Model training method, and cell capacity prediction method and device

Abstract

The invention discloses a model training method and a cell capacity prediction method and device, and relates to the field of batteries. The model training method comprises the steps that a first feature set is acquired, the first feature set is used for representing features associated with cell capacity in production parameters of sample cells, and the sample cells comprise normal cells and abnormal cells; screening features of which the distinguishing degree of the normal battery cell and the abnormal battery cell exceeds a first preset threshold value from the first feature set, and determining a second feature set; balancing the number of sample battery cells of which the battery cell capacities are located in each preset capacity interval in the second feature set, and determining a third feature set; and screening features having contribution to model training from the third feature set, determining a fourth feature set, and training the cell capacity prediction model through the fourth feature set. According to the method, the data for model training can be optimized, and the data quality is improved, so that the trained model can accurately predict the cell capacity of various capacity intervals.

Description

Technical Field

[0001] The present application relates to the field of battery technology, and in particular to a model training method, a battery cell capacity prediction method and a device. Background Art

[0002] With the rapid development of new energy vehicles, the cost, safety, efficiency and intelligence of new energy vehicles are the top priorities of various companies, and these are closely related to the development of power batteries. The production process of power battery cells is very complicated, usually including many production processes such as stirring, coating, rolling, etc. Since the time and cost of cell production related technologies and processes are relatively high, some battery manufacturers have introduced various technologies to predict cell capacity in order to achieve the expectation of eliminating the capacity division process or reducing the capacity division ratio at the cell production site, thereby achieving the purpose of reducing costs and increasing efficiency.

[0003] Commonly used battery cell capacity prediction methods mainly rely on empirical models and simplified electrochemical equations. However, empirical models are mainly established using machine learning algorithms and big data technologies, and often lack accuracy and adaptability. They can usually only predict the battery cell capacity in a capacity range with a large number of sample categories, and cannot ensure the prediction accuracy of each type of capacity. Summary of the invention

[0004] The embodiments of the present application provide a model training method, a battery cell capacity prediction method and a device to optimize the data used for model training and improve data quality, so that the trained model can accurately predict the battery cell capacity of various capacity ranges.

[0005] In order to solve the above technical problems, the embodiments of the present application disclose the following technical solutions:

[0006] In a first aspect, a model training method is provided for training a cell capacity prediction model; the method comprises:

[0007] Acquire a first feature set, where the first feature set is used to characterize features associated with the cell capacity in production parameters of each sample cell, where each sample cell includes a normal cell and an abnormal cell;

[0008] Selecting features whose degree of differentiation between the normal battery cell and the abnormal battery cell exceeds a first preset threshold from the first feature set to determine a second feature set;

[0009] Balancing the number of sample cells whose cell capacities are in each preset capacity interval in the second feature set, and determining a third feature set;

[0010] Features that contribute to model training are screened from the third feature set to determine a fourth feature set, so as to train the battery cell capacity prediction model through the fourth feature set.

[0011] In combination with the first aspect, balancing the number of sample cells whose cell capacities are in each preset capacity interval in the second feature set to determine a third feature set includes:

[0012] dividing the second feature set into a plurality of feature subsets based on a plurality of preset capacity intervals;

[0013] The number of the sample cells in each of the feature subsets is adjusted so that the number of the sample cells in each of the feature subsets meets a preset condition.

[0014] In combination with the first aspect, the adjusting the number of the sample cells in each of the feature subsets includes:

[0015] Determine a first feature subset whose quantity exceeds a second preset threshold and a second feature subset whose quantity is less than a third preset threshold, wherein the second preset threshold is greater than or equal to the third preset threshold;

[0016] Downsampling the first feature subset;

[0017] And / or, upsampling the second feature subset.

[0018] In combination with the first aspect, selecting from the first feature set features whose degree of differentiation between the normal battery cell and the abnormal battery cell exceeds a first preset threshold value includes:

[0019] For each feature in the first feature set, determine a first probability distribution of the feature in the normal battery cell and a second probability distribution of the feature in the abnormal battery cell;

[0020] Determine, based on the first probability distribution and the second probability distribution, a KL divergence of the feature between the normal battery cell and the abnormal battery cell;

[0021] The feature whose KL divergence exceeds the first preset threshold is selected from the first feature set.

[0022] In combination with the first aspect, the first preset threshold is set by the following steps:

[0023] determining a plurality of candidate threshold values;

[0024] For any feature in the first feature set, based on the KL divergence of the feature between the normal battery cell and the abnormal battery cell, respectively select from the first feature set candidate feature sets whose KL divergence exceeds each candidate threshold;

[0025] Using each of the candidate feature sets respectively, training and evaluating the battery cell capacity prediction model based on a cross-validation and grid search method, to determine the performance of the battery cell capacity prediction model;

[0026] The alternative threshold corresponding to the battery cell capacity prediction model with the best performance is determined as the first preset threshold.

