Lithium-ion battery early life prediction method for scenarios with scarce labeled data

Abstract

The present application relates to the technical field of lithium ion battery prediction and health management, and discloses a lithium ion battery early life prediction method for a labeled data scarce scene, which comprises the following steps: using the charging and discharging data of the first 70 cycle periods of the lithium ion battery, extracting degradation features of a preset dimension, and performing subsequent prediction; using labeled samples to train a plurality of prediction models, using the trained prediction models to obtain pseudo-labels of unlabeled samples; clustering the unlabeled samples to obtain corresponding cluster labels; based on the pseudo-labels and the cluster labels, calculating the confidence score of each unlabeled sample, and selecting the unlabeled sample with a high score and its corresponding pseudo-label to add to the initial training set; and simultaneously, by the change of the average absolute percentage error of the training set, dynamically adjusting the preset confidence threshold and the preset sampling ratio, updating the training set, and training a more accurate target prediction model to predict the early life of the lithium ion battery.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of lithium ion battery prediction and health management, in particular to a lithium ion battery early life prediction method for a labeled data scarce scenario. BACKGROUND

[0002] Lithium ion batteries are widely used in electric vehicles, energy storage systems and other fields due to their high energy density, long cycle life and other advantages. The service life of a lithium ion battery is defined based on its cycle period, and a cycle period refers to a complete charging and discharging process of a lithium ion battery. As the cycle period of a lithium ion battery increases, the performance of the lithium ion battery shows a degradation trend with a periodic characteristic.

[0003] Accurate prediction of the early life of a lithium ion battery is crucial for battery management, maintenance cost control and system reliability improvement. Existing early life prediction methods are mainly divided into two categories: physical model-based methods and data-driven methods. Physical model-based methods, such as equivalent circuit models and electrochemical impedance spectroscopy, rely on accurate modeling of internal battery parameters. However, in practical applications, these parameters are difficult to obtain accurately, resulting in poor universality of physical model-based methods. Data-driven methods, such as support vector machines, random forests and deep learning, can capture complex nonlinear relationships. However, these methods usually rely on a large amount of labeled data. In the early use stage of the battery, the scarcity of data labels severely restricts the performance of the prediction model. In addition, existing methods usually require full life cycle data or a large amount of historical data for modeling, making it difficult to achieve high-precision prediction within the cycle period of the early use stage of the battery.

[0004] In summary, existing lithium battery life prediction methods rely on a large amount of labeled data throughout the cycle, and the scarcity of labeled data leads to insufficient feature extraction of existing models, resulting in insufficient prediction accuracy and difficulty in predicting the early life of a lithium ion battery. SUMMARY

[0005] Therefore, the technical problem to be solved by the present application is to overcome the problem of insufficient feature extraction of the prediction model and low prediction accuracy in the prior art, which relies on scarce full-cycle labeled data for training.

[0006] To solve the above technical problems, the present application provides a lithium ion battery early life prediction method for a labeled data scarce scenario, comprising:

[0007] Using the battery life value of a lithium ion battery as a label, labeled charging and discharging data for the first 70 cycle periods is obtained to construct an initial training set;

[0008] Obtain the unlabeled charge-discharge data of the lithium ion battery in the first 70 cycle periods, and construct a candidate training set;

[0009] Based on the difference between the charge-discharge data corresponding to any two cycle periods in the first 70 cycle periods of the training sample, extract a pre-set dimension of degradation features to form a candidate feature set;

[0010] Based on the candidate feature set, all unlabeled samples are clustered into multiple clusters, and each cluster is assigned a corresponding cluster label; the initial training set is used to train multiple prediction models, and the trained multiple prediction models are used to predict the unlabeled samples to obtain the pseudo label of each unlabeled sample; based on the pseudo label and the cluster label of each unlabeled sample, the confidence score of each unlabeled sample is constructed;

[0011] In the unlabeled samples with a confidence score greater than a pre-set confidence threshold, use a stratified sampling method to select multiple unlabeled samples and their corresponding pseudo labels based on a pre-set sampling ratio to add to the initial training set to update the training set;

[0012] Calculate the change of the mean absolute percentage error of the updated training set and the mean absolute percentage error of the training set before updating, dynamically adjust the pre-set confidence threshold and the pre-set sampling ratio, update the training set, until the mean absolute percentage error of the updated training set converges, and obtain multiple target prediction models;

[0013] Obtain the candidate feature set of the lithium ion battery to be predicted, input it into multiple target prediction models respectively, obtain the corresponding multiple prediction results, and perform weighted fusion to obtain the early life prediction result of the lithium ion battery to be predicted.

