Method for optimizing lithium battery charging process by multi-input neural network model

Abstract

This application relates to a training method for a multi-input neural network model for optimizing the lithium battery charging process. The method includes: during a charging test of a lithium battery test cell equipped with a reference electrode, collecting a dataset of the reference voltage and electrochemical impedance spectroscopy (EIS) of the lithium battery test cell; dividing the collected dataset into a training set, a validation set, and a test set; training the multi-input neural network model using the training set; evaluating the performance parameters of the trained multi-input neural network model using the validation set; and completing the training process only when the mean square error between the predicted reference voltage output by the trained multi-input neural network model based on the EIS in the validation set and the reference voltage in the validation set is less than a preset threshold. This application also relates to a method for optimizing the lithium battery charging process using a multi-input neural network model, a battery management system, a vehicle, and a computer program product.

Description

Technical Field

[0001] This application relates to the field of lithium battery charging, and in particular to a training method for a multi-input neural network model for optimizing the lithium battery charging process, a method for optimizing the lithium battery charging process through a multi-input neural network model, a battery management system, a vehicle including the battery management system according to this application, and a computer program product. Background Technology

[0002] Lithium-ion batteries are widely used in new energy vehicles. However, a common issue in their actual use is the gradual degradation of their performance. One major reason for this is the lithium deposited at the negative electrode during fast charging. This lithium not only reacts with the electrolyte but also gradually develops into dead lithium, ultimately leading to irreversible loss of active lithium and battery capacity. Furthermore, the lithium dendrites formed during lithium deposition may penetrate the separator, causing internal short circuits and other safety issues.

[0003] Therefore, how to detect and provide early warning of lithium plating on the negative electrode of lithium batteries has become a technical challenge that needs to be solved. Summary of the Invention

[0004] The purpose of this application is to provide a training method for a multi-input neural network model for optimizing the lithium battery charging process, a method for optimizing the lithium battery charging process through a multi-input neural network model, a battery management system, a vehicle including the battery management system according to this application, and a computer program product, to at least partially solve the problems in the prior art.

[0005] According to a first aspect of this application, a training method for a multi-input neural network model for optimizing the lithium battery charging process is provided, the training method comprising: - During the charging test of a lithium battery test cell equipped with a reference electrode, data sets of reference voltage and electrochemical impedance spectroscopy of the lithium battery test cell are collected; The collected dataset is divided into a training set, a validation set, and a test set. The multi-input neural network model is trained using the training set, and the performance parameters of the trained multi-input neural network model are evaluated using the validation set. The training process is completed when the mean square error between the reference voltage prediction value output by the trained multi-input neural network model based on the electrochemical impedance spectroscopy in the validation set and the reference voltage in the validation set is less than a preset threshold.

[0006] The core concept of this application includes at least the following: introducing a reference electrode into the structure of a lithium battery test cell to form a three-electrode battery for monitoring the negative electrode potential; dividing the dataset of reference voltage and electrochemical impedance spectrum of the lithium battery test cell collected during the charging test into a training set, a validation set, and a test set; using the training set to train a multi-input neural network model; and using the validation set to evaluate the performance parameters of the trained multi-input neural network model so that the performance parameters of the trained multi-input neural network model meet the requirements.

[0007] According to an optional embodiment of this application, the method may further include: - Evaluate the performance parameters of the trained multi-input neural network model using the test set, wherein the performance parameters may include the mean square error and the coefficient of determination between the reference voltage prediction output by the trained multi-input neural network model based on the electrochemical impedance spectroscopy in the test set and the reference voltage in the test set.

[0008] According to another optional embodiment of this application, the collected electrochemical impedance spectroscopy data may include the amplitude and phase angle of the electrochemical impedance spectrum of the lithium battery test cell, or the real and imaginary parts of the electrochemical impedance spectrum of the lithium battery test cell, or battery model parameters determined based on the electrochemical impedance spectrum of the lithium battery test cell, and key frequency points and / or key frequency bands related to the battery model parameters are marked in the collected electrochemical impedance spectrum, wherein the battery model parameters include one or more of the following parameters, for example: battery internal resistance, charge transfer resistance, double layer capacitance and diffusion impedance.

[0009] According to another optional embodiment of this application, during the charging test of the lithium battery test cell at various charging rates, a dataset of reference voltage and electrochemical impedance spectrum of the lithium battery test cell can be collected, and the collected dataset can be labeled with the corresponding charging rate of the charging test.

[0010] According to another optional embodiment of this application, during the charging test of lithium battery test cells at various ambient temperatures, a dataset of reference voltage and electrochemical impedance spectroscopy of the lithium battery test cells can be collected, and the collected dataset can be labeled with the corresponding ambient temperature of the charging test.

[0011] According to another optional embodiment of this application, during the charging test of the lithium battery test cell, a dataset of the reference voltage and electrochemical impedance spectrum of the lithium battery test cell, as well as the state of charge and / or battery health of the lithium battery test cell, can be collected, and the collected dataset can be labeled with the corresponding state of charge and / or battery health.

[0012] According to another optional embodiment of this application, during the charging test of the lithium battery test cell, the electrochemical impedance of the lithium battery test cell can be measured based on the state of charge of the battery, or the electrochemical impedance of the lithium battery test cell can be measured dynamically in situ.

