Lithium battery health state estimation method based on relaxation voltage and interpretability algorithm

Abstract

The invention provides a lithium battery health state estimation method based on relaxation voltage and an interpretability algorithm, and relates to the technical field of lithium batteries, and the method comprises the steps: carrying out the pre-analysis processing of a lithium battery relaxation voltage data set, converting the relaxation voltage into a two-dimensional grayscale image through a recurrence plot algorithm, and carrying out the calculation of the relaxation voltage. Inputting the grayscale image into a convolutional neural network model based on interpretability algorithm feedback, and performing accurate lithium battery health state estimation; and meanwhile, a second-order equivalent circuit model is constructed to simulate battery aging, a battery aging mechanism represented by relaxation voltage is explained by combining a class activation image generated by an interpretability algorithm, and a guiding significance is provided for a lithium battery health state estimation strategy based on an explanation result, so that the method has a practical application value.

Description

Technical Field

[0001] The present invention relates to the technical field of lithium batteries, and in particular to a method for estimating the health state of a lithium battery based on relaxation voltage and an interpretability algorithm. Background Art

[0002] In recent years, the rapid development of new energy vehicles has driven progress in related industries. Lithium-ion batteries, as core components of new energy electric vehicles, are widely used in new energy vehicles and energy storage power stations due to their high specific energy, high power density, long cycle life, and environmentally friendly advantages. However, during the charge and discharge cycles of lithium-ion batteries, the ongoing chemical reactions can cause corrosion of the battery's internal electrode materials, decomposition and deposition of solid electrolyte interface films, and precipitation of metal materials in the electrolyte. These factors increase the battery's internal resistance, leading to a decrease in battery capacity and a deterioration in its state of health (SOH). Failure to promptly and accurately assess the battery's SOH and replace it promptly can lead to safety accidents such as battery combustion and explosion. Therefore, accurate SOH estimation is crucial to ensuring the safe and efficient operation of lithium-ion batteries. In recent years, many researchers have proposed various SOH estimation methods, which can be summarized into three categories: test-based methods, model-based methods, and data-driven methods. Test-based methods typically use the ampere-hour integration method, the ohmic internal resistance method, and the electrochemical impedance method. While these methods can directly estimate SOH through measurement, they are difficult to measure and are time- and costly, hindering widespread application. Model-based methods typically use an electrochemical model or equivalent circuit model for initialization, then use the least squares method or its improved algorithms for parameter identification, and finally use filtering algorithms to estimate the battery capacity. However, the effectiveness of these methods depends heavily on the quality of the constructed model. Furthermore, before online application, these methods also require appropriate model simplification and accurate parameter identification.

[0003] Data-driven methods can avoid complex electrochemical process analysis and solution processes, directly estimating battery SOH by constructing nonlinear relationships between health characteristics and aging parameters. Most existing methods require extensive charging process data or reconstruction of partial temperature differences or incremental capacity (IC). However, these methods require either a large amount of charging data that is relatively susceptible to charging conditions or complex reconstruction methods and multidimensional data, resulting in considerable complexity and data access difficulties. In comparison, relaxation voltage-based methods offer advantages due to their readily available data and lack of complex feature extraction methods. Furthermore, data-driven methods rely on "black box" models, making it impossible for users to understand how the model makes its decisions and arrives at its results. Failures caused by this unpredictability in risk-sensitive areas such as battery systems can result in significant losses. Therefore, this study's SOH estimation based on relaxation voltage and an interpretable algorithm has significant practical significance. Summary of the Invention

[0004] In order to solve the above technical problems, the goal of the present invention is to provide a lithium battery health state estimation method based on relaxation voltage and interpretable algorithm, which uses the relaxation voltage converted by the recursive graph to accurately estimate the health state of the lithium battery. At the same time, based on the interpretable algorithm, the model is improved and the battery aging mechanism represented by the relaxation voltage is explained, providing guidance for the health state estimation strategy.

[0005] The first technical solution adopted by the present invention is: a method for estimating the health status of a lithium battery based on relaxation voltage and an interpretability algorithm, the method comprising the following steps:

[0006] Obtain charge and discharge cycle data of lithium batteries and obtain relaxation voltage data sets from relaxation process data;

[0007] performing data preprocessing on the relaxation voltage data set to obtain a preprocessed data set;

[0008] Converting the preprocessed data into a two-dimensional grayscale image using a recursive graph algorithm to obtain an image dataset;

[0009] Divide the image dataset into training set, validation set and test set;

[0010] Construct a lithium battery health status estimation model based on convolutional neural network;

[0011] Feeding back the importance weight output by the interpretability algorithm to the lithium battery health state estimation model to obtain the lithium battery health state estimation model fed back by the interpretability algorithm;

[0012] Inputting the training set of the image data set into the lithium battery health state estimation model fed back by the interpretability algorithm for training, thereby obtaining the trained lithium battery health state estimation model fed back by the interpretability algorithm;

[0013] The test set is input into the lithium battery health status estimation model fed back by the trained interpretability algorithm to obtain the estimation results and the class activation image generated by the interpretability algorithm;

[0014] Construct a second-order equivalent circuit model to simulate the aging process of lithium batteries and analyze the battery aging mechanism;

[0015] Combining the battery aging mechanism with the class activation image to explain the battery aging mechanism reflected by the relaxation voltage;

[0016] The analysis of class activation images provides guidance for the health status estimation strategy of lithium batteries.

[0017] In some embodiments, performing data preprocessing on the relaxation voltage dataset to obtain a preprocessed dataset comprises the following steps:

[0018] Performing data cleaning on the relaxation voltage data set to obtain a cleaned charge and discharge data set;

[0019] The cleaned relaxation voltage dataset is smoothed to obtain a preprocessed dataset.