[0014] Preferably, based on the difference between the charge-discharge data corresponding to any two cycle periods in the first 70 cycle periods of the training sample, a pre-set dimension of degradation features is extracted to form a candidate feature set, including:

[0015] Based on the difference between the charge-discharge data corresponding to any two cycle periods in the first 70 cycle periods of the training sample, a multi-dimensional degradation feature is extracted;

[0016] Use the recursive feature elimination training elastic net regression model to obtain the ranking score of each dimension of the degradation feature;

[0017] Use the random forest model to predict the importance score of each dimension of the degradation feature;

[0018] Weighted fusion of the ranking score and the importance score of each dimension of the degradation feature is performed to obtain the comprehensive score of each dimension of the degradation feature;

[0019] Select a pre-set dimension of degradation features with the highest comprehensive score to construct a candidate feature set for each sample.

[0020] Preferably, based on the difference of the charge-discharge data of any two cycle periods in the first 70 cycle periods of the training sample, a multi-dimensional degradation feature is extracted, including:

[0021] Any two cycle periods are selected in the first 70 cycle periods of the training sample as the start period and the end period;

[0022] Based on the difference of the charge-discharge data of the start period and the end period, the capacity-voltage curve, the temperature-voltage curve, and the incremental capacity curve are constructed;

[0023] Based on the capacity-voltage curve, the capacity difference value of the start period and the end period is calculated, and the variance, root mean square value, area difference, first derivative standard deviation and second derivative standard deviation of the capacity difference value are calculated;

[0024] Based on the temperature-voltage curve, the temperature difference value of the start period and the end period is calculated, and the first derivative standard deviation and the second derivative standard deviation of the temperature difference value are calculated;

[0025] Based on the incremental capacity curve, the incremental capacity difference value of the start period and the end period is calculated, and the minimum value, extreme point voltage offset and root mean square value of the incremental capacity difference value are calculated;

[0026] The variance of the charge-discharge polarization voltage difference of the start period and the end period is calculated , which is represented as: ; and respectively represent the voltage-normalized capacity curve of the charging stage and the discharging stage;

[0027] The dynamic internal resistance coefficient variance of the start period and the end period is calculated , which is represented as: ; and respectively represent the current absolute value-normalized capacity curve of the charging stage and the discharging stage;

[0028] The standard deviation of the discharge time difference and the charging curve slope difference of the start period and the end period is calculated.

[0029] Preferably, based on the candidate feature set, all unlabeled samples are clustered and divided into multiple clusters, and the elbow rule is used to obtain the best clustering number, including:

[0030] Based on the elbow rule, when the clustering number is , the sum of squared errors of all samples and their cluster centers , which is represented as: ;

[0031] Sort the number of clusters from small to large, calculate the change rate of the error sum of squares under the number of adjacent two clusters, denoted as: ;

[0032] Until the error sum of squares under the number of adjacent two clusters converges, the latter cluster number is obtained as the optimal cluster number;

[0033] Wherein, Indicates the total number of samples in the initial training set; Is an indicator function, when the first Lithium ion battery sample belongs to the first Cluster, its value is 1, otherwise 0; Indicates the candidate feature set of the first Lithium ion battery sample in the initial training set, Indicates the centroid of the first Cluster.

[0034] Preferably, after obtaining the pseudo label of each unlabeled sample, the pseudo label is further smoothed and optimized by using the improved KNN algorithm, including:

[0035] For each unlabeled sample, the unlabeled samples belonging to the same cluster as it are taken as its neighbor samples;

[0036] Calculate the Mahalanobis distance Between each unlabeled sample And its neighbor samples , denoted as:

[0037] ;

[0038] Using KNN label smoothing algorithm, according to the data density in the cluster and the similarity between samples, the pseudo label after smoothing optimization , denoted as:

[0039] ;

[0040] Wherein, Indicates the total number of neighbor samples of the unlabeled sample , is the pseudo label of the neighbor sample, Is an adaptive smoothing coefficient, Is a preset coefficient for controlling the steepness of the smoothing coefficient.

[0041] Preferably, the plurality of prediction models include random forest model, elastic network, support vector machine, gradient boosting regression tree and Gaussian process regression model.