[0013] According to another optional embodiment of this application, the method may further include: - Match the reference voltage and electrochemical impedance spectrum in the collected dataset; - Clean the abnormal data in the collected dataset and normalize the cleaned dataset. The cleaning methods include, for example, noise reduction and / or outlier removal.

[0014] According to another optional embodiment of this application, hyperparameters of a multi-input neural network model can be set, such as learning rate, batch size, number of iterations, number of hidden layer neurons, and / or dropout rate; during the training of the multi-input neural network model, the hyperparameters of the multi-input neural network model can be adjusted based on the performance parameters of the multi-input neural network model evaluated using the validation set.

[0015] According to another optional embodiment of this application, a patience value and a minimum improvement threshold can be set for the training process of a multi-input neural network model; the multi-input neural network model is trained using the training set, and the performance parameters of the trained multi-input neural network model are evaluated using the validation set; if the change in the performance parameters relative to the historical best performance parameters of the multi-input neural network model is less than the minimum improvement threshold, one no-improvement epoch is accumulated, and the model parameters of the multi-input neural network model are not saved; if the change in the performance parameters relative to the historical best performance parameters of the multi-input neural network model is greater than or equal to the minimum improvement threshold, the model parameters of the multi-input neural network model are saved; if the number of consecutively accumulated no-improvement epochs is greater than or equal to the patience value, the training process of the multi-input neural network model is stopped.

[0016] According to a second aspect of this application, a method is provided for optimizing a lithium battery charging process using a multi-input neural network model, wherein the multi-input neural network model is trained using the training method according to any one of the preceding claims, the method comprising: - Collect the electrochemical impedance spectroscopy of the lithium battery cell during the charging process; - The reference voltage of the lithium battery cell is predicted at least based on the acquired electrochemical impedance spectroscopy using a trained multi-input neural network model; - Monitor the charging boundary conditions of the lithium battery cell based on the predicted reference voltage.

[0017] The core concept of this application includes at least the following: during the lithium battery charging process, the reference voltage of the lithium battery cell is predicted by a trained multi-input neural network model based on the collected electrochemical impedance spectrum, and the charging boundary conditions of the lithium battery cell are automatically monitored based on the predicted reference voltage. When the charging boundary conditions are detected, a prompt message is sent and / or the charging process is automatically terminated, thereby optimizing the lithium battery charging process, preventing lithium plating in the lithium battery cell during charging, and effectively extending the service life of the lithium battery.

[0018] According to another optional embodiment of this application, when the predicted reference voltage is less than or equal to zero, the lithium battery cell can be monitored to reach the charging boundary condition of lithium plating.

[0019] According to another optional embodiment of this application, the collected electrochemical impedance spectroscopy data may include the amplitude and phase angle of the electrochemical impedance spectrum of the lithium battery cell, or the real and imaginary parts of the electrochemical impedance spectrum of the lithium battery cell, or battery model parameters determined based on the electrochemical impedance spectrum of the lithium battery cell, wherein the battery model parameters include one or more of the following parameters, such as battery internal resistance, charge transfer resistance, double layer capacitance, and diffusion impedance.

[0020] According to another optional embodiment of this application, key frequency points and / or key frequency bands related to battery model parameters are marked in the collected electrochemical impedance spectroscopy, and the reference voltage of the lithium battery cell is predicted by a trained multi-input neural network model based at least on the collected electrochemical impedance spectroscopy and the key frequency points and / or the key frequency bands.

[0021] According to another optional embodiment of this application, the charging rate of the lithium battery cell during the charging process can be obtained, and the reference voltage of the lithium battery cell can be predicted by a trained multi-input neural network model based at least on the charging rate of the lithium battery cell during the charging process and the collected electrochemical impedance spectrum.

[0022] According to another optional embodiment of this application, the ambient temperature during the charging process of the lithium battery cell can be collected, and the reference voltage of the lithium battery cell can be predicted by a trained multi-input neural network model based at least on the ambient temperature during the charging process of the lithium battery cell and the collected electrochemical impedance spectrum.

[0023] According to another optional embodiment of this application, the state of charge and / or state of health of the lithium battery cell can be collected during the charging process of the lithium battery cell, and the reference voltage of the lithium battery cell can be predicted by a trained multi-input neural network model based on the collected electrochemical impedance spectrum and the state of charge and / or state of health of the lithium battery cell during the charging process.

[0024] According to another optional embodiment of this application, the method may further include: - If the lithium battery cell is detected to have reached the charging boundary condition for lithium plating, a warning message about lithium plating in the lithium battery cell is sent and / or the charging process of the lithium battery cell is terminated.

[0025] According to a third aspect of this application, a battery management system is provided, the battery management system including a control unit for performing a method according to this application for optimizing a lithium battery charging process using a multi-input neural network model.