[0027] In combination with the first aspect, features that contribute to model training are selected from the third feature set to determine a fourth feature set, including:

[0028] Determine the importance of each feature in the third feature set and rank them;

[0029] Selecting the feature with the lowest importance from the third feature set as the current feature, and using each feature in the third feature set to train and evaluate the battery cell capacity prediction model;

[0030] If the model performance before removing the current feature is better than the model performance after removing the current feature, determining that the current feature has a contribution to the model training, and retaining the current feature in the third feature set;

[0031] If the model performance after removing the current feature is better than the model performance before removing the current feature, it is determined that the current feature does not contribute to the model training, and the current feature is removed from the third feature set. Continue to select the feature with the lowest importance from the remaining features of the third feature set in order from low to high importance as the current feature, and repeat the training and evaluation steps until the judgment of each feature in the third feature set is completed to obtain the fourth feature set.

[0032] In combination with the first aspect, before screening the features whose degree of differentiation between the normal battery cell and the abnormal battery cell exceeds a first preset threshold from the first feature set, the method further includes:

[0033] Data cleaning is performed on the first feature set to remove missing data and abnormal data in the first feature set.

[0034] In combination with the first aspect, the method further includes:

[0035] Using the fourth feature set, the battery cell capacity prediction model is trained and evaluated based on a cross-validation and grid search method to obtain optimal hyperparameters of the battery cell capacity prediction model;

[0036] Based on the optimal hyperparameters, the trained cell capacity prediction model is determined.

[0037] In combination with the first aspect, the battery cell capacity prediction model is a gradient boosting decision tree model.

[0038] In a second aspect, a model training device is provided for training a cell capacity prediction model; the device comprises:

[0039] A first acquisition module, used for acquiring a first feature set, where the first feature set is used for characterizing features associated with the cell capacity in production parameters of each sample cell, where each sample cell includes a normal cell and an abnormal cell;

[0040] A preprocessing module, configured to select features whose degree of differentiation between the normal battery cell and the abnormal battery cell exceeds a first preset threshold from the first feature set, and determine a second feature set;

[0041] An equivalent processing module, used for balancing the number of sample cells whose cell capacities are in each preset capacity interval in the second feature set, and determining a third feature set;

[0042] The incremental elimination module is used to select features that contribute to model training from the third feature set, determine a fourth feature set, and train the battery cell capacity prediction model through the fourth feature set.

[0043] In a third aspect, a method for predicting cell capacity is provided, the method comprising:

[0044] Obtain the features to be processed of the battery cell to be tested;

[0045] Based on the feature to be processed, a cell capacity prediction value of the cell to be tested is generated via a cell capacity prediction model, wherein the cell capacity prediction model is generated by training using the model training method as described in any one of the first aspects.

[0046] In a fourth aspect, a battery cell capacity prediction device is provided, the device comprising:

[0047] The second acquisition module is used to obtain the features to be processed of the battery cell to be tested;

[0048] A capacity prediction module is used to generate a predicted value of the cell capacity of the cell to be tested based on the feature to be processed via a cell capacity prediction model, wherein the cell capacity prediction model is generated by training using the model training method as described in any one of the first aspects.

[0049] One of the above technical solutions has the following advantages or beneficial effects:

[0050] Compared with the prior art, a model training method of the present application includes: obtaining a first feature set, the first feature set is used to characterize the features associated with the cell capacity in the production parameters of each sample cell, and each sample cell includes a normal cell and an abnormal cell; screening the features whose degree of distinction between normal cells and abnormal cells exceeds a first preset threshold from the first feature set, and determining a second feature set; balancing the number of sample cells whose cell capacities are located in each preset capacity interval in the second feature set, and determining a third feature set; screening the features that contribute to model training from the third feature set, and determining a fourth feature set, so as to train the cell capacity prediction model through the fourth feature set. The model training method provided in the present application can screen out features with a large degree of distinction between normal cells and abnormal cells, and balance the number of samples whose cell capacities are located in each capacity interval, and finally screen out features that contribute to model training for training, so as to optimize the data used for model training, improve data quality, and enable the trained model to accurately predict the cell capacity of various capacity intervals.

[0051] A battery cell capacity prediction method of the present application uses a trained battery cell capacity prediction model to predict the processing features to obtain a battery cell capacity prediction value of the battery cell to be tested, wherein the battery cell capacity prediction model is trained with high-quality data, thereby being able to accurately predict the battery cell capacity of various capacity ranges and improve prediction accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For those skilled in the art, other drawings can be obtained based on these drawings without creative work.

[0053] Figure 1 A schematic diagram of the overall process of the model training method provided in the embodiment of the present application;

[0054] Figure 2 A schematic diagram of a specific process of the model training method provided in an embodiment of the present application;

[0055] Figure 3 A schematic diagram of the overall process of the battery cell capacity prediction method provided in an embodiment of the present application;

[0056] Figure 4 A schematic diagram of the structure of a model training device provided in an embodiment of the present application;

[0057] Figure 5 A schematic diagram of the structure of a battery cell capacity prediction device provided in an embodiment of the present application;

[0058] Figure 6 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application;

[0059] Reference numerals:

[0060] 401 - first acquisition module; 402 - pre-processing module; 403 - equivalent processing module; 404 - increment elimination module; 501 - second acquisition module; 502 - capacity prediction module; 801 - memory; 802 - processor. DETAILED DESCRIPTION

[0061] The technical solutions in the embodiments of the present application will be described clearly and completely below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all of the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present application.