[0042] ​Preferably, based on the pseudo-labels and the cluster labels of each unlabeled sample, a confidence score of each unlabeled sample is constructed, including:

[0043] For the random forest model, the consistency is measured by the variance of the prediction results between trees as the confidence score ; represents the total number of trees in the random forest model, represents the prediction value of the mth tree; represents the overall prediction value of the random forest;

[0044] For elastic network, support vector machine and gradient boosting regression tree, a multi-model consensus mechanism is adopted to generate the confidence score ; represents the prediction value of the current prediction model, represents the prediction mean value of all reference models of the current model, expressed as , M represents the total number of reference models, is the prediction value of the mth reference model; represents the prediction stability of the model;

[0045] For the Gaussian process regression model, the confidence score is calculated based on the prediction variance ; represents the prediction variance, represents the prediction value of the Gaussian process regression model, represents the mean value of the labeled training data.

[0046] Preferably, after obtaining a plurality of target prediction models, the method further includes optimizing the plurality of target prediction models respectively, including:

[0047] The TPE algorithm is used to perform sequential hyperparameter optimization on the random forest model;

[0048] The Gaussian process regression model-based Bayesian optimizer is used to perform hyperparameter space search on the elastic network, support vector machine, gradient boosting regression tree and Gaussian process regression model respectively.

[0049] Preferably, the mean absolute percentage error of the current training set and the change of the mean absolute percentage error of the training set before updating are calculated, and the pre-set confidence threshold and the pre-set sampling ratio are dynamically adjusted, including:

[0050] The pre-set confidence threshold is dynamically adjusted, expressed as:

[0051] ;

[0052] wherein, represents the mth reference model; ​The preset confidence threshold is used when updating the training set. and The preset adjustment coefficient; This represents the change in the mean absolute percentage error of the current training set compared to the mean absolute percentage error of the training set before the update. Indicates the threshold for loss deterioration. Indicates the threshold for loss improvement;

[0053] The preset sampling ratio is dynamically adjusted, as shown below:

[0054] ;

[0055] in, Indicates the first The preset sampling ratio when updating the training set. The growth factor is greater than 1. This indicates the upper limit of the preset sampling ratio. For attenuation coefficients with values ​​less than 1, This indicates the lower limit of the preset sampling ratio.

[0056] Preferably, after obtaining multiple target prediction models, the method further includes ensemble learning of the multiple target prediction models, including:

[0057] For each labeled sample, obtain its predicted label as a meta-feature, which is output by multiple target prediction models.

[0058] A ridge regression model with cross-validation is used as a meta-learner to perform weighted fusion of multiple meta-features to obtain an integrated label;

[0059] Minimize the mean squared error between the real labels of labeled samples and the ensemble labels to obtain the target weight corresponding to each meta-feature;

[0060] Based on the prediction model for each target, the predicted label and its corresponding target weight are output to obtain the ensemble learning prediction result.

[0061] Compared with the prior art, the above-described technical solution of the present invention has the following advantages:

[0062] The early life prediction method for lithium ion batteries in the annotation data scarce scene provided by the application selects the charge and discharge data of any two cycle periods in the first 70 cycle periods of the lithium ion battery to capture the key features of the end of the battery performance stable period and the beginning of the nonlinear degradation starting period, clusters and divides by using the life, and ensures the consistent distribution of the training set and the unannotated data in the cluster by using stratified sampling; after the prediction model is trained by using part of the labeled samples, the unlabeled samples are predicted to obtain pseudo labels; and the unlabeled samples are clustered and divided to assign cluster labels; the confidence score is calculated based on the pseudo labels and the cluster labels; meanwhile, by dynamically adjusting the confidence threshold and the sampling ratio, the pseudo label data with high confidence is gradually added to the training set to avoid noise accumulation, so that the training set is updated and subsequent optimization training is performed. The application can start training based on a small amount of labeled samples, solve the problem that the prior art relies on a large amount of labeled data, effectively utilize limited annotated data and a large amount of unlabeled data for early life prediction of lithium ion batteries, reduce the dependence on annotated data, and effectively improve the accuracy and stability of prediction.

[0063] Meanwhile, the application calculates the change of the mean absolute percentage error of the current training set and the mean absolute percentage error of the training set before updating, dynamically adjusts the preset confidence threshold and the preset sampling ratio, gradually expands the training set, improves the data utilization rate, and reduces the annotation dependence; through the dynamic clustering division and the self-training strategy driven by the confidence, the model trained based on a small amount of labeled samples has higher prediction accuracy, and the accuracy of the early life prediction of the lithium ion battery is improved.