[0026] According to another optional embodiment of this application, the control unit may include at least one processor and a memory, wherein the memory stores program instructions executable by the at least one processor, and when the program instructions are executed by the at least one processor, a multi-input neural network model trained by the training method for optimizing the lithium battery charging process according to this application is deployed, wherein the multi-input neural network model may include: - Input layer, the input layer includes multiple neurons, the data corresponding to the neurons of the input layer includes data on the electrochemical impedance spectroscopy of the lithium battery cell, wherein the data on the electrochemical impedance spectroscopy of the lithium battery cell includes the amplitude and phase angle of the electrochemical impedance spectroscopy of the lithium battery cell, or the real part and imaginary part of the electrochemical impedance spectroscopy of the lithium battery cell, or battery model parameters determined based on the electrochemical impedance spectroscopy of the lithium battery cell, the battery model parameters including, for example, battery internal resistance, and / or charge transfer resistance, and / or double layer capacitance and / or diffusion impedance; optionally, the data corresponding to the neurons of the input layer also includes various key frequency points and / or key frequency bands of the lithium battery cell related to the battery model parameters, and / or the state of charge of the lithium battery cell, and / or the state of health of the battery, and / or the charging rate and / or ambient temperature of the charging process of the lithium battery cell; - Hidden layers, which are used for feature extraction from the input data of the input layer; - Output layer, the output layer includes a neuron, and the data corresponding to the neuron of the output layer includes the reference voltage of the battery cell.

[0027] According to a fourth aspect of this application, a vehicle is provided, the vehicle including a battery management system according to this application.

[0028] According to a fifth aspect of this application, a computer program product, such as a computer-readable program carrier, is provided, comprising or storing computer program instructions that, when executed by a processor, at least assist in implementing the steps of the method described in this application. Attached Figure Description

[0029] The principles, features, and advantages of this application can be better understood by describing it in more detail below with reference to the accompanying drawings. The drawings show: Figure 1 A flowchart illustrating a training method for a multi-input neural network model for optimizing a lithium battery charging process, according to an exemplary embodiment of this application, is shown. Figure 2 A flowchart illustrating a training method for a multi-input neural network model for optimizing a lithium battery charging process, according to another exemplary embodiment of this application, is shown. Figure 3 A schematic structural diagram of a multi-input neural network model according to an exemplary embodiment of this application is shown. Figure 4 A flowchart illustrating a training method for a multi-input neural network model for optimizing a lithium battery charging process, according to another exemplary embodiment of this application, is shown. Figure 5 A flowchart illustrating a method for optimizing a lithium battery charging process using a multi-input neural network model, according to an exemplary embodiment of this application, is shown. Figure 6 A flowchart illustrating a method for optimizing a lithium battery charging process using a multi-input neural network model, according to another exemplary embodiment of this application, is shown; and Figure 7 A schematic structural diagram of a vehicle according to an exemplary embodiment of this application is shown. Detailed Implementation

[0030] To make the technical problems to be solved, the technical solutions, and the beneficial technical effects of this application clearer, the application will be further described in detail below with reference to the accompanying drawings and several exemplary embodiments. It should be understood that the specific embodiments described herein are only for explaining this application and are not intended to limit the scope of protection of this application.

[0031] Figure 1 A flowchart illustrating a training method for a multi-input neural network model for optimizing a lithium battery charging process, according to an exemplary embodiment of this application, is shown. The following exemplary embodiments describe the training method according to this application in more detail.

[0032] like Figure 1As shown, the training method may include steps S1 and S2. In step S1, during the charging test of a lithium battery test cell equipped with a reference electrode, a dataset of the reference voltage and electrochemical impedance spectroscopy (EIS) of the lithium battery test cell can be collected. During the charging process of lithium batteries, especially during fast charging at a high charging rate, lithium plating is easily caused when the negative electrode potential of the lithium battery is below 0V. However, voltage measuring devices can only measure the full cell voltage (i.e., the potential difference between the positive and negative electrodes) and cannot monitor the negative electrode potential in real time. Therefore, the lithium battery test cell can be modified before performing the charging test by introducing a reference electrode into the structure of the lithium battery test cell, thereby forming a three-electrode battery for monitoring the negative electrode potential.

[0033] The environmental factors, such as temperature, for charging tests of lithium-ion battery test cells can be set and kept constant according to the charging requirements of the test cells. The reference voltage and electrochemical impedance spectroscopy of the lithium-ion battery test cells are measured periodically in an electrochemical workstation. First, the reference electrode can be calibrated using a metallic lithium electrode as an absolute reference to eliminate potential drift and ensure accuracy. Then, in an anhydrous and oxygen-free environment, the potential difference between the negative electrode and the reference electrode—i.e., the reference voltage—is measured using high input impedance, for example. Since no current flows in the reference electrode, its potential remains stable, thus achieving accurate measurement of the negative electrode potential.

[0034] In an electrochemical workstation, AC impedance mode can be enabled, from, for example, 10... 5 Hz to 10 -2 Multiple sampling frequencies are selected at regular intervals within a frequency range of Hz. AC excitation is applied sequentially from high to low frequency, and the potential and current responses at each sampling frequency are collected. The electrochemical impedance of the lithium-ion battery test cell is then calculated, and Nyquist and Boden plots of the electrochemical impedance of the lithium-ion battery test cell are plotted as a function of frequency, thus obtaining the electrochemical impedance spectrum of the lithium-ion battery test cell. During the charging test of the lithium-ion battery test cell, the electrochemical impedance can be measured dynamically in situ, i.e., the next electrochemical impedance spectroscopy test is performed immediately after the first one, thereby collecting the electrochemical impedance spectrum dataset of the lithium-ion battery test cell in real time. Alternatively, the electrochemical impedance of the lithium-ion battery test cell can be measured based on the battery's state of charge, for example, measuring the electrochemical impedance of the lithium-ion battery test cell every 10% of the state of charge to collect the electrochemical impedance spectrum dataset of the lithium-ion battery test cell, thereby reducing the workload and data volume of the electrochemical impedance spectroscopy test.