[0020] In some embodiments, converting the preprocessed dataset into a two-dimensional grayscale image using a recursive graph algorithm to obtain an image dataset comprises the following steps:

[0021] Converting the preprocessed data set into a two-dimensional grayscale image using a target recursive graph algorithm to obtain an image data set;

[0022] The expression of the target recursive graph algorithm is:

[0023]

[0024] Among them, R i,j represents a two-dimensional grayscale image, represents the trajectory extracted from the time series data; D represents the given threshold distance; θ(·) represents the Heaviside function; ‖·‖ represents the Euclidean distance function; T represents the number of timestamps in the time series.

[0025] In some embodiments, the construction of a lithium battery health status estimation model based on a convolutional neural network includes the following steps:

[0026] Construct convolutional layers based on a two-dimensional convolutional neural network;

[0027] Construct activation functions based on the LeakyReLU neural network layer;

[0028] Construct a pooling layer based on a two-dimensional Maxpool neural network layer;

[0029] Construct a batch normalization layer based on the two-dimensional Batchnorm neural network layer;

[0030] The lithium battery health status estimation model is constructed based on the convolutional layer, the activation function, the pooling layer, the batch normalization layer and the fully connected layer.

[0031] In some embodiments, constructing the lithium battery health status estimation model based on the convolutional layer, the activation function, the pooling layer, the batch normalization layer, and the fully connected layer includes the following steps:

[0032] The lithium battery health status estimation model is constructed based on the convolution layer, the activation function, the pooling layer, the batch normalization layer and the fully connected layer as follows:

[0033]

[0034] Among them, x c Represents the input feature matrix of the convolutional layer, w (k) It is represented as the weight matrix of the k-th convolution kernel, and the size of the weight matrix is ​​L; t is the convolution step; LeakyReLU(·) is the activation function, α is the leakage coefficient, and its value is between 0 and 1; Maxpool(·) is the maximum pooling layer, and the output of the maximum pooling layer is the maximum value of the input feature map elements; Batchnorm(·) is the batch normalization function, μ is the mean, σ is the variance, ∈ is the minimum number so that the denominator is not 0, and γ and β are learnable parameters.

[0035] In some embodiments, feeding back the importance weight output by the interpretability algorithm to the lithium battery health state estimation model to obtain the lithium battery health state estimation model fed back by the interpretability algorithm includes the following steps:

[0036] The interpretable algorithm GradCAM++ is used to calculate the contribution weight w of the feature map output by the last convolutional layer of the lithium battery health status estimation model based on the convolutional neural network to the model output. k ;

[0037] Contribution weight w k The expression is:

[0038]

[0039] in, Represents the model output; Represents activation map A kThe importance of positions i and j in the matrix; ReLU represents the activation function, and the output is the maximum value between 0 and the input; d and h represent the length and width of the feature map.

[0040] The contribution weight w k Normalized to Then process and obtain the network model weight w;

[0041] The expression of the network model weight w is:

[0042]

[0043] The network model weight w is multiplied by the features output by the last convolutional layer of the lithium battery health status estimation model based on the convolutional neural network to obtain the expression-enhanced features;

[0044] The expression-enhanced features are input into the fully connected layer of the lithium battery health state estimation model based on convolutional neural network to obtain the lithium battery health state estimation model and model output value fed back by the interpretable algorithm.

[0045] In some embodiments, inputting the training set of the image data set into the lithium battery health state estimation model fed back by the interpretability algorithm for training to obtain the trained lithium battery health state estimation model fed back by the interpretability algorithm includes the following steps:

[0046] Design a loss function to calculate the model output value and the true value of the training set to obtain the error value;

[0047] The expression of the loss function is:

[0048]

[0049] Among them, y represents the true value, represents the initial output of the model, represents the output of the model after feedback from the interpretability algorithm; λ represents an adjustable parameter with a value between 0 and 1; represents the cross entropy loss;

[0050] Based on the error value, the model parameters are optimized through the back propagation algorithm and gradient descent method to obtain the optimized state estimation model with minimized error value;

[0051] The training optimization steps are repeated until the predetermined number of training times is reached, and a lithium battery health status estimation model based on the trained interpretable algorithm feedback is obtained.

[0052] In some embodiments, constructing a second-order equivalent circuit model to simulate the aging process of a lithium battery and analyzing the battery aging mechanism includes the following steps:

[0053] A second-order RC equivalent circuit model is established to simulate the battery aging model, with different components representing different chemical reactions of the battery;

[0054] Based on the analysis of experimental results in literature, the parameters of each component of the equivalent circuit model change when the battery ages.

[0055] In some embodiments, inputting the test set into a lithium battery health state estimation model fed back by a trained interpretability algorithm to obtain an estimation result and a class activation image generated by the interpretability algorithm comprises the following steps:

[0056] The test set is input into the lithium battery health state estimation model fed back by the trained interpretability algorithm to obtain the estimated lithium battery health state;

[0057] The estimated lithium battery health status and the gradient and features of the last convolutional layer of the lithium battery health status estimation model fed back by the interpretability algorithm are input into the interpretability algorithm GradCAM++ to generate a class activation image;

[0058] The expression of the explainability algorithm GradCAM++ is:

[0059]

[0060] Among them, w k represents the contribution weight; Represents activation map A k The values ​​at positions i and j in .