[0064] The application selects the key features highly related to the degradation of the battery by the weighted fusion of the elastic network regression, the recursive feature elimination and the random forest, effectively eliminates redundant information, constitutes a candidate feature set, effectively reduces the feature dimension, improves the sensitivity of the prediction model to the early degradation signal, and further improves the prediction accuracy and efficiency of the model.

[0065] The application uses multiple prediction models including the random forest model, the elastic network, the support vector machine, the gradient boosting regression tree and the Gaussian process regression model to perform heterogeneous model learning; and based on the ensemble learning strategy, the RidgeCV meta-learner is used to perform weighted fusion on the prediction results of the multiple prediction models, the advantages of the models are integrated, and more accurate ensemble learning prediction results are obtained.

[0066] The application uses hyperparameter optimization on the trained target prediction model, further optimizes the model parameters, improves the prediction accuracy of the model, so as to obtain a more accurate stage of the early life of the lithium ion battery, facilitate accurate prediction of the battery degradation trend, optimize the charge and discharge strategy, and prolong the cycle life of the battery. BRIEF DESCRIPTION OF DRAWINGS

[0067] In order to make the content of the present application more easily understood, the present application will be further described in detail below according to specific embodiments of the present application and in conjunction with the accompanying drawings, in which:

[0068] Figure 1 is a step flow chart of the lithium ion battery early life prediction method provided by the present application for the labeled data scarce scenario;

[0069] Figure 2 is a candidate feature set construction flow chart;

[0070] Figure 3 is a step flow chart of model verification;

[0071] Figure 4 is a working principle diagram of the semi-supervised learning framework;

[0072] Figure 5 is a box plot of the RMSE index of various prediction models;

[0073] Figure 6 is a box plot of the MAE index of various prediction models;

[0074] Figure 7 is a box plot of the MAPE index of various prediction models;

[0075] Figure 8 is a box plot of the index of various prediction models. DETAILED DESCRIPTION

[0076] The present application will be further described below in conjunction with the accompanying drawings and specific embodiments, so that those skilled in the art can better understand the present application and implement it, but the embodiments are not limiting to the present application.

[0077] Referring to Figure 1 , it is a step flow chart of the lithium ion battery early life prediction method provided by the present application for the labeled data scarce scenario, and the specific steps include:

[0078] S101: Taking the battery life value of the lithium ion battery as a label, obtaining the labeled charge-discharge data of the first 70 cycle periods, and constructing an initial training set;

[0079] S102: Obtaining the unlabeled charge-discharge data of the lithium ion battery in the first 70 cycle periods, and constructing a candidate training set;

[0080] S103: Based on the difference between the charge-discharge data corresponding to any two cycle periods in the first 70 cycle periods of the training sample, extracting degradation features of a preset dimension, and composing a candidate feature set;

[0081] S104: Cluster all unlabeled samples based on the candidate feature set, divide them into multiple clusters, and assign corresponding cluster labels to each cluster;

[0082] S105: Train multiple prediction models using the initial training set, and use the trained multiple prediction models to predict the unlabeled samples to obtain the pseudo labels of each unlabeled sample;

[0083] S106: Based on the pseudo label and the cluster label of each unlabeled sample, a confidence score of each unlabeled sample is constructed;

[0084] S107: In the unlabeled samples with a confidence score greater than a preset confidence threshold, use a stratified sampling method to select multiple unlabeled samples and their corresponding pseudo labels based on a preset sampling ratio to join the initial training set, and update the training set;

[0085] S108: Calculate the change of the mean absolute percentage error of the updated training set and the mean absolute percentage error of the training set before updating, dynamically adjust the preset confidence threshold and the preset sampling ratio, update the training set, until the mean absolute percentage error of the updated training set converges, and obtain multiple target prediction models;

[0086] S109: Obtain the candidate feature set of the lithium ion battery to be predicted, input it into multiple target prediction models respectively, obtain corresponding multiple prediction results, and perform weighted fusion to obtain the early life prediction result of the lithium ion battery to be predicted.