[0035] The collected electrochemical impedance spectroscopy (EIS) data may include the amplitude and phase angle of the EIS of a lithium-ion battery test cell, or the real and imaginary parts of the EIS of a lithium-ion battery test cell, or battery model parameters determined based on the EIS of the lithium-ion battery test cell. Key frequency points and / or key frequency ranges related to the battery model parameters are marked in the collected EIS. The battery model parameters include, for example, one or more of the following: battery internal resistance, charge transfer resistance, double-layer capacitance, and diffusion impedance. The key frequency points and / or key frequency ranges are, for example, a certain number of frequency points and / or frequency intervals from 1 Hz to 1000 Hz. Since EIS measurement is very time-consuming, marking key frequency points and / or key frequency ranges in advance can effectively improve the efficiency of EIS testing and calculation.

[0036] Considering the differences in battery parameter changes during slow charging (e.g., 0.2C to 1C) and fast charging (e.g., 2C to 5C), datasets of reference voltage and electrochemical impedance spectroscopy of lithium battery test cells can be collected during charging tests at various charging rates (e.g., multiple charging rates selected between 0.2C and 5C), and the collected datasets can be labeled with the corresponding charging rates of the charging tests.

[0037] Since ambient temperature also affects the changes in battery parameters during the charging process, datasets of reference voltage and electrochemical impedance spectroscopy of lithium battery test cells can be collected during charging tests at various ambient temperatures, and the collected datasets can be labeled with the corresponding ambient temperature of the charging test.

[0038] Since the state of charge and / or state of health of the battery can also affect the changes in battery parameters during the charging process, data sets of reference voltage and electrochemical impedance spectroscopy of the lithium battery test cell, as well as the state of charge and / or state of health of the lithium battery test cell, can be collected during the charging test of the lithium battery test cell, and the corresponding state of charge and / or state of health can be labeled on the collected data sets.

[0039] like Figure 2The flowchart shown below illustrates a training method for a multi-input neural network model for optimizing the charging process of a lithium battery according to another exemplary embodiment of this application. The method may further include steps S11 and S12 for preprocessing the collected dataset. In step S11, reference voltage and electrochemical impedance spectra in the collected dataset may be matched such that the matched reference voltage and electrochemical impedance are obtained under the same test conditions of the lithium battery test cell. In step S12, outlier data in the collected dataset may be cleaned, and the cleaned dataset may be normalized. The cleaning methods may include, for example, noise reduction and / or outlier removal, to avoid interference from noisy data and / or outlier data in the training of the multi-input neural network model.

[0040] In step S2, the collected dataset is divided into a training set, a validation set, and a test set, for example, in a ratio of 7:2:1. The training set can be used to train the multi-input neural network model, and the validation set can be used to evaluate the performance parameters of the trained multi-input neural network model. The training process of the multi-input neural network model is completed when the mean square error between the reference voltage prediction value output by the trained multi-input neural network model based on the electrochemical impedance spectroscopy in the validation set and the reference voltage in the validation set is less than a preset threshold.

[0041] Multi-input neural network models can simultaneously receive input data from multiple sources, with multiple modalities or multiple structures, and independently extract features from these input data, then perform feature fusion and joint prediction, thereby improving the accuracy and robustness of the model's output results. Figure 3 A schematic structural diagram of a multi-input neural network model 2 according to an exemplary embodiment of this application is shown. The multi-input neural network model 2 may include: - Input layer 21, the input layer comprising multiple neurons ( Figure 3 The example shows three neurons. The data corresponding to the neurons in the input layer can include data about the electrochemical impedance spectroscopy of the lithium battery cell. The data about the electrochemical impedance spectroscopy of the lithium battery cell includes, for example, the amplitude and phase angle of the electrochemical impedance spectroscopy of the lithium battery cell, or the real and imaginary parts of the electrochemical impedance spectroscopy of the lithium battery cell, or battery model parameters determined based on the electrochemical impedance spectroscopy of the lithium battery cell. The battery model parameters include, for example, the battery internal resistance, and / or charge transfer resistance, and / or double-layer capacitance and / or diffusion impedance, etc. Optionally, the data corresponding to the neurons in the input layer can also include various key frequency points and / or key frequency bands of the lithium battery cell related to the battery model parameters, and / or the state of charge of the lithium battery cell, and / or the battery health state, and / or the charging rate and / or ambient temperature of the charging process of the lithium battery cell, etc. - Hidden layer 22, which is used to extract features from the input data of the input layer 21, wherein the number of neurons in the hidden layer can be set as a hyperparameter of the multi-input neural network model 2. Figure 3 (Two neurons are shown as an example). - Output layer 23, the output layer includes a neuron, and the output data of the output layer 23 is the reference voltage of the lithium battery cell.