[0061] In some embodiments, combining the battery aging mechanism with the class activation image to interpret the battery aging mechanism reflected by the relaxation voltage includes the following steps:

[0062] Corresponding the change of the class activation image as the battery ages to the change of the component parameters of the second-order equivalent circuit model to obtain an analysis result;

[0063] Combined with the analysis results, the battery aging mechanism reflected by the relaxation voltage is explained, and the phenomenon of capacity recovery effect is explained.

[0064] In some embodiments, analyzing the class activation image to provide guidance for a lithium battery health status estimation strategy includes the following steps:

[0065] Analyzing the region on the class activation image that is most strongly correlated with aging, corresponding to the early segment of relaxation voltage;

[0066] Repeat the lithium battery health state estimation step, input the early relaxation voltage segment into the lithium battery health state estimation model fed back by the interpretable algorithm to estimate the lithium battery health state, and shorten the length of the relaxation voltage segment used;

[0067] It is proposed to reconstruct data through the relaxation voltage segment with the strongest correlation, and then estimate the health status of the lithium battery based on the reconstructed relaxation voltage.

[0068] The second technical solution adopted by the present invention is:

[0069] A lithium battery health status estimation system based on relaxation voltage and interpretability algorithm, the system comprising:

[0070] A data acquisition module is used to obtain charge and discharge cycle data of the lithium battery and obtain a relaxation voltage data set from the relaxation process data;

[0071] A data processing module, configured to perform data preprocessing on the relaxation voltage data set to obtain a preprocessed data set;

[0072] a data conversion module, configured to convert the preprocessed data set into a two-dimensional grayscale image by a recursive graph algorithm to obtain an image data set;

[0073] A training set partitioning module is used to partition the image data set into a training set, a validation set, and a test set;

[0074] A model building module for building a lithium battery health status estimation model based on a convolutional neural network;

[0075] A model reconstruction module is used to feed back the importance weight output by the interpretability algorithm to the lithium battery health state estimation model to obtain the lithium battery health state estimation model fed back by the interpretability algorithm;

[0076] A model training module is used to input the training set of the image data set into the lithium battery health state estimation model fed back by the interpretability algorithm for training, so as to obtain a trained lithium battery health state estimation model fed back by the interpretability algorithm;

[0077] A testing module, configured to input the test set into a lithium battery health state estimation model fed back by a trained interpretability algorithm, to obtain an estimation result and a class activation image generated by the interpretability algorithm;

[0078] Equivalent model analysis module, used to construct a second-order equivalent circuit model to simulate the aging process of lithium batteries and analyze the battery aging mechanism;

[0079] an aging mechanism analysis module, configured to combine the battery aging mechanism with the class activation image to interpret the battery aging mechanism reflected by the relaxation voltage;

[0080] The guidance module is used to analyze the class activation image and provide guidance for the lithium battery health status estimation strategy.

[0081] The beneficial effects of the method and system of the present invention are as follows: the present invention pre-analyzes and processes the lithium battery relaxation voltage data set, converts the relaxation voltage into a two-dimensional grayscale image through a recursive graph algorithm, and inputs the grayscale image into a convolutional neural network model based on interpretable algorithm feedback to accurately estimate the health status of the lithium battery; at the same time, a second-order equivalent circuit model is constructed to simulate battery aging, and the class activation image generated by the interpretable algorithm is combined to explain the battery aging mechanism represented by the relaxation voltage, and based on the interpretation results, provides guidance for the lithium battery health status estimation strategy, and therefore has practical application value. BRIEF DESCRIPTION OF THE DRAWINGS

[0082] Figure 1 This is a flowchart of a method for estimating the health status of a lithium battery based on relaxation voltage and an interpretability algorithm according to the present invention;

[0083] Figure 2 This is a structural diagram of a health state estimation model of a lithium battery health state estimation system based on relaxation voltage and interpretability algorithm of the present invention;

[0084] Figure 3 is a flow chart of the interpretability algorithm feedback used in the embodiment of the present invention;

[0085] Figure 4 is a diagram of health status estimation results according to an embodiment of the present invention;

[0086] Figure 5 is a structural diagram of a second-order RC equivalent circuit model constructed in an embodiment of the present invention;

[0087] Figure 6 is a graph showing parameter changes of components in an equivalent circuit model of a document used in an embodiment of the present invention;

[0088] Figure 7 is a class activation image generated by the interpretability algorithm used in the embodiment of the present invention;

[0089] Figure 8 This is a structural block diagram of a lithium battery health status estimation system based on relaxation voltage and explainability algorithm of the present invention. DETAILED DESCRIPTION

[0090] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The step numbers in the following embodiments are provided for ease of description only and do not limit the order of the steps. The order of execution of the steps in the embodiments can be adaptively adjusted based on the understanding of those skilled in the art.

[0091] Reference Figure 1 The present invention provides a method for estimating the health status of a lithium battery based on relaxation voltage and an interpretability algorithm, the method comprising the following steps:

[0092] S1. Obtain charge and discharge cycle data of a lithium battery, and obtain a relaxation voltage data set from the relaxation process data.

[0093] A certain number of samples from mass-produced lithium-ion battery packs are selected as test subjects. Initial performance testing, including capacity, internal resistance, and voltage, is performed on these samples to ensure they meet test standards. Repeated charge and discharge cycles are then used to simulate the performance degradation of lithium-ion batteries over long-term use. BMS sensors or monitoring equipment are used to collect relaxation voltage data every 30 seconds during the battery's charge cycle, which is then used as the data for that charge cycle. Data from multiple consecutive charge cycles is then concatenated to generate the raw data.

[0094] S2. Perform data preprocessing on the relaxation voltage dataset to obtain a preprocessed dataset.