[0087] The early life prediction method of the lithium ion battery facing the scenario of labeled data scarcity provided by the application selects the charge and discharge data of any two cycle periods in the first 70 cycle periods of the lithium ion battery to capture the key features of the end of the battery performance stable period and the beginning of the nonlinear degradation starting period, clusters and divides the life, and ensures the consistent distribution of the training set and the unlabeled data in the cluster through stratified sampling. After training the prediction model using part of the labeled samples, the pseudo labels of the unlabeled samples are obtained. The unlabeled samples are clustered and divided, and the cluster labels are assigned. Based on the pseudo label and the cluster label, the confidence score is calculated. At the same time, by dynamically adjusting the confidence threshold and the sampling ratio, the pseudo label data with high confidence is gradually added to the training set to avoid noise accumulation, so as to update the training set and perform subsequent optimization training. The application can start training based on only a small amount of labeled samples, solve the problem of relying on a large amount of labeled data in the prior art, effectively utilize limited labeled data and a large amount of unlabeled data for early life prediction of the lithium ion battery, reduce the dependence on labeled data, and effectively improve the prediction accuracy and stability.

[0088] Specifically, referring to Figure 2The diagram shows the flowchart for constructing the candidate feature set. In step S102 of this embodiment, the construction of the candidate feature set includes:

[0089] S102-1: Extract multidimensional degradation features based on the difference in charge and discharge data between any two cycles in the first 70 cycles of the training samples;

[0090] S102-2: Use recursive feature elimination to train the elastic network regression model and obtain the ranking score of each degenerate feature.

[0091] S102-3: Using a random forest model, predict the importance score of each degenerate feature;

[0092] S102-4: The ranking score and importance score of each degenerate feature are weighted and fused to obtain the comprehensive score of each degenerate feature.

[0093] S102-5: Select the pre-defined degenerate feature with the highest comprehensive score to construct a candidate feature set for each sample.

[0094] Among them, the extracted multidimensional degradation features in S102-1 include:

[0095] Two cycles are randomly selected from the first 70 cycles of the training samples as the start cycle and the end cycle;

[0096] Based on the differences in charge and discharge data between the start and end cycles, capacity-voltage curves, temperature-voltage curves, and incremental capacity curves are constructed.

[0097] Based on the capacity-voltage curve, calculate the capacity difference between the start and end cycles, and calculate the variance, root mean square value, area difference, standard deviation of the first derivative, and standard deviation of the second derivative of the capacity difference.

[0098] Based on the temperature-voltage curve, the temperature difference between the start and end periods is calculated, and the standard deviation of the first derivative and the standard deviation of the second derivative of the temperature difference are calculated.

[0099] Based on the incremental capacity curve, calculate the incremental capacity difference between the start and end periods, and calculate the minimum value, extreme point voltage offset, and root mean square value of the incremental capacity difference.

[0100] Calculate the variance of the charge / discharge polarization voltage difference between the start and end of the cycle. , is represented as: ; and The voltage-normalized capacity curves for the charging and discharging phases are shown respectively.

[0101] Calculate the variance of the dynamic internal resistance coefficient between the start and end periods. , is expressed as: ; and respectively represent the current absolute value-normalized capacity curves of the charging stage and the discharging stage;

[0102] The standard deviation of the difference between the discharging time of the start period and the end period and the difference between the slopes of the charging curves is calculated.

[0103] Before constructing the corresponding feature curve based on the charging and discharging data, the embodiment further includes cleaning the charging and discharging data, including: removing high-frequency noise in the charging and discharging data by using a low-pass filter with a preset cutoff frequency; identifying and removing outliers in the charging and discharging data by using the interquartile range method.

[0104] The application filters out key features highly related to battery degradation by weighted fusion of elastic network regression, recursive feature elimination and random forest, effectively removes redundant information, forms a candidate feature set, effectively reduces the feature dimension, at the same time improves the sensitivity of the prediction model to early degradation signals, and further improves the prediction accuracy and efficiency of the model.

[0105] In the embodiment, the K-mean++ clustering algorithm is used to cluster and divide the unlabeled data, and before the division, the elbow rule is used to obtain the best cluster number, including:

[0106] Based on the elbow rule, when the cluster number is , the sum of squared errors of all samples and their corresponding cluster centers , is expressed as: ;

[0107] Sort the cluster numbers from small to large, calculate the change rate of the sum of squared errors under the adjacent two cluster numbers, and express it as: ;

[0108] Until the sum of squared errors under the adjacent two cluster numbers converges, the latter cluster number is obtained as the best cluster number;

[0109] Wherein, Indicates the total number of samples in the initial training set; is an indicator function, and when the th lithium ion battery sample belongs to the th cluster, its value is 1, otherwise it is 0; Indicates the candidate feature set of the th lithium ion battery sample in the initial training set, Indicates the centroid of the th cluster.