[0062] In the description of the present application, it should be understood that the terms "upper", "lower", "front", "back", "left", "right", "top", "bottom", "inside", "outside" and the like indicate positions or positional relationships based on the positions or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present application and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation, and therefore cannot be understood as a limitation on the present application. In addition, the terms "first" and "second" are used for descriptive purposes only, and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Thus, the features defined as "first" and "second" may explicitly or implicitly include one or more features. In the description of the present application, "multiple" means two or more, and at least one means one, two or more, unless otherwise clearly and specifically defined.

[0063] With the rapid development of new energy vehicles, more and more car companies are focusing on the development of this field under the broad market prospects. Among them, the cost, safety, efficiency and intelligence of new energy vehicles are the top priorities of each company's research, and these are closely related to the development of new energy power batteries. However, the production process of power battery cells is very complicated, including many production processes such as stirring, coating, and rolling. The current cell capacity is strongly related to the formation and capacity separation processes, that is, the cell is activated after the formation process and the capacity is tested in the capacity separation process.

[0064] However, the time and cost of battery cell production-related technologies and processes are relatively high. Therefore, most battery manufacturers have introduced various technologies to predict battery cell capacity, hoping to achieve the goal of eliminating the capacity division process or reducing the capacity division ratio at the battery cell production site, thereby achieving the goal of reducing costs and increasing efficiency. However, the commonly used battery cell capacity prediction methods mainly rely on empirical models and simplified electrochemical equations. The empirical models are mainly established using machine learning algorithms and big data technologies, which often lack accuracy and adaptability and cannot effectively process large-scale and high-dimensional battery cell data and battery cell production data with a large number of nonlinear characteristics; secondly, it is impossible to fully utilize the rich information contained in the historical production data of the battery cell to establish an accurate prediction model; in addition, the battery cell capacity prediction model established using machine learning algorithms and big data technologies cannot guarantee the prediction accuracy of each type of capacity.

[0065] In view of this, an embodiment of the present application provides a model training method, which optimizes the data used for model training, improves data quality, and then constructs an accurate prediction model, so that the trained model can accurately predict the capacity of battery cells in various capacity ranges, thereby improving the accuracy and reliability of the prediction.

[0066] See also Figure 1 , Figure 1 The overall flow diagram of the model training method provided in the embodiment of the present application. Before predicting the cell capacity of the cell to be tested by the cell capacity prediction model, the cell capacity prediction model can be trained by the following steps:

[0067] Step 101: Obtain a first feature set, where the first feature set is used to characterize features associated with cell capacity in production parameters of each sample cell, where each sample cell includes a normal cell and an abnormal cell.

[0068] Specifically, a normal cell refers to a sample cell whose cell capacity is within a normal range, and an abnormal cell refers to a cell whose cell capacity is low, high or abnormal.

[0069] In some examples, the process production parameter characteristics and their derived parameter characteristics of the battery cell can be collected, where the process production parameter characteristics refer to various parameters used to describe and control the production process in the industrial production process, which may include the result characteristics and process data of each process that have a certain mechanism relationship with the capacity, such as weight before injection, baking temperature, baking time, baking moisture, formation pressure, formation voltage, formation storage location temperature, etc. Derived parameter characteristics are parameters calculated or derived from the process production parameter characteristics. These characteristics have a mechanism relationship with the battery cell capacity from various dimensions and can be used to evaluate the battery cell capacity status.

[0070] Specifically, a complete battery cell production data includes basic information of the battery cell barcode, production result characteristics, curve process parameter characteristics and actual capacity value, etc. Parameters related to capacity can be obtained from engineers or material mechanism groups. These parameters mainly include process parameters of the pre-process, mid-process and post-process of battery cell production. Among them, the process parameters of the pre-process, mid-process and post-process of battery cell production can include more than 40 result parameter characteristics and curve process parameter characteristics such as coating surface density of the pre-process, the positive electrode stage length and the negative electrode stage length of the mid-process, the stage water content of the post-process, and the battery cell temperature. It can be understood that all parameter characteristics form the first feature set.

[0071] See also Figure 2 , Figure 2 A specific flow chart of the model training method provided in the embodiment of the present application. In some embodiments, after executing step 101 and before executing step 102, the model training method in the embodiment of the present application may also include:

[0072] Step 105: Perform data cleaning on the first feature set to remove missing data and abnormal data in the first feature set.

[0073] Specifically, because the data used to train the model must contain data of multiple categories, it is necessary to delete data with obvious detection errors, that is, to delete data with negative actual capacity of the battery cell, etc. In this way, each parameter feature in the first feature set is preprocessed, and the battery cells with missing data and outliers are processed through data cleaning to clean up the dirty and messy data to ensure the quality and applicability of the data for machine learning training.

[0074] Step 102: Screening features whose degree of differentiation between normal cells and abnormal cells exceeds a first preset threshold from the first feature set to determine a second feature set.

[0075] In some embodiments, the following steps may be performed to filter out features whose degree of differentiation between normal cells and abnormal cells exceeds a first preset threshold from the first feature set:

[0076] Step 1: for each feature in the first feature set, determine a first probability distribution of the feature in a normal cell and a second probability distribution of the feature in an abnormal cell.