[0095] Data preprocessing of the charge and discharge dataset is a critical step in ensuring the accuracy of subsequent data analysis, model training, and performance evaluation. It aims to extract, clean, transform, and normalize data from the raw relaxation voltage dataset to generate a preprocessed dataset. Raw charge and discharge data, typically including timestamps, voltage, current, capacity, and other parameters, are exported from the battery testing equipment or database. This data is imported into a computer using Python (such as the pandas library) or Excel for subsequent processing. The dataset is inspected for missing values ​​or outliers (such as NaN, Inf, or values ​​significantly deviating from the normal range). Missing values ​​can be addressed through deletion, interpolation, or model-based prediction. Duplicate records in the dataset should be deleted to avoid interference with subsequent analysis. Outliers significantly deviating from the normal range can be identified and addressed using statistical methods (such as the 3σ principle) or business logic. After data cleaning, the cleaned dataset is saved in an appropriate format (such as CSV, Excel, etc.) for subsequent analysis.

[0096] S3. Convert the preprocessed data set into a two-dimensional grayscale image using a recursive graph algorithm to obtain an image data set.

[0097] Specifically, the present invention uses a recurrence graph (RP) algorithm to convert a one-dimensional relaxation voltage curve into a two-dimensional grayscale image. The RP algorithm calculates the distance between the trajectories extracted from the original time series data. The algorithm reveals the internal structure of the data, provides information such as data similarity and predictability, and enriches the data features. It is denoted as x = (x1, x2, ..., x T ) time series data, from which the trajectory is extracted in The RP matrix is ​​calculated using the following formula:

[0098]

[0099] in, Represents the trajectory extracted from the time series data; D represents the given threshold distance; θ(·) represents the Heaviside function; ‖·‖ represents the Euclidean distance function; and T represents the number of timestamps in the time series. The horizontal and vertical axes of the converted RP matrix represent the timestamps of the original time series data, and each element of the matrix represents the Euclidean distance between the values ​​of two timestamps calculated according to the RP algorithm. The resulting RP matrix contains the texture and topological information of the time series data, which exhibits characteristic repetitive patterns at both small and large scales. In this invention, we omit the threshold processing step and directly convert the RP matrix into a grayscale image.

[0100] S4. Divide the image dataset into a training set, a validation set, and a test set.

[0101] Specifically, the present invention uses two data sets for experiments: a self-test data set and an open source data set. The self-test data set contains 8 batteries, denoted as Cell 1 to Cell 8. K-fold cross-validation is used, with one battery used as the test set, 80% of the battery data from the remaining 7 batteries used as the training set, and 20% of the battery data used as the test set. A total of 8 sets of training / validation / test sets can be obtained. For the open source data set, it contains 23 batteries, denoted as CY 1 to CY 23. 20 of these batteries are selected and evenly divided into 5 groups, each containing 4 batteries, denoted as CYs I to CYs V. K-fold cross-validation is also used, with one group of batteries used as the test set, 80% of the battery data from the remaining 4 groups used as the training set, and 20% of the battery data used as the test set. A total of 5 sets of training / validation / test sets can be obtained. Through multiple sets of experiments with different battery data sets, the effectiveness and generalization ability of the model can be judged, and necessary adjustments and improvements can be made.

[0102] S5. Construct a lithium battery health status estimation model based on convolutional neural network.

[0103] In practical applications, the charging and discharging conditions of electric vehicles are quite complex. The diversity of multi-level current charging strategies and the uncertainty of the charging range pose challenges to methods for estimating the state of health of lithium batteries based on charging data. In comparison, the relaxation voltage is relatively unaffected by the charging process, and the data is readily available. Furthermore, previous studies have shown that the relaxation voltage is closely related to battery aging. Therefore, methods for estimating the state of health (SOH) based on the relaxation voltage are more easily implemented in real-world scenarios. However, existing methods lack estimation accuracy, and the "black box" nature of deep learning models makes it impossible for users to understand how the model makes its decisions. To address these issues, a CNN-based partial charging curve reconstruction model is employed. This model reconstructs charging data from a low-voltage range to a high-voltage range, thereby enhancing the correlation between the extracted HFs and the SOH. By capturing the local features of the input low-correlation range segment data through convolution operations, the CNN model achieves accurate reconstruction.

[0104] Reference Figure 2 The lithium battery health state estimation model based on convolutional neural network consists of six convolution modules, a flattening layer and a fully connected layer. The convolution layer in each convolution module adopts the convolution kernel size kernel_size = 3; the activation function used is LeakyReLU, which is used to enhance the nonlinear fitting ability of the model and avoid overfitting; the batch normalization layer transforms the feature distribution output by the intermediate layer into a distribution of the same type with a mean of 0 and a variance of 1, which is used to optimize the network gradient flow, accelerate the training process and improve the model performance; the maximum pooling layer is used to reduce the spatial size of the feature map to improve the model operation efficiency; the flattening layer adjusts the features extracted by the convolution module to an input size suitable for the fully connected layer; and the last fully connected layer is used to output the final SOH estimation result. The formula of the lithium battery health state estimation model based on convolutional neural network is expressed as follows:

[0105]

[0106] Among them, x c Represents the input feature matrix of the convolutional layer, w (k) It is represented as the weight matrix of the k-th convolution kernel, whose size is L; t is the convolution step; LeakyReLU(·) is the activation function, α is the leakage coefficient, and its value is between 0 and 1; Maxpool(·) is the maximum pooling layer, whose output is the maximum value of the input feature map elements; Batchnorm(·) is the batch normalization function, μ is the mean, σ is the variance, ∈ is the minimum number to prevent the denominator from being 0, and γ and β are learnable parameters.