[0042] Here, the multi-input neural network model can be trained using the training set, and its model parameters can be adjusted during training. Then, the performance parameters of the trained multi-input neural network model can be evaluated using the validation set. These performance parameters may include the mean square error (MSE) between the predicted reference voltage output by the trained multi-input neural network model based on the electrochemical impedance spectroscopy data in the validation set and the actual reference voltage value in the validation set; that is, the average of the sum of squares of the differences between the predicted and actual reference voltage values. The smaller the MSE, the higher the accuracy of the predicted reference voltage output by the multi-input neural network model. When the MSE is greater than or equal to a preset threshold, the multi-input neural network model can be retrained using the training set, and its model parameters can be further adjusted during training. The performance parameters of the trained multi-input neural network model are then re-evaluated using the validation set. When the MSE is less than the preset threshold, the trained multi-input neural network model is considered to have reached convergence, thus completing the training process of the multi-input neural network model.

[0043] To avoid overfitting of model parameters during training, the optimal model structure of the multi-input neural network model 2 can be determined through hyperparameter tuning during training. Specifically, hyperparameters of the multi-input neural network model 2 can be set, including, for example, learning rate, batch size, number of iterations, number of hidden layer neurons, and / or dropout rate. During the training of the multi-input neural network model 2, the hyperparameters of the multi-input neural network model 2 are adjusted based on the performance parameters of the multi-input neural network model 2 evaluated using the validation set, in order to determine the optimal model structure of the multi-input neural network model 2.

[0044] Furthermore, the model training process can be optimized using early stopping during training. Specifically, a patience value and a minimum improvement threshold can be set for the training process of the multi-input neural network model 2. Here, the "patience value" represents the number of consecutive rounds during which the model performance parameters evaluated based on the validation set are allowed to remain unchanged. Only when the improvement in the model performance parameters evaluated based on the validation set exceeds the set minimum improvement threshold is it determined that the model performance parameters have improved in the current round of training. The multi-input neural network model 2 is trained using the training set, and the performance parameters of the trained multi-input neural network model 2 are evaluated using the validation set. If the change in the performance parameters relative to the historical best performance parameters of the multi-input neural network model 2 is less than the minimum improvement threshold, one round of no improvement can be accumulated, and the model parameters of the multi-input neural network model 2 are not saved; if the change in the performance parameters relative to the historical best performance parameters of the multi-input neural network model 2 is greater than or equal to the minimum improvement threshold, the model parameters of the multi-input neural network model 2 can be saved. When the number of consecutively accumulated no-boost rounds is greater than or equal to the patience value, the training process of the multi-input neural network model 2 can be stopped. This allows the optimal number of training rounds to be determined automatically without changing the model structure or adding complex penalty terms, saving computational resources and effectively avoiding overfitting of model parameters during model training.

[0045] In the above embodiments of this application, a reference electrode is introduced into the structure of the lithium battery test cell to form a three-electrode battery for monitoring the negative electrode potential. The dataset of reference voltage and electrochemical impedance spectrum of the lithium battery test cell collected during the charging test is divided into a training set, a validation set, and a test set. The multi-input neural network model is trained using the training set, and the performance parameters of the trained multi-input neural network model are evaluated using the validation set, so that the performance parameters of the trained multi-input neural network model meet the requirements.

[0046] Figure 4 A flowchart illustrating a training method for a multi-input neural network model for optimizing a lithium battery charging process, according to another exemplary embodiment of this application, is shown. The following only describes the method in relation to... Figure 1 The differences between the embodiments shown are omitted for brevity, and the same steps will not be repeated.

[0047] like Figure 4As shown, the training method may further include step S3. In step S3, the performance parameters of the trained multi-input neural network model 2 can be evaluated using the test set. These performance parameters may include the mean square error and the coefficient of determination between the predicted reference voltage output by the trained multi-input neural network model 2 based on the electrochemical impedance spectroscopy data in the test set and the reference voltage in the test set. Here, the mean square error is the average of the sum of squares of the differences between the predicted reference voltage output by the multi-input neural network model 2 based on the test set and the true reference voltage. The coefficient of determination R... 2 It is a core metric used to evaluate simulation fit, representing the proportion of the dependent variable's variation that the model's independent variables can explain, i.e., the proportion of the reference voltage prediction that can be explained by the multi-input neural network model. In this way, the true generalization ability of the model can be independently and accurately evaluated after the model structure, hyperparameters, and / or training strategy of the multi-input neural network model 2 are fixed.

[0048] After the training process of the multi-input neural network model 2 is completed, the trained multi-input neural network model 2 can be deployed in the control unit 100 of the battery management system 10 to execute a method for optimizing the lithium battery charging process through the multi-input neural network model 2.

[0049] Figure 5 A flowchart illustrating a method for optimizing a lithium battery charging process using a multi-input neural network model, according to an exemplary embodiment of this application, is shown. Figure 5 As shown, the method may include steps S1' to S3'. In step S1', the electrochemical impedance spectroscopy (EIS) of the lithium battery cell can be acquired during the charging process of the lithium battery cell. In the context of this application, the lithium battery cell is a cell in a lithium battery pack actually used during vehicle operation, and no reference electrode is provided in the lithium battery cell. Similarly, during the charging process of the lithium battery cell, the EIS of the lithium battery cell can be measured dynamically in situ, that is, the next EIS test is performed immediately after the first EIS test, thereby acquiring a dataset of the EIS of the lithium battery cell in real time; alternatively, the EIS of the lithium battery cell can be measured based on the state of charge of the battery, for example, the EIS of the lithium battery cell can be measured every 10% of the state of charge to acquire a dataset of the EIS of the lithium battery cell, thereby reducing the workload of the EIS test and the amount of data collected.

