabeled source domain data to drive the encoder to learn general low-level feature representations of EIS data (such as impedance trends and frequency response patterns), a common knowledge base across battery types is established, significantly reducing the dependence on labeled data in the target domain. During the fine-tuning stage, some residual block layers (such as the 2nd and 4th) are frozen. If all parameters are fine-tuned, small sample target domains may easily destroy the general features obtained through pre-training; if completely frozen, it cannot adapt to new tasks. Deep feature extraction modules (mid-to-low-level residual blocks) are frozen to retain the learned general electrochemical features; only high-level networks are fine-tuned to achieve a balance between "knowledge transfer + task adaptation," preventing catastrophic forgetting and improving model stability and convergence speed. Layered learning rates are set: a large learning rate for the encoder and a small learning rate for the predictor. A uniform learning rate is difficult to balance the optimization needs of different modules, and the encoder may be over-updated or under-updated. High-level predictors require fine-tuning to match SOC output, so a small learning rate is used; the encoder already has good feature extraction capabilities and only needs minor adjustments, so a larger learning rate is used to accelerate convergence, achieving differentiated optimization and improving training efficiency and final accuracy. During the fine-tuning phase, the Huber loss function is used. MSE loss is sensitive to outliers, while MAE loss is not differentiable at zero, affecting optimization stability. Huber loss uses a squared term to ensure convergence when the error is small, and a linear term to suppress the influence of outliers when the error is large, improving the model's robustness to measurement noise and outliers, and ensuring the stability of SOC estimation.

[0127] Although embodiments of the present invention have been described above in conjunction with the accompanying drawings, the present invention is not limited to the specific embodiments and application fields described above. The specific embodiments described above are merely illustrative and instructive, and not restrictive. Those skilled in the art can make many other forms based on the guidance of this specification and without departing from the scope of protection of the claims of the present invention, and all of these are within the scope of protection of the present invention.

Claims

1. A method for estimating the state of charge (SOC) of a lithium-ion battery based on electrochemical impedance spectroscopy and parameter fine-tuning migration, characterized in that, Includes the following steps: S100: Collect electrochemical impedance spectroscopy data of lithium-ion batteries under different states of charge. The electrochemical impedance spectroscopy data includes sampling frequency, real part of impedance and imaginary part of impedance. The electrochemical impedance spectroscopy data of one type of lithium-ion battery is used as the target domain data, and its corresponding SOC is known. The electrochemical impedance spectroscopy data of other types of lithium-ion batteries are used as the target domain data, and their corresponding SOC is unknown. S200: Based on the electrochemical impedance spectroscopy data, extract the equivalent circuit model parameters and relaxation time distribution characteristics, and use correlation analysis to select features that are more correlated with the state of charge than a preset threshold as input features. S300: Construct a neural network model that includes an encoder for extracting abstract representations of input features, a decoder for reconstructing input features, and a predictor for outputting estimates of the state of charge. S400: Normalize the input features of the source domain data and input them into the neural network model for unsupervised or self-supervised pre-training to adjust the model parameters of the encoder and decoder; S500: After pre-training, freeze at least some residual block layers in the encoder, set the hierarchical learning rates of the encoder and predictor, input the input features of the target domain data into the frozen model, perform supervised fine-tuning training, and obtain the trained neural network model for SOC estimation.

2. The method for estimating the state of charge (SOC) of a lithium-ion battery based on electrochemical impedance spectroscopy and parameter fine-tuning migration as described in claim 1, characterized in that, Preferably, the electrochemical impedance spectroscopy data collected from a lithium-ion battery under a certain state of charge. , For sampling frequency, For the real part of the impedance, Let N be the imaginary part of the impedance and N be the number of sampling frequencies. The source domain data can be represented as: ， in, It is the source domain feature vector. It is the number of samples in the source domain; The target domain data can be represented as: ， in, It is the feature vector of the target domain. It is the target domain tag. It is the number of samples in the target domain.

3. The method for estimating the state of charge (SOC) of a lithium-ion battery based on electrochemical impedance spectroscopy and parameter fine-tuning migration as described in claim 2, characterized in that, The equivalent circuit model is fitted using a nonlinear least squares optimization algorithm. The objective function that achieves simultaneous fitting of the real and imaginary parts of the impedance with uniform weights during the fitting process is: ， in, Indicates frequency, Represents frequency The real part of the corresponding original impedance spectrum, Represents frequency The real part of the impedance spectrum of the corresponding equivalent circuit model. Represents frequency The imaginary part of the corresponding original impedance spectrum, Represents frequency The imaginary part of the impedance spectrum of the corresponding equivalent circuit model, where N is the frequency number; If the optimal equivalent circuit model is known, it is directly selected; if the optimal equivalent circuit model is unknown, it is evaluated and selected from the Randles model, the Randles CPE model, the second-order integer circuit, the second-order fractional circuit model, and the user-defined equivalent circuit model. The selection is based on the regression evaluation index of the coefficient of determination, root mean square error, and mean absolute percentage error. After selecting the optimal model based on the fitting effect, the parameters of each circuit element of the equivalent circuit model are obtained by reverse engineering.