[0107] S6. Feedback the importance weight output by the interpretability algorithm to the lithium battery health state estimation model to obtain the lithium battery health state estimation model fed back by the interpretability algorithm.

[0108] Reference Figure 3 , the lithium battery health status estimation model based on convolutional neural network is decoupled into an encoder and a predictor. The encoder contains the six convolution modules mentioned above to extract input features, and the encoder contains a flattening layer and a connection layer to estimate SOH. The RP matrix is ​​input into the model to obtain a preliminary estimated SOH, that is, Then, the contribution weight w is calculated according to the interpretability algorithm GradCAM++ k , its formula is as follows:

[0109]

[0110] in, Represents the model output; Represents activation map A k The importance of position i and j in the matrix; ReLU represents the activation function, and the output is the maximum value between 0 and input; d and h represent the length and width of the feature map. Then the contribution weight w k Normalized to Then process it to get the network model weight w, which is expressed as follows:

[0111]

[0112] The network model weight w is multiplied by the feature extracted by the encoder to obtain the expression enhancement feature, which is then input into the predictor for estimation to obtain the SOH estimate after feedback from the interpretability algorithm. Finally, a loss function is designed to calculate the model output value and the true value of the training set. Its formula is as follows:

[0113]

[0114] Among them, y represents the true value, represents the initial output of the model, represents the output of the model after feedback from the interpretability algorithm; λ represents an adjustable parameter with a value between 0 and 1; Denotes the cross entropy loss. The above design results in a lithium battery health status estimation model with interpretable algorithm feedback.

[0115] S7. Input the training set of the image data set into the lithium battery health state estimation model fed back by the interpretability algorithm for training, to obtain a trained lithium battery health state estimation model fed back by the interpretability algorithm.

[0116] The training set of the two-dimensional grayscale image dataset after the relaxation voltage is converted by the RP algorithm is input into the lithium battery health state estimation model fed back by the interpretability algorithm. The validation set is used to verify whether the model training effect meets the standard, and the trained lithium battery health state estimation model fed back by the interpretability algorithm is obtained. In the specific example of the present invention, the dimension of the input dataset is cycle_size*feature_size*feature_size, where cycle_size is the number of charging cycles of each battery, feature size = 61 is the dimension of the input features. K-fold cross-validation is used to evaluate the performance of the reconstructed model. The error between the model output and the true value of the training set is calculated using the loss function. Three error functions are selected as indicators for model evaluation. The formulas are as follows:

[0117] Root mean square error:

[0118] Mean absolute error:

[0119] Mean absolute percentage error:

[0120] in is the estimated value of the final model, y i is the true value of SOH, and N is the number of samples.

[0121] The specific test of the embodiment of the present invention is to select the Adam optimizer for model training, with a learning rate of lr = 0.001, where the learning rate is decayed by a decay coefficient of 0.5 after every 30 training cycles, and the number of iterations epoch = 100. Based on the BP algorithm, which uses forward information propagation and backward error propagation, the model weight parameters are updated and trained. The training optimization steps are repeated, and when the loss decreases, the validation set is input for verification. If the model's evaluation indicators on the validation set have improved, the model is saved; if not, training is continued. If the model loss does not decrease within 40 training cycles after saving the model, training is stopped to reduce computational costs.

[0122] S8. Input the test set into the lithium battery health status estimation model fed back by the trained interpretability algorithm to obtain the estimation result and the class activation image generated by the interpretability algorithm.

[0123] The image dataset test set is input into the lithium battery health status estimation model fed back by the trained interpretability algorithm, and the estimation results are compared as follows: Figure 4 As shown in the figure, it can be observed that the SOH estimation result after feedback from the explainability algorithm is closer to the actual SOH value than that without feedback, and the error is significantly reduced, which proves the effectiveness of the explainability feedback algorithm proposed in this invention. Figure 4(a) shows the SOH curve of Cell 2. Figure 4 (b) to (e) are class activation images under different SOH on the SOH curve. Figure 4 (f) shows the SOH curve of battery CY 5. Figure 4 (g) to (j) in the figure show the class activation images at different SOH levels on the SOH curve. The horizontal and vertical axes of the class activation images represent the timestamps of the original relaxation voltage curve. The color of the region represents the correlation with battery aging, with colors closer to dark red indicating stronger correlation and colors closer to dark blue indicating weaker correlation.

[0124] By analyzing the class activation image, we can know which part of the relaxation voltage the model focuses on to obtain the SOH estimation result, which facilitates the subsequent explanation of the battery aging mechanism.

[0125] S9. Construct a second-order equivalent circuit model to simulate the aging process of lithium batteries and analyze the battery aging mechanism.

[0126] The constructed second-order equivalent circuit model is as follows Figure 5 As shown in Figure 1, the equivalent model circuit simulates the battery aging model. Each component corresponds to a different chemical reaction inside the battery, and the change in component parameters represents the change in the effect of different chemical reactions on battery aging. Among them, the resistor R0 represents the ohmic depolarization process of the battery, which is represented by the resistor R a and capacitor C a The parallel network formed by the resistor R represents the electrochemical depolarization process of the battery. c and capacitor C c The resulting parallel network represents the battery's concentration depolarization process. With respect to the relaxation voltage, since the current is zero at this point, the ohmic depolarization process ends instantaneously. Therefore, the relaxation process is primarily influenced by electrochemical depolarization and concentration depolarization. According to electrochemical principles, the electrochemical depolarization process typically lasts about 10 seconds, while concentration depolarization continues throughout the relaxation process.