[0050] In step S2', the reference voltage of the lithium battery cell can be predicted at least based on the acquired electrochemical impedance spectroscopy (EIS) using a trained multi-input neural network (MIN) model 2. The data corresponding to the neurons in the input layer of the MNN model 2 includes data about the EIS of the lithium battery cell. Therefore, the EIS data of the lithium battery cell can be input into the trained MNN model 2. This data may include, for example, the amplitude and phase angle of the EIS of the lithium battery cell, or the real and imaginary parts of the EIS of the lithium battery cell, or battery model parameters determined based on the EIS of the lithium battery cell. These battery model parameters may include one or more of the following: battery internal resistance, charge transfer resistance, double-layer capacitance, and diffusion impedance.

[0051] Optionally, the data corresponding to the neurons in the input layer of the multi-input neural network model 2 may also include key frequency points and / or key frequency bands of the lithium battery cell related to the battery model parameters. The key frequency points and / or key frequency bands related to the battery model parameters can be marked in the collected electrochemical impedance spectroscopy, and the reference voltage of the lithium battery cell can be predicted by the trained multi-input neural network model 2 based at least on the collected electrochemical impedance spectroscopy and the key frequency points and / or key frequency bands.

[0052] Optionally, the data corresponding to the neurons in the input layer of the multi-input neural network model 2 may also include the charging rate of the lithium battery cell during the charging process. The changes in battery parameters vary under different charging rates, and the multi-input neural network model 2 has been trained using a dataset of charging rates labeled for charging tests. Therefore, the ambient temperature during the charging process of the lithium battery cell can be collected, and the reference voltage of the lithium battery cell can be predicted by the trained multi-input neural network model 2 based at least on the ambient temperature during the charging process of the lithium battery cell and the collected electrochemical impedance spectrum, thereby minimizing the influence of the charging rate on the prediction of the reference voltage.

[0053] Optionally, the data corresponding to the neurons in the input layer of the multi-input neural network model 2 may also include the ambient temperature during the charging process of the lithium battery cell. The changes in battery parameters during the charging process vary to some extent under different ambient temperatures. Furthermore, the multi-input neural network model 2 has been trained using a dataset of labeled ambient temperatures during the charging process. Therefore, it can collect the ambient temperature during the charging process of the lithium battery cell and predict the reference voltage of the lithium battery cell based at least on the ambient temperature during the charging process of the lithium battery cell and the collected electrochemical impedance spectrum through the trained multi-input neural network model 2.

[0054] Optionally, the data corresponding to the neurons in the input layer of the multi-input neural network model 2 may also include the state of charge (SOC) and / or state of health of the lithium battery cell. The SOC and / or state of health of the battery will also have a certain impact on the changes in battery parameters during the charging process. The multi-input neural network model 2 has been trained using a dataset containing the SOC and / or state of health of the lithium battery cell. Therefore, the SOC and / or state of health of the lithium battery cell can be collected during the charging process of the lithium battery cell. Based on the collected electrochemical impedance spectroscopy and the SOC and / or state of health of the lithium battery cell during the charging process, the reference voltage of the lithium battery cell can be predicted by the trained multi-input neural network model 2.

[0055] In step S3', the charging boundary conditions of the lithium battery cell can be monitored based on the predicted reference voltage. During the charging process of a lithium battery, especially during fast charging at a high charging rate, lithium plating is easily caused when the negative electrode potential of the lithium battery is below 0V. The reference voltage is equal to the potential difference between the negative electrode and the reference electrode, and the reference voltage of the reference electrode is calibrated to 0V. Therefore, when the predicted reference voltage is less than or equal to zero, it can be determined that the lithium battery cell has reached the charging boundary condition for lithium plating.

[0056] Figure 6 A flowchart illustrating a method for optimizing a lithium battery charging process using a multi-input neural network model, according to another exemplary embodiment of this application, is shown. The following only describes the process in relation to... Figure 5 The differences between the embodiments shown are omitted for brevity, and the same steps will not be repeated.

[0057] like Figure 6 As shown, the method may include step S4'. In step S4', when the lithium battery cell is detected to have reached the charging boundary condition for lithium plating, a prompt message regarding lithium plating of the lithium battery cell can be sent, for example, by displaying the prompt message on the vehicle's in-vehicle display screen or by sending the prompt message to a relevant APP on the user's mobile terminal; alternatively, the charging process of the lithium battery cell can be terminated to spontaneously prevent the occurrence of lithium plating.

[0058] According to the above embodiments of this application, during the lithium battery charging process, the reference voltage of the lithium battery cell is predicted by a trained multi-input neural network model based on the collected electrochemical impedance spectrum. Based on the predicted reference voltage, the charging boundary conditions of the lithium battery cell are automatically monitored. If charging boundary conditions are detected, a prompt message is sent and / or the charging process is automatically terminated, thereby optimizing the lithium battery charging process, preventing lithium plating in the lithium battery cell during charging, and effectively extending the service life of the lithium battery.