[0077] Specifically, statistical analysis can be performed on each feature of the normal battery cell and each feature of the abnormal battery cell, and corresponding probability distribution models can be selected to fit each feature of the normal battery cell and each feature of the abnormal battery cell, respectively, and a first probability distribution P of the feature in the normal battery cell and a second probability distribution Q of the feature in the abnormal battery cell can be obtained, respectively.

[0078] Step 2: Based on the first probability distribution and the second probability distribution, determine the KL divergence of the feature between the normal battery cell and the abnormal battery cell.

[0079] Specifically, KL divergence (Kullback-Leibler divergence) can compare the distribution of characteristic parameters of normal cell capacity and abnormal cell capacity to screen out parameter features with greater distribution differentiation between normal cells and abnormal cells. Specifically, the KL divergence is used to compare the difference between the probability density functions of any feature in normal cells and abnormal cells. A feature with a larger KL divergence value indicates that there is a large difference between the distribution of the characteristic parameter in normal cells and abnormal cells. This feature may be one of the causal features of the abnormal cell, so it is retained in the second feature set, while features with smaller KL divergence values ​​are regarded as irrelevant features and are removed.

[0080] Exemplarily, the KL divergence of the first probability distribution P and the second probability distribution Q is determined by the following formula (1):

[0081]

[0082] In formula (1), D KL (P∥Q) is the KL divergence, n is the size of the sample space, P(i) is the probability of the i-th event in the sample space in the first probability distribution, and Q(i) is the probability of the i-th event in the sample space in the second probability distribution.

[0083] Accordingly, the KL divergence of the i-th feature between normal cells and abnormal cells is determined by the following formula (2):

[0084]

[0085] In formula (2), D KL (P(i)∥Q(i)) is the KL divergence of the ith feature between normal cells and abnormal cells, P(i) is the probability of the ith feature in the sample space in the first probability distribution, and Q(i) is the probability of the ith feature in the sample space in the second probability distribution.

[0086] Step three: filter out features whose KL divergence exceeds a first preset threshold from the first feature set.

[0087] Specifically, the first preset threshold is a hyperparameter of the model, and the optimal value needs to be obtained through subsequent training.

[0088] Exemplarily, the first preset threshold may be set by the following steps:

[0089] The first step is to determine multiple candidate thresholds.

[0090] Exemplarily, the multiple candidate thresholds may be multiple values ​​selected from 0.1 to 1.0, such as 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, and 1.0.

[0091] In the second step, for any feature in the first feature set, based on the KL divergence of the feature between normal cells and abnormal cells, the candidate feature sets whose KL divergence exceeds each candidate threshold are respectively screened out from the first feature set.

[0092] That is to say, each candidate threshold can filter out a candidate feature set.

[0093] In the third step, each alternative feature set is used to train and evaluate the battery cell capacity prediction model based on cross-validation and grid search methods to determine the performance of the battery cell capacity prediction model.

[0094] In the fourth step, the alternative threshold corresponding to the battery cell capacity prediction model with the best performance is determined as the first preset threshold.

[0095] Through the above scheme, a better first preset threshold can be screened out, thereby screening out a second feature set that is more beneficial to training, thereby improving the quality of training data and facilitating improving the accuracy of subsequent training.

[0096] It is understandable that the model training and evaluation methods may also adopt other methods besides cross-validation and grid search, and the embodiments of the present application do not specifically limit this.

[0097] Through the above scheme, the KL divergence method is a regression task for quantitatively predicting capacity values, which can better handle the various nonlinear changes and parameter feature interaction processes involved in the capacity activation process of the lithium iron phosphate system, so that it can more accurately screen the parameter features with large distribution differentiation between normal cells and abnormal cells, that is, screen out the features most relevant to capacity and filter out the features without differentiation and irrelevant features, thereby improving the prediction accuracy of the model and reducing training time and memory usage.

[0098] In other embodiments, other methods may be used to obtain features with greater discrimination, such as chi-square test, L1 regularization, information gain, correlation coefficient, etc., which are not specifically limited in the embodiments of the present application.

[0099] Step 103: Balancing the number of sample cells whose cell capacities are in each preset capacity interval in the second feature set to determine a third feature set.

[0100] Specifically, in the second feature set, samples with cell capacities within the normal range may usually account for 80% of the total samples, while cells with lower or higher capacities and cells with abnormal capacities account for relatively fewer samples, and the training sample data set is relatively unbalanced.

[0101] In some embodiments, step 103 may be implemented by the following steps:

[0102] Step 1: divide the second feature set into multiple feature subsets based on multiple preset capacity intervals.

[0103] Specifically, the capacity intervals can be divided according to the specifications of the battery cells. For example, the sample batteries in the second feature set can be divided into three intervals of low-capacity batteries, normal batteries and high-capacity batteries according to the battery cell capacity. The features corresponding to the sample batteries in each interval constitute the corresponding feature subset.

[0104] Step 2: Adjust the number of sample cells in each feature subset so that the number of sample cells in each feature subset meets a preset condition.

[0105] In some examples, the number of sample cells in each feature subset satisfying the preset condition may include that the difference between the number of sample cells in each feature subset is less than a preset threshold. That is, the number of sample cells in each feature subset may be substantially the same.