[0127] As the battery ages, the resistance changes representing different depolarization processes are as follows Figure 6 As shown. In the early stage of battery aging, R0 and R a It gradually decreases at the beginning, and R c Gradually increase. If R0 and R a The reduction degree is greater than R c The increase is large. At this time, the internal resistance of the battery will decrease, the capacity will increase, and the so-called "capacity recovery" effect will occur. At this time, the relaxation process is dominated by the electrochemical depolarization process. As the battery ages further, R0 and R c Gradually increases, R aThe battery's internal resistance gradually increases, the capacity gradually declines, and the relaxation process is dominated by the concentration depolarization process. Based on the above analysis, we hope to observe this phenomenon from the class activation image, realizing the observation of the battery aging mechanism based on the external characteristics of the battery.

[0128] S10. Combining the battery aging mechanism with the activation-like image, interpreting the battery aging mechanism reflected by the relaxation voltage.

[0129] Class activation images such as Figure 7 As shown, Figure 7 (a) shows the SOH curve of Cell 2. Figure 7 (b) to (e) are class activation images under different SOH on the SOH curve. Figure 7 (f) shows the SOH curve of battery CY 5. Figure 7 (g) to (j) in the figure show the class activation images at different SOH levels on the SOH curve. The horizontal and vertical axes of the class activation images represent the timestamps of the original relaxation voltage curve. The color of the region represents the correlation with battery aging, with colors closer to dark red indicating stronger correlation and colors closer to dark blue indicating weaker correlation.

[0130] For Cell 2, the “capacity recovery” phenomenon can be observed, indicating that for this battery in the early stage of battery aging, R0 and R a The reduction degree is greater than R c The increase is large, resulting in a decrease in the battery's internal resistance and an increase in capacity. At this point, the electrochemical polarization process dominates the relaxation voltage. The 10-second square region in the upper left corner of the class activation map represents the combined effects of electrochemical depolarization and concentration depolarization. However, this region consistently appears dark red, making it difficult to discern the changing trend. Conversely, the dominance of electrochemical depolarization indicates a weakening influence of concentration depolarization, leading to a gradual weakening of the correlation with the concentration depolarization region on the class activation map, corresponding to region S1. Figure 7 (b) and (c) in the figure show that the color of the region gradually becomes lighter, indicating that the effect of concentration depolarization becomes smaller. As the battery ages further, R0 and R c Gradually increases, R a Keep it steady, at this time the internal resistance of the battery gradually increases, the capacity gradually declines, and the relaxation process is dominated by the concentration depolarization process. Figure 7 In Figures (d) and (e), we can observe that the color of region S1 gradually brightens, indicating that the influence of concentration depolarization increases, which is consistent with the battery aging mechanism. In addition, we can also observe that the range of strong correlation in region S2 first increases and then decreases, indicating that the range of concentration depolarization process, which is strongly correlated with battery aging, first increases and then decreases. For the batteries in the self-test data set, the relaxation voltage correlation is always strong in the first fifteen minutes.

[0131] For CY 5, there is no “capacity recovery” phenomenon, indicating that for this battery, R c The increase is always greater than R a The decrease in is greater, so concentration depolarization always dominates. Figure 7 From (g) to (e), we can observe that region S3 corresponding to concentration depolarization gradually brightens, indicating that the influence of concentration depolarization gradually becomes stronger, which is consistent with theoretical analysis. We can also observe that region S4, indicating a gradually increasing range of strong correlations, indicates that the range of the concentration depolarization process, which is strongly correlated with battery aging, first increases and then decreases. Furthermore, for batteries in the public dataset, the relaxation voltage correlation remains strong for the first six minutes.

[0132] In summary, the battery aging mechanism represented by the relaxation voltage can be explained by analyzing the activation-like images. It can also be clearly seen which specific area of ​​the battery depolarization reaction the model focuses on to make decisions.

[0133] S11. Analyze the class activation image to provide guidance for the lithium battery health status estimation strategy.

[0134] Based on the above analysis, for the batteries in the public data set, the correlation of the relaxation voltage in the first six minutes is always very strong, so the health status of the lithium battery can be estimated using only the relaxation voltage fragments in the first six minutes. In order to verify the effectiveness of the method proposed in the present invention, a comparative experiment was designed to perform health status estimation based on the relaxation fragments of the first four minutes, six minutes, and eight minutes, respectively. The RMSE (%) of the estimation results is shown in Table 1. The experimental results show that accurate SOH estimation can be achieved based on the relaxation voltage of the six-minute fragment, while the estimation based on the four-minute fragment has a decrease in accuracy due to the short fragment length and the lack of some important information. The estimation based on the eight-minute fragment is better than that based on the six-minute fragment, but the improvement is smaller, indicating that the first six-minute relaxation fragment already contains sufficient battery aging information, which is consistent with the above explanation.

[0135]

[0136] Table 1

[0137] In addition, the present invention also compares SOH estimation based on six-minute segments distributed in different relaxation voltage intervals, namely [0,6], [6,12], [12,18] and [18,24]. The experimental results shown in Table 2 show that the estimation effect using the [0,6] segment is the best, and the estimation accuracy of other segments is significantly reduced. Although it is easy to subjectively believe that the end relaxation voltage changes the most from the trend of the relaxation voltage curve changing with battery aging, and therefore the end relaxation voltage should be paid more attention to, the experimental results and the class activation image show that the front relaxation voltage is more worthy of attention. Therefore, analysis through interpretable algorithms can avoid similar subjective misjudgments and has practical application value.