[0059] In addition, it should be noted that the step numbers described herein do not necessarily represent the order of steps, but are merely a reference numeral. The order may be changed depending on the specific circumstances, as long as the technical objective of this application can be achieved.

[0060] Figure 7 A schematic structural diagram of a vehicle according to an exemplary embodiment of this application is shown. Figure 7 As shown, vehicle 1 is equipped with a battery management system 10, which includes a control unit 100. The control unit 100 is used to execute the method according to this application for optimizing the lithium battery charging process using a multi-input neural network model. The control unit 100 may include at least one processor and a memory, in which program instructions executable by the at least one processor are stored. When the program instructions are executed by the at least one processor, a multi-input neural network model 2 trained using the training method according to this application for optimizing the lithium battery charging process is deployed.

[0061] It should be understood that the terms “first,” “second,” “third,” etc., used in this document are for descriptive purposes only and should not be construed as indicating or implying relative importance, nor should they be construed as implicitly specifying the number of technical features indicated.

[0062] If an embodiment includes an "and / or" association between a first feature and a second feature, it should be interpreted as follows: according to one implementation, the embodiment has not only the first feature but also the second feature; according to another implementation, the embodiment has either only the first feature or only the second feature.

[0063] Although specific embodiments have been described above, these embodiments are not intended to limit the scope of this application, even when only a single embodiment is described with respect to a particular feature. The feature examples provided in this application are intended for illustrative purposes and not for limitation, unless otherwise stated. In practice, multiple features may be combined with each other as needed and where technically feasible. Various substitutions, modifications, and alterations are also conceived without departing from the spirit and scope of this application.

Claims

1. A training method for a multi-input neural network model (2) used to optimize the charging process of a lithium battery, the training method comprising: During the charging test of a lithium battery test cell equipped with a reference electrode, data sets of reference voltage and electrochemical impedance spectroscopy of the lithium battery test cell are collected. The collected dataset is divided into a training set, a validation set, and a test set. The multi-input neural network model (2) is trained using the training set, and the performance parameters of the trained multi-input neural network model (2) are evaluated using the validation set. The training process of the multi-input neural network model (2) is completed only when the mean square error between the reference voltage prediction value output by the trained multi-input neural network model (2) based on the electrochemical impedance spectrum in the validation set and the reference voltage in the validation set is less than a preset threshold.

2. The training method according to claim 1, wherein, The method further includes: The performance parameters of the trained multi-input neural network model (2) are evaluated using the test set, wherein the performance parameters include the mean square error and the coefficient of determination between the reference voltage prediction value output by the trained multi-input neural network model (2) based on the electrochemical impedance spectroscopy in the test set and the reference voltage in the test set.

3. The training method according to any one of the preceding claims, wherein, The collected electrochemical impedance spectroscopy data includes the amplitude and phase angle of the electrochemical impedance spectrum of the lithium battery test cell, or the real and imaginary parts of the electrochemical impedance spectrum of the lithium battery test cell, or battery model parameters determined based on the electrochemical impedance spectrum of the lithium battery test cell. Key frequency points and / or key frequency bands related to the battery model parameters are marked in the collected electrochemical impedance spectrum. The battery model parameters include, for example, one or more of the following parameters: battery internal resistance, charge transfer resistance, double layer capacitance, and diffusion impedance.

4. The training method according to any one of the preceding claims, wherein, During the charging test of the lithium battery test cell at various charging rates, a dataset of reference voltage and electrochemical impedance spectrum of the lithium battery test cell was collected, and the corresponding charging rate of the charging test was labeled for the collected dataset. and / or During the charging test of lithium battery test cells under various ambient temperatures, data sets of reference voltage and electrochemical impedance spectroscopy of the lithium battery test cells were collected, and the corresponding ambient temperature of the charging test was labeled for the collected data sets.

5. The training method according to any one of the preceding claims, wherein, During the charging test of the lithium battery test cell, datasets of the reference voltage and electrochemical impedance spectroscopy of the lithium battery test cell, as well as the state of charge and / or state of health of the lithium battery test cell, are collected, and the collected datasets are labeled with the corresponding state of charge and / or state of health; and / or During the charging test of the lithium battery test cell, the electrochemical impedance of the lithium battery test cell is measured based on the state of charge of the battery, or the electrochemical impedance of the lithium battery test cell is measured dynamically in situ.

6. The training method according to any one of the preceding claims, wherein, The method further includes: The reference voltage and electrochemical impedance spectroscopy in the collected dataset are matched. The abnormal data in the collected dataset is cleaned and the cleaned dataset is normalized. The cleaning methods include, for example, noise reduction and / or outlier removal.

7. The training method according to any one of the preceding claims, wherein, Set the hyperparameters of the multi-input neural network model (2), such hyperparameters including, for example, learning rate, and / or batch size, and / or number of iterations, and / or number of hidden layer neurons, and / or dropout rate; During the training of the multi-input neural network model (2), the hyperparameters of the multi-input neural network model (2) are adjusted based on the performance parameters of the multi-input neural network model (2) evaluated using the validation set.