[0106] In some examples, step 2 can be implemented as follows:

[0107] In the first step, a first feature subset whose quantity exceeds a second preset threshold and a second feature subset whose quantity is less than a third preset threshold are determined, and the second preset threshold is greater than or equal to the third preset threshold.

[0108] Specifically, a feature subset with a larger number of samples and a feature subset with a smaller number of samples are determined from each feature subset.

[0109] In the second step, down-sample the first feature subset and / or up-sample the second feature subset.

[0110] Specifically, downsampling can balance the data set by reducing the number of samples in the majority class to make it close to the number of samples in the minority class. Methods such as random downsampling, clustered downsampling, or feature-based downsampling can be used. Upsampling can balance the data set by increasing the number of samples in the minority class to make it close to the number of samples in the majority class. Methods such as repeated replication, interpolation between samples, and adding noise can be used.

[0111] It is understandable that, according to actual conditions, you can choose to downsample only the first feature subset, or upsample only the second feature subset, or downsample the first feature subset and upsample the second feature subset at the same time, and the embodiments of the present application are not limited to this.

[0112] Through the above scheme, the number of samples of majority cell capacity and minority cell capacity can be balanced, the impact of sample imbalance on modeling can be reduced, and it can be ensured that each category of capacity data is fairly considered during the training process of the machine learning model. The importance of cell capacity with a small number of samples in machine learning algorithm modeling is increased to avoid the model being overly sensitive to the majority category and ignoring the minority category, so as to balance the performance and generalization ability of the model, avoid the model's prediction results being seriously biased towards cell capacity data with a large sample size, and pay less attention to the training of cell capacity with a small sample size of lower capacity, higher capacity and abnormal measurement, resulting in a decrease in the prediction accuracy of the prediction model and an inability to effectively predict the cell capacity.

[0113] Step 104: Screen features that contribute to model training from the third feature set, determine a fourth feature set, and train the cell capacity prediction model through the fourth feature set.

[0114] Specifically, after obtaining the third feature set, if the process production parameter feature subset is not yet optimal, the third feature set can be combined with recursive feature elimination to obtain the optimal feature subset as much as possible.

[0115] In some embodiments, step 104 may be implemented by:

[0116] Step 1: Determine the importance of each feature in the third feature set and rank them.

[0117] In some examples, there are multiple ways to evaluate feature importance. For example, each feature in the third feature set can be traversed in the extreme boosting decision tree, and methods such as calculating the correlation between the feature and the target variable or variance analysis can be used. The embodiments of the present application do not specifically limit this.

[0118] Step 2: Select the feature with the lowest importance from the third feature set as the current feature, and use the features in the third feature set to train and evaluate the cell capacity prediction model.

[0119] Exemplarily, the model training and evaluation method can be based on the method of cross-validation and grid search, that is, the cross-validation grid search method. Specifically, first, a candidate set of hyperparameters is defined, which are parameters that need to be manually set during the model training process. Then, all possible hyperparameter combinations are traversed by exhaustive search. For each hyperparameter combination, a model training and evaluation will be performed. In order to accurately evaluate the performance of each hyperparameter combination, cross-validation is used to divide the training set and the validation set. A common cross-validation method is K-fold cross-validation, which divides the training set into K subsets, selects one of the subsets as the validation set each time, and the remaining subsets as the training set. Repeat K times, select a different validation set each time, and finally obtain the average value of the K model performance evaluation results. For each hyperparameter combination, the training set and validation set obtained by cross-validation are used for model training, and the model performance is evaluated on the validation set. Usually, some evaluation indicators (such as accuracy, precision, recall, etc.) are used to measure the performance of the model. According to the results of cross-validation, the hyperparameter combination with the best performance is selected as the final model configuration.

[0120] Step three: if the model performance before removing the current feature is better than the model performance after removing the current feature, it is determined that the current feature has a contribution to the model training, and the current feature is retained in the third feature set.

[0121] In some examples, the ways to evaluate the contribution of the current feature to model training include Mean Absolute Error (MAE) and R 2 Coefficient. If the current feature is used, the R 2 The coefficient increases, while MAE decreases, indicating that the current feature has a contribution in the model learning process. 2 The coefficient decreases, while MAE increases, indicating that the current feature does not contribute to the model learning process. 2 The coefficient can be calculated by the relevant formula, which will not be described here.

[0122] Step 4: If the model performance after removing the current feature is better than the model performance before removing the current feature, it is determined that the current feature does not contribute to the model training, and the current feature is removed from the third feature set. Continue to select the feature with the lowest importance from the remaining features of the third feature set in order from low to high importance as the current feature, and repeat the training and evaluation steps until the judgment of each feature in the third feature set is completed to obtain the fourth feature set.

[0123] It can be understood that each feature in the fourth feature set is a feature that contributes to model training.

[0124] Through the above solution, the optimal fourth feature set can be obtained, which can greatly optimize the data used for model training and improve data quality, so that the trained model can accurately predict the capacity of battery cells in various capacity ranges.

[0125] In some embodiments, after acquiring the fourth feature set, the method of the embodiment of the present application may further include the following steps:

[0126] Step 106: Using the fourth feature set, the battery cell capacity prediction model is trained and evaluated based on the cross-validation and grid search methods to obtain the optimal hyperparameters of the battery cell capacity prediction model.