[0138]

[0139] Table 2

[0140] like Figure 8 As shown, the present invention provides a lithium battery health status estimation system based on relaxation voltage and interpretability algorithm, the system comprising:

[0141] A data acquisition module is used to obtain charge and discharge cycle data of the lithium battery and obtain a relaxation voltage data set from the relaxation process data;

[0142] A data processing module, configured to perform data preprocessing on the relaxation voltage data set to obtain a preprocessed data set;

[0143] a data conversion module, configured to convert the preprocessed data set into a two-dimensional grayscale image by a recursive graph algorithm to obtain an image data set;

[0144] A training set partitioning module is used to partition the image data set into a training set, a validation set, and a test set;

[0145] A model building module for building a lithium battery health status estimation model based on a convolutional neural network;

[0146] A model reconstruction module is used to feed back the importance weight output by the interpretability algorithm to the lithium battery health state estimation model to obtain the lithium battery health state estimation model fed back by the interpretability algorithm;

[0147] A model training module is used to input the training set of the image data set into the lithium battery health state estimation model fed back by the interpretability algorithm for training, so as to obtain a trained lithium battery health state estimation model fed back by the interpretability algorithm;

[0148] A testing module, configured to input the test set into a lithium battery health state estimation model fed back by a trained interpretability algorithm, to obtain an estimation result and a class activation image generated by the interpretability algorithm;

[0149] Equivalent model analysis module, used to construct a second-order equivalent circuit model to simulate the aging process of lithium batteries and analyze the battery aging mechanism;

[0150] an aging mechanism analysis module, configured to combine the battery aging mechanism with the class activation image to interpret the battery aging mechanism reflected by the relaxation voltage;

[0151] The guidance module is used to analyze the class activation image and provide guidance for the lithium battery health status estimation strategy.

[0152] The contents of the above method embodiments are all applicable to the present system embodiments. The functions specifically implemented by the present system embodiments are the same as those of the above method embodiments, and the beneficial effects achieved are also the same as those achieved by the above method embodiments.

[0153] The above is a specific description of the preferred implementation of the present invention, but the invention is not limited to the embodiments. Those skilled in the art can make various equivalent modifications or substitutions without violating the spirit of the present invention. These equivalent modifications or substitutions are all included in the scope defined by the claims of this application.

Claims

1. A lithium battery health status estimation method based on relaxation voltage and interpretability algorithm, characterized in that: The method comprises the following steps: Obtain charge and discharge cycle data of lithium batteries and obtain relaxation voltage data sets from relaxation process data; performing data preprocessing on the relaxation voltage data set to obtain a preprocessed data set; Converting the preprocessed data set into a two-dimensional grayscale image using a recursive graph algorithm to obtain an image data set; Dividing the image dataset into a training set, a validation set, and a test set; Construct a lithium battery health status estimation model based on convolutional neural network; Feeding back the importance weight output by the interpretability algorithm to the lithium battery health state estimation model to obtain the lithium battery health state estimation model fed back by the interpretability algorithm; Inputting the training set of the image data set into the lithium battery health state estimation model fed back by the interpretability algorithm for training, thereby obtaining a trained lithium battery health state estimation model fed back by the interpretability algorithm; Inputting the test set into the lithium battery health state estimation model fed back by the trained interpretability algorithm to obtain an estimation result and a class activation image generated by the interpretability algorithm; Construct a second-order equivalent circuit model to simulate the aging process of lithium batteries and analyze the battery aging mechanism; Combining the battery aging mechanism with the class activation image to explain the battery aging mechanism reflected by the relaxation voltage; Analyzing the class activation images provides guidance for the health status estimation strategy of lithium batteries.

2. The method for estimating the health status of a lithium battery based on relaxation voltage and interpretability algorithm according to claim 1, characterized in that: The method of converting the preprocessed data set into a two-dimensional grayscale image by a recursive graph algorithm to obtain an image data set includes the following steps: Converting the preprocessed data set into a two-dimensional grayscale image using a target recursive graph algorithm to obtain an image data set; The expression of the target recursive graph algorithm is: Among them, R i,j represents a two-dimensional grayscale image, represents the trajectory extracted from the time series data; D represents the given threshold distance; θ(·) represents the Heaviside function; ‖·‖ represents the Euclidean distance function; T represents the number of timestamps in the time series.

3. The method for estimating the health status of a lithium battery based on relaxation voltage and interpretability algorithm according to claim 1, wherein: The construction of a lithium battery health status estimation model based on a convolutional neural network includes the following steps: Construct convolutional layers based on a two-dimensional convolutional neural network; Construct activation functions based on the LeakyReLU neural network layer; Construct a pooling layer based on a two-dimensional Maxpool neural network layer; Construct a batch normalization layer based on the two-dimensional Batchnorm neural network layer; The lithium battery health status estimation model is constructed based on the convolutional layer, the activation function, the pooling layer, the batch normalization layer and the fully connected layer.

4. The method for estimating the health status of a lithium battery based on relaxation voltage and interpretability algorithm according to claim 3, characterized in that: The lithium battery health state estimation model is constructed based on the convolutional layer, the activation function, the pooling layer, the batch normalization layer and the fully connected layer, comprising the following steps: The lithium battery health status estimation model is constructed based on the convolution layer, the activation function, the pooling layer, the batch normalization layer and the fully connected layer as follows: Among them, x c Represents the input feature matrix of the convolutional layer, w (k) It is represented as the weight matrix of the k-th convolution kernel, and the size of the weight matrix is ​​L; t is the convolution step; LeakyReLU(·) is the activation function, α is the leakage coefficient, and its value is between 0 and 1; Maxpool(·) is the maximum pooling layer, and the output of the maximum pooling layer is the maximum value of the input feature map elements; Batchnorm(·) is the batch normalization function, μ is the mean, σ is the variance, ∈ is the minimum number so that the denominator is not 0, and γ and β are learnable parameters.