8. The training method according to any one of the preceding claims, wherein, Set the patience value and minimum improvement threshold for the training process of the multi-input neural network model (2); The multi-input neural network model (2) is trained using the training set, and the performance parameters of the trained multi-input neural network model (2) are evaluated using the validation set. If the change in the performance parameter relative to the historical best performance parameter of the multi-input neural network model (2) is less than the minimum improvement threshold, the number of no-improvement rounds is accumulated, and the model parameters of the multi-input neural network model (2) are not saved. If the change in the performance parameter relative to the historical best performance parameter of the multi-input neural network model (2) is greater than or equal to the minimum improvement threshold, the model parameters of the multi-input neural network model (2) are saved; If the number of consecutively accumulated no-boost rounds is greater than or equal to the patience value, the training process of the multi-input neural network model (2) is stopped.

9. A method for optimizing a lithium battery charging process using a multi-input neural network model (2), said multi-input neural network model (2) being trained using the training method according to any one of the preceding claims, said method comprising: Electrochemical impedance spectroscopy of the lithium battery cell was collected during the charging process. The reference voltage of the lithium battery cell is predicted at least based on the collected electrochemical impedance spectroscopy by a trained multi-input neural network model (2); The charging boundary conditions of the lithium battery cell are monitored based on the predicted reference voltage.

10. The method according to claim 9, wherein, When the predicted reference voltage is less than or equal to zero, the lithium battery cell is monitored to have reached the charging boundary condition for lithium plating.

11. The method according to claim 9 or 10, wherein, The collected electrochemical impedance spectroscopy data includes the amplitude and phase angle of the electrochemical impedance spectrum of the lithium-ion battery cell, or the real and imaginary parts of the electrochemical impedance spectrum of the lithium-ion battery cell, or battery model parameters determined based on the electrochemical impedance spectrum of the lithium-ion battery cell. These battery model parameters include, for example, one or more of the following parameters: battery internal resistance, charge transfer resistance, double-layer capacitance, and diffusion impedance; and / or Mark key frequency points and / or key frequency bands related to battery model parameters in the collected electrochemical impedance spectroscopy, and predict the reference voltage of the lithium battery cell using a trained multi-input neural network model (2) based at least on the collected electrochemical impedance spectroscopy and the key frequency points and / or the key frequency bands; and / or Obtain the charging rate of the lithium battery cell during the charging process, and predict the reference voltage of the lithium battery cell by a trained multi-input neural network model (2) based at least on the charging rate of the lithium battery cell during the charging process and the collected electrochemical impedance spectrum; and / or The ambient temperature during the charging process of the lithium battery cell is collected, and the reference voltage of the lithium battery cell is predicted by a trained multi-input neural network model (2) based at least on the ambient temperature during the charging process of the lithium battery cell and the collected electrochemical impedance spectrum; and / or During the charging process of the lithium battery cell, the state of charge and / or state of health of the lithium battery cell are collected, and the reference voltage of the lithium battery cell is predicted by a trained multi-input neural network model (2) based on the collected electrochemical impedance spectrum and the state of charge and / or state of health of the lithium battery cell during the charging process.

12. The method according to any one of claims 9 to 11, wherein, The method further includes: If the lithium battery cell is detected to have reached the charging boundary condition for lithium plating, a warning message about lithium plating in the lithium battery cell is sent and / or the charging process of the lithium battery cell is terminated.

13. A battery management system (10) comprising a control unit (100) for performing the method according to any one of claims 9 to 12.

14. The battery management system (10) according to claim 13, wherein, The control unit (100) includes at least one processor and a memory, wherein the memory stores program instructions executable by the at least one processor, and when the program instructions are executed by the at least one processor, a multi-input neural network model (2) trained by the training method according to any one of claims 1 to 8 is deployed, wherein the multi-input neural network model (2) includes: The input layer (21) includes multiple neurons. The data corresponding to the neurons of the input layer includes data on the electrochemical impedance spectroscopy of the lithium battery cell. The data on the electrochemical impedance spectroscopy of the lithium battery cell includes the amplitude and phase angle of the electrochemical impedance spectroscopy of the lithium battery cell, or the real and imaginary parts of the electrochemical impedance spectroscopy of the lithium battery cell, or battery model parameters determined based on the electrochemical impedance spectroscopy of the lithium battery cell. The battery model parameters include, for example, the battery internal resistance, and / or charge transfer resistance, and / or double layer capacitance and / or diffusion impedance. Optionally, the data corresponding to the neurons of the input layer also includes various key frequency points and / or key frequency bands of the lithium battery cell related to the battery model parameters, and / or the state of charge of the lithium battery cell, and / or the battery health state, and / or the charging rate and / or ambient temperature of the charging process of the lithium battery cell. A hidden layer (22) is used to extract features from the input data of the input layer (21); The output layer (23) includes a neuron, and the data corresponding to the neuron of the output layer (23) includes the reference voltage of the battery cell.

15. A vehicle (1) comprising a battery management system (10) according to claim 13 or 14.

16. A computer program product, such as a computer-readable program carrier, comprising or storing computer program instructions that, when executed by a processor, at least auxiliaryly implement the steps of the method according to any one of claims 1 to 12.

Images