[0127] Exemplarily, the cell capacity prediction model is a gradient boosting decision tree model. Specifically, the machine learning algorithm models mainly include four types of models: extreme gradient boosting decision tree, random forest, vector machine, and LightGBM (Light Gradient Boosting Machine). The four types of models are trained and evaluated, and the mean square absolute error (MAE), R 2 Coefficients and model training time, choose the smallest MAE, R 2 The machine learning algorithm with the largest coefficient and the shortest training time can obtain the ultimate gradient boosting decision tree.

[0128] Specifically, the method based on cross-validation and grid search can be found in the above-mentioned related content, that is, to screen hyperparameters under multiple candidate hyperparameter values, and verify them on the validation set to obtain the hyperparameter value with the best model evaluation performance. During the model training process, appropriate hyperparameters and regularization terms can be set, and the model can be tuned through cross-validation and grid search techniques to achieve better performance and generalization ability. First, based on the sample size of the battery cell and the parameter dimension, the appropriate hyperparameter range is estimated, such as the number of trees used in the model. The initial estimated range is [800, 1200, 1800, 2200], and the optimal value in the initial range is obtained through cross-validation. Next, based on the optimal value obtained above, the range is expanded. If the optimal initial value of the model is 1200, the range is set to [1000, 1200, 1400, 1600] at this time, and the above steps are iterated until the optimal hyperparameter value of the model is selected.

[0129] Step 107: Determine the trained cell capacity prediction model based on the optimal hyperparameters.

[0130] It can be understood that the model training method of the embodiment of the present application can screen out features that have a greater degree of distinction between normal battery cells and abnormal battery cells, and balance the number of samples whose battery cell capacities are in each capacity interval. Finally, the features that contribute to the model training are screened for training, thereby optimizing the data used for model training and improving data quality, so that the trained model can achieve accurate prediction of battery cell capacities in various capacity intervals.

[0131] Accordingly, see Figure 3 , Figure 3 The overall flow diagram of the cell capacity prediction method provided in the embodiment of the present application is as follows. After the cell capacity prediction model is obtained through training in the above embodiment, the cell capacity prediction can be performed through the following steps:

[0132] Step 301: Obtain the features to be processed of the battery cell to be tested.

[0133] It is understandable that the feature to be processed includes at least part of the process production parameter feature of the cell to be tested and its derived parameter feature. In other words, at least part of the process production parameter feature of the cell to be tested and its derived parameter feature can be collected as the feature to be processed.

[0134] Step 302: Based on the feature to be processed, a cell capacity prediction value of the cell to be tested is generated via a cell capacity prediction model.

[0135] It should be noted that, for related features not described in the embodiments of the present application, reference can be made to the related introductions of the aforementioned embodiments, and the embodiments can refer to each other.

[0136] It can be understood that the battery cell capacity prediction method provided in the embodiment of the present application uses a trained battery cell capacity prediction model to predict the features to be processed, and obtains the battery cell capacity prediction value of the battery cell to be tested, wherein the battery cell capacity prediction model is trained with high-quality data, and thus can achieve accurate prediction of the battery cell capacity in various capacity ranges, with high prediction accuracy.

[0137] Accordingly, see Figure 4 , Figure 4 The schematic diagram of the structure of the model training device provided in the embodiment of the present application is as follows: The model training device provided in the embodiment of the present application comprises a first acquisition module 401 , a preprocessing module 402 , an equivalent processing module 403 , and an increment elimination module 404 .

[0138] The first acquisition module 401 is used to obtain a first feature set, where the first feature set is used to characterize features associated with the cell capacity in production parameters of each sample cell, where each sample cell includes a normal cell and an abnormal cell.

[0139] The pre-processing module 402 is used to select features whose degree of differentiation between normal cells and abnormal cells exceeds a first preset threshold from the first feature set to determine a second feature set.

[0140] The equivalent processing module 403 is used to balance the number of sample cells whose cell capacities are in each preset capacity interval in the second feature set to determine a third feature set.

[0141] The incremental elimination module 404 is used to select features that contribute to model training from the third feature set, determine a fourth feature set, and train the cell capacity prediction model through the fourth feature set.

[0142] In some embodiments, the equivalent processing module 403 is specifically used to:

[0143] The second feature set is divided into a plurality of feature subsets based on a plurality of preset capacity intervals.

[0144] The number of sample cells in each feature subset is adjusted so that the number of sample cells in each feature subset meets a preset condition.

[0145] In some embodiments, the equivalent processing module 403 is specifically used to:

[0146] A first feature subset whose quantity exceeds a second preset threshold and a second feature subset whose quantity is less than a third preset threshold are determined, and the second preset threshold is greater than or equal to the third preset threshold.

[0147] The first feature subset is downsampled.

[0148] And / or, upsampling the second feature subset.

[0149] In some embodiments, the pre-processing module 402 is specifically used to:

[0150] For each feature in the first feature set, a first probability distribution of the feature in normal cells and a second probability distribution of the feature in abnormal cells are determined.