5. The method for estimating the health status of a lithium battery based on relaxation voltage and interpretability algorithm according to claim 1, characterized in that: Feeding back the importance weight output by the interpretability algorithm to the lithium battery health state estimation model to obtain the lithium battery health state estimation model fed back by the interpretability algorithm includes the following steps: The interpretable algorithm GradCAM++ is used to calculate the contribution weight w of the feature map output by the last convolutional layer of the lithium battery health status estimation model based on the convolutional neural network to the model output. k ; Contribution weight w k The expression is: in, Represents the model output; Represents activation map A k The importance of positions i and j in the matrix; ReLU represents the activation function, and the output is the maximum value between 0 and the input; d and h represent the length and width of the feature map. The contribution weight w k Normalized to Then process and obtain the network model weight w; The expression of the network model weight w is: The network model weight w is multiplied by the features output by the last convolutional layer of the lithium battery health status estimation model based on the convolutional neural network to obtain the expression-enhanced features; The expression-enhanced features are input into the fully connected layer of the lithium battery health state estimation model based on convolutional neural network to obtain the lithium battery health state estimation model and model output value fed back by the interpretable algorithm.

6. The method for estimating the health status of a lithium battery based on relaxation voltage and interpretability algorithm according to claim 5, characterized in that: Inputting the training set of the image data set into the lithium battery health state estimation model fed back by the interpretability algorithm for training to obtain the trained lithium battery health state estimation model fed back by the interpretability algorithm includes the following steps: Design a loss function to calculate the model output value and the true value of the training set to obtain the error value; The expression of the loss function is: Among them, y represents the true value, represents the initial output of the model, represents the output of the model after feedback from the interpretability algorithm; λ represents an adjustable parameter with a value between 0 and 1; represents the cross entropy loss; Based on the error value, the model parameters are optimized through the back propagation algorithm and gradient descent method to obtain the optimized state estimation model with minimized error value; The training optimization steps are repeated until the predetermined number of training times is reached, and a lithium battery health status estimation model based on the trained interpretable algorithm feedback is obtained.

7. The method for estimating the health status of a lithium battery based on relaxation voltage and interpretability algorithm according to claim 1, characterized in that: Inputting the test set into the lithium battery health status estimation model fed back by the trained interpretability algorithm to obtain the estimation result and the class activation image generated by the interpretability algorithm includes the following steps: The test set is input into the lithium battery health state estimation model fed back by the trained interpretability algorithm to obtain the estimated lithium battery health state; The estimated lithium battery health status and the gradient and features of the last convolutional layer of the lithium battery health status estimation model fed back by the interpretability algorithm are input into the interpretability algorithm GradCAM++ to generate a class activation image; The expression of the explainability algorithm GradCAM++ is: Among them, w k represents the contribution weight; Represents activation map A k The values ​​at positions i and j in .

8. The method for estimating the health status of a lithium battery based on relaxation voltage and interpretability algorithm according to claim 1, characterized in that: The method of combining the battery aging mechanism with the class activation image to interpret the battery aging mechanism reflected by the relaxation voltage includes the following steps: Corresponding the change of the class activation image as the battery ages to the change of the component parameters of the second-order equivalent circuit model to obtain an analysis result; Combined with the analysis results, the battery aging mechanism reflected by the relaxation voltage is explained, and the phenomenon of capacity recovery effect is explained.

9. The method for estimating the health status of a lithium battery based on relaxation voltage and interpretability algorithm according to claim 1, characterized in that: The analyzing of the class activation image to provide guidance for the lithium battery health status estimation strategy includes the following steps: Analyzing the region on the class activation image that is most strongly correlated with aging, corresponding to the early segment of relaxation voltage; Repeat the lithium battery health state estimation step, input the early relaxation voltage segment into the lithium battery health state estimation model fed back by the interpretable algorithm to estimate the lithium battery health state, and shorten the length of the relaxation voltage segment used; It is proposed to reconstruct data through the relaxation voltage segment with the strongest correlation, and then estimate the health status of the lithium battery based on the reconstructed relaxation voltage.

10. A lithium battery health status estimation system based on relaxation voltage and interpretability algorithm, characterized in that: The system comprises: A data acquisition module is used to obtain charge and discharge cycle data of the lithium battery and obtain a relaxation voltage data set from the relaxation process data; A data processing module, configured to perform data preprocessing on the relaxation voltage data set to obtain a preprocessed data set; a data conversion module, configured to convert the preprocessed data set into a two-dimensional grayscale image by a recursive graph algorithm to obtain an image data set; A training set partitioning module is used to partition the image data set into a training set, a validation set, and a test set; A model building module for building a lithium battery health status estimation model based on a convolutional neural network; A model reconstruction module is used to feed back the importance weight output by the interpretability algorithm to the lithium battery health state estimation model to obtain the lithium battery health state estimation model fed back by the interpretability algorithm; A model training module is used to input the training set of the image data set into the lithium battery health state estimation model fed back by the interpretability algorithm for training, so as to obtain a trained lithium battery health state estimation model fed back by the interpretability algorithm; A testing module, configured to input the test set into a lithium battery health state estimation model fed back by a trained interpretability algorithm, to obtain an estimation result and a class activation image generated by the interpretability algorithm; Equivalent model analysis module, used to construct a second-order equivalent circuit model to simulate the aging process of lithium batteries and analyze the battery aging mechanism; an aging mechanism analysis module, configured to combine the battery aging mechanism with the class activation image to interpret the battery aging mechanism reflected by the relaxation voltage; The guidance module is used to analyze the class activation image and provide guidance for the lithium battery health status estimation strategy.

Images