4. The method for estimating the state of charge (SOC) of a lithium-ion battery based on electrochemical impedance spectroscopy and parameter fine-tuning migration as described in claim 1, characterized in that, Based on the relaxation time distribution, the electrochemical impedance spectroscopy data are represented as follows: ， in, This is a high-frequency limiting resistor; The relaxation time distribution function represents the contribution weight of different relaxation processes to the total impedance and is an unknown DRT function that needs to be solved. The relaxation time distribution analysis is known. Reverse solving The process involves DRT analysis, which yields a continuous distribution function. Further analysis reveals two characteristics: the high-frequency limiting resistance corresponding to the time constant approaching zero in the relaxation time distribution spectrum. Another characteristic is the total polarization resistance. Its value is a function The integral over the entire time constant characterizes the total energy consumption of all polarization processes within a lithium-ion battery, and the calculation formula is as follows: 。 5. The method for estimating the state of charge (SOC) of a lithium-ion battery based on electrochemical impedance spectroscopy and parameter fine-tuning migration as described in claim 1, characterized in that, The correlation analysis uses Spearman's rank correlation coefficient to calculate the correlation between each extracted feature and the true state of charge. The preset threshold is 0.8, and features with an absolute correlation coefficient greater than 0.8 are selected as input features.

6. The method for estimating the state of charge (SOC) of a lithium-ion battery based on electrochemical impedance spectroscopy and parameter fine-tuning migration according to claim 1, characterized in that, The encoder is composed of an initial convolutional layer and multiple one-dimensional residual blocks connected sequentially. The initial convolutional layer includes a one-dimensional convolutional layer, a batch normalization layer, a ReLU activation layer, and a max pooling layer. Each one-dimensional residual block includes a main path and a shortcut path. The main path contains two one-dimensional convolutional layers and an intermediate batch normalization and activation layer. The shortcut path performs dimension matching through 1×1 convolution when the input and output dimensions are inconsistent. The decoder and encoder have a symmetrical structure and are composed of multiple one-dimensional transposed convolutional layers. The predictor includes a Flatten layer, a fully connected layer, and a ReLU activation layer, and finally outputs a single-value SOC estimate.

7. The method for estimating the state of charge (SOC) of a lithium-ion battery based on electrochemical impedance spectroscopy and parameter fine-tuning migration according to claim 1, characterized in that, When freezing some neural network layers, select the 2nd and 4th residual block layers in the hierarchical residual learning module. The Huber loss function is chosen as the loss function for model training, and its calculation formula is as follows: ， in, Represents the true value. Indicates the predicted value. It is a hyperparameter greater than 0.

8. An evaluation system for estimating the state of charge (SOC) of a lithium-ion battery based on electrochemical impedance spectroscopy and parameter fine-tuning migration as described in any one of claims 1-7, characterized in that, It includes: The acquisition module acquires electrochemical impedance spectroscopy data of lithium-ion batteries under different states of charge. The electrochemical impedance spectroscopy data includes the sampling frequency, the real part of the impedance, and the imaginary part of the impedance. The extraction module extracts the equivalent circuit model parameters and relaxation time distribution characteristics based on the electrochemical impedance spectroscopy data, and selects features with a correlation higher than a preset threshold as input features through correlation analysis. The building module constructs a neural network model that includes an encoder for extracting abstract representations of input features, a decoder for reconstructing input features, and a predictor for outputting estimates of the state of charge. The training module normalizes the input features of the source domain data and inputs them into the neural network model for unsupervised or self-supervised pre-training to adjust the model parameters of the encoder and decoder. In the estimation module, after pre-training, at least some residual block layers in the encoder are frozen, the hierarchical learning rates of the encoder and predictor are set, the input features of the target domain data are input into the frozen model, and supervised fine-tuning training is performed to obtain the trained neural network model for SOC estimation.

9. A computer storage medium, characterized in that, The storage medium includes computer instructions that, when executed on a computer, cause the computer to perform the method as described in any one of claims 1-7.

10. An electronic device, characterized in that, The electronic device includes: Memory, processor, and computer programs stored in memory and executable on the processor, wherein, When the processor executes the program, it implements the method as described in any one of claims 1-7.

Images