[0151] Based on the first probability distribution and the second probability distribution, a KL divergence of the feature between the normal battery cell and the abnormal battery cell is determined.

[0152] Features whose KL divergence exceeds a first preset threshold are selected from the first feature set.

[0153] In some embodiments, the pre-processing module 402 is specifically configured to set the first preset threshold through the following steps:

[0154] A number of candidate thresholds are determined.

[0155] For any feature in the first feature set, based on the KL divergence of the feature between normal cells and abnormal cells, candidate feature sets whose KL divergence exceeds each candidate threshold are respectively screened out from the first feature set.

[0156] Each candidate feature set is used to train and evaluate the battery cell capacity prediction model based on cross-validation and grid search methods to determine the performance of the battery cell capacity prediction model.

[0157] The alternative threshold corresponding to the battery cell capacity prediction model with the best performance is determined as the first preset threshold.

[0158] In some embodiments, the increment elimination module 404 is specifically used to:

[0159] Determine the importance of each feature in the third feature set and rank them.

[0160] The feature with the lowest importance is selected from the third feature set as the current feature, and the battery cell capacity prediction model is trained and evaluated using each feature in the third feature set.

[0161] If the model performance before removing the current feature is better than the model performance after removing the current feature, it is determined that the current feature has a contribution to the model training, and the current feature is retained in the third feature set.

[0162] If the model performance after removing the current feature is better than the model performance before removing the current feature, it is determined that the current feature does not contribute to the model training, and the current feature is removed from the third feature set. Continue to select the feature with the lowest importance from the remaining features of the third feature set in order from low to high importance as the current feature, and repeat the training and evaluation steps until the judgment of each feature in the third feature set is completed to obtain the fourth feature set.

[0163] In some embodiments, before screening out features whose degree of differentiation between normal cells and abnormal cells exceeds a first preset threshold from the first feature set, the pre-processing module 402 is further configured to:

[0164] Data cleaning is performed on the first feature set to remove missing data and abnormal data in the first feature set.

[0165] In some embodiments, the apparatus further comprises a training module;

[0166] The training module is used to train and evaluate the cell capacity prediction model based on the fourth feature set and the cross-validation and grid search method to obtain the optimal hyperparameters of the cell capacity prediction model.

[0167] Based on the optimal hyperparameters, the trained cell capacity prediction model is determined.

[0168] In some embodiments, the cell capacity prediction model is a gradient boosting decision tree model.

[0169] It can be understood that the model training device of the embodiment of the present application can screen out features that have a greater degree of distinction between normal battery cells and abnormal battery cells, and balance the number of samples with battery cell capacities in each capacity interval, and finally screen out features that contribute to model training for training, thereby optimizing the data used for model training and improving data quality, so that the trained model can achieve accurate prediction of battery cell capacities in various capacity intervals.

[0170] Accordingly, see Figure 5 , Figure 5 The schematic diagram of the structure of the battery cell capacity prediction device provided in the embodiment of the present application is as follows: The battery cell capacity prediction device provided in the embodiment of the present application comprises a second acquisition module 501 and a capacity prediction module 502 .

[0171] The second acquisition module 501 is used to obtain the to-be-processed features of the battery cell to be tested.

[0172] The capacity prediction module 502 is used to generate a predicted value of the cell capacity of the cell to be tested based on the feature to be processed via the cell capacity prediction model of the embodiment of the present application.

[0173] It can be understood that the battery cell capacity prediction device of the embodiment of the present application can use the trained battery cell capacity prediction model to predict the processing features and obtain the battery cell capacity prediction value of the battery cell to be tested, wherein the battery cell capacity prediction model is trained with high-quality data, and thus can achieve accurate prediction of battery cell capacities in various capacity ranges, thereby improving prediction accuracy.

[0174] Accordingly, see Figure 6 , Figure 6 A schematic diagram of the structure of an electronic device provided in an embodiment of the present application. An embodiment of the present application also provides an electronic device, including a memory 801 and a processor 802, wherein the memory 801 stores a computer program executable by the processor 802, and when the computer program is executed by the processor 802, the steps of the model training method in the aforementioned embodiment are executed, or the steps of the battery cell capacity prediction method in the aforementioned embodiment are executed.

[0175] Correspondingly, an embodiment of the present application also provides a computer-readable storage medium, which stores computer program instructions. When the computer program instructions are executed by a processor, the steps of the model training method in the aforementioned embodiment are executed, or the steps of the battery cell capacity prediction method in the aforementioned embodiment are executed.

[0176] The above is a detailed introduction to a model training method, a battery cell capacity prediction method and a device provided in the embodiments of the present application. Specific examples are used in this article to illustrate the principles and implementation methods of the present application. The description of the above embodiments is only used to help understand the technical solution and its core idea of ​​the present application. Ordinary technicians in this field should understand that they can still modify the technical solutions recorded in the aforementioned embodiments, or replace some of the technical features therein with equivalents; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A model training method, characterized in that: Used to train a cell capacity prediction model; the method comprises: Acquire a first feature set, where the first feature set is used to characterize features associated with the cell capacity in production parameters of each sample cell, where each sample cell includes a normal cell and an abnormal cell; Selecting features whose degree of differentiation between the normal battery cell and the abnormal battery cell exceeds a first preset threshold from the first feature set to determine a second feature set; Balancing the number of sample cells whose cell capacities are in each preset capacity interval in the second feature set, and determining a third feature set; Features that contribute to model training are screened from the third feature set to determine a fourth feature set, so as to train the battery cell capacity prediction model through the fourth feature set.

2. The model training method according to claim 1, characterized in that: Balancing the number of sample cells whose cell capacities are in each preset capacity interval in the second feature set to determine a third feature set includes: dividing the second feature set into a plurality of feature subsets based on a plurality of preset capacity intervals; The number of the sample cells in each of the feature subsets is adjusted so that the number of the sample cells in each of the feature subsets meets a preset condition.

3. The model training method according to claim 2, characterized in that: The adjusting the number of the sample cells in each of the feature subsets comprises: Determine a first feature subset whose quantity exceeds a second preset threshold and a second feature subset whose quantity is less than a third preset threshold, wherein the second preset threshold is greater than or equal to the third preset threshold; Downsampling the first feature subset; And / or, upsampling the second feature subset.

4. The model training method according to claim 1, characterized in that: Selecting a feature whose degree of differentiation between the normal battery cell and the abnormal battery cell exceeds a first preset threshold from the first feature set includes: For each feature in the first feature set, determine a first probability distribution of the feature in the normal battery cell and a second probability distribution of the feature in the abnormal battery cell; Determine, based on the first probability distribution and the second probability distribution, a KL divergence of the feature between the normal battery cell and the abnormal battery cell; The feature whose KL divergence exceeds the first preset threshold is selected from the first feature set.

5. The model training method according to claim 4, characterized in that: The first preset threshold is set by the following steps: determining a plurality of candidate threshold values; For any feature in the first feature set, based on the KL divergence of the feature between the normal battery cell and the abnormal battery cell, respectively select from the first feature set candidate feature sets whose KL divergence exceeds each candidate threshold; Using each of the candidate feature sets respectively, training and evaluating the battery cell capacity prediction model based on a cross-validation and grid search method, to determine the performance of the battery cell capacity prediction model; The alternative threshold corresponding to the battery cell capacity prediction model with the best performance is determined as the first preset threshold.

6. The model training method according to claim 1, characterized in that: Selecting features that contribute to model training from the third feature set to determine a fourth feature set includes: Determine the importance of each feature in the third feature set and rank them; Selecting the feature with the lowest importance from the third feature set as the current feature, and using each feature in the third feature set to train and evaluate the battery cell capacity prediction model; If the model performance before removing the current feature is better than the model performance after removing the current feature, determining that the current feature has a contribution to the model training, and retaining the current feature in the third feature set; If the model performance after removing the current feature is better than the model performance before removing the current feature, it is determined that the current feature does not contribute to the model training, and the current feature is removed from the third feature set. Continue to select the feature with the lowest importance from the remaining features of the third feature set in order from low to high importance as the current feature, and repeat the training and evaluation steps until the judgment of each feature in the third feature set is completed to obtain the fourth feature set.

7. The model training method according to any one of claims 1 to 6, characterized in that: Before screening the features whose degree of differentiation between the normal battery cell and the abnormal battery cell exceeds a first preset threshold from the first feature set, the method further includes: Data cleaning is performed on the first feature set to remove missing data and abnormal data in the first feature set.

8. The model training method according to any one of claims 1 to 6, characterized in that: The method further comprises: Using the fourth feature set, the battery cell capacity prediction model is trained and evaluated based on a cross-validation and grid search method to obtain optimal hyperparameters of the battery cell capacity prediction model; Based on the optimal hyperparameters, the trained cell capacity prediction model is determined.

9. The model training method according to claim 8, characterized in that: The battery cell capacity prediction model is a gradient boosting decision tree model.

10. A model training device, characterized in that: Used to train a cell capacity prediction model; the device comprises: A first acquisition module, used for acquiring a first feature set, where the first feature set is used for characterizing features associated with the cell capacity in production parameters of each sample cell, where each sample cell includes a normal cell and an abnormal cell; A preprocessing module, configured to select features whose degree of differentiation between the normal battery cell and the abnormal battery cell exceeds a first preset threshold from the first feature set, and determine a second feature set; An equivalent processing module, used for balancing the number of sample cells whose cell capacities are in each preset capacity interval in the second feature set, and determining a third feature set; The incremental elimination module is used to select features that contribute to model training from the third feature set, determine a fourth feature set, and train the battery cell capacity prediction model through the fourth feature set.

11. A method for predicting battery cell capacity, characterized in that: The method comprises: Obtain the features to be processed of the battery cell to be tested; Based on the feature to be processed, a cell capacity prediction value of the cell to be tested is generated via a cell capacity prediction model, wherein the cell capacity prediction model is generated by training using the model training method according to any one of claims 1 to 9.

12. A battery cell capacity prediction device, characterized in that: The device comprises: The second acquisition module is used to obtain the features to be processed of the battery cell to be tested; A capacity prediction module is used to generate a predicted value of the cell capacity of the cell to be tested based on the feature to be processed via a cell capacity prediction model, wherein the cell capacity prediction model is generated by training using the model training method described in any one of claims 1 to 9.

Images