[0110] After obtaining the pseudo label of each unlabeled sample in the embodiment, the pseudo label is further smoothed and optimized by using an improved KNN algorithm, including:

[0111] For each unlabeled sample, the unlabeled samples belonging to the same cluster as the unlabeled sample are taken as neighbor samples of the unlabeled sample;

[0112] The Mahalanobis distance of each unlabeled sample and its neighbor samples is calculated , which is expressed as:

[0113] ;

[0114] The pseudo label after smoothing optimization is obtained by using a KNN label smoothing algorithm according to the data density in the cluster and the similarity between samples , which is expressed as:

[0115] ;

[0116] wherein, the total number of neighbor samples of the unlabeled sample , the pseudo label of the neighbor sample, is an adaptive smoothing coefficient, is a preset coefficient for controlling the steepness of the smoothing coefficient.

[0117] Based on the above embodiment, the various prediction models selected by the embodiment of the application include a random forest model, an elastic network, a support vector machine, a gradient boosting regression tree and a Gaussian process regression model.

[0118] For different types of prediction models, the embodiment provides different confidence score calculation methods, including:

[0119] For the random forest model, the consistency is measured by the variance of the prediction results between trees, and the consistency is taken as the confidence score ; wherein, T represents the total number of trees in the random forest model, represents the prediction value of the i-th tree, represents the overall prediction value of the random forest;

[0120] For the elastic network, the support vector machine and the gradient boosting regression tree, a multi-model consensus mechanism is adopted to generate the confidence score ; represents the prediction value of the current prediction model, represents the prediction mean value of all reference models of the current model, and the expression is , and M represents the total number of reference models, ​​is the prediction value of the mth reference model; represents the model prediction stability;

[0121] For a Gaussian process regression model, a confidence score is calculated based on the prediction variance ; represents the prediction variance, represents the prediction value of the Gaussian process regression model, represents the mean of the labeled training data.

[0122] Specifically, in step S107, the change amount of the mean absolute percentage error of the current training set and the mean absolute percentage error of the training set before updating is calculated, and the pre-set confidence threshold and the pre-set sampling ratio are dynamically adjusted, including:

[0123] The pre-set confidence threshold is dynamically adjusted and represented as:

[0124] ;

[0125] wherein, represents the pre-set confidence threshold when the training set is updated for the mth time, and is a pre-set adjustment coefficient; represents the change amount of the mean absolute percentage error of the current training set and the mean absolute percentage error of the training set before updating, represents the loss deterioration threshold, represents the loss improvement threshold; The pre-set sampling ratio is dynamically adjusted and represented as:

[0126]

[0127] ;

[0128] wherein, represents the pre-set sampling ratio when the training set is updated for the mth time, is a growth coefficient with a value greater than 1, represents the upper limit of the pre-set sampling ratio, is a decay coefficient with a value less than 1, represents the lower limit of the pre-set sampling ratio.

[0129] Meanwhile, the change amount of the mean absolute percentage error of the current training set and the mean absolute percentage error of the training set before updating is calculated, the pre-set confidence threshold and the pre-set sampling ratio are dynamically adjusted, and the training set is gradually expanded, thereby improving data utilization and reducing labeling dependence; through dynamic clustering division and confidence-driven self-training strategy, the model trained based on a small amount of labeled samples has higher prediction accuracy, thereby improving the accuracy of lithium-ion battery early life prediction.​​

[0130] The embodiment of the present application further comprises optimizing the plurality of target prediction models respectively after obtaining the plurality of target prediction models, comprising:

[0131] The TPE algorithm is used to perform sequence hyperparameter optimization on the random forest model.

[0132] The Bayesian optimizer based on the Gaussian process regression model is used to perform hyperparameter space search on the elastic network, the support vector machine, the gradient boosting regression tree and the Gaussian process regression model respectively.

[0133] The present application uses hyperparameter optimization on the trained target prediction model to further optimize the model parameters and improve the model prediction accuracy, so as to obtain a more accurate early life stage of the lithium ion battery, facilitate accurate prediction of the battery degradation trend, optimize the charging and discharging strategy, and prolong the cycle life of the battery.

[0134] In the present embodiment, after obtaining the plurality of target prediction models, the plurality of target prediction models are further subjected to ensemble learning, comprising:

[0135] For each labeled sample, the prediction label output by the plurality of target prediction models is obtained as a meta-feature;

[0136] The Ridge regression model with cross-validation is used as a meta-learner to perform weighted fusion on the plurality of meta-features to obtain an ensemble label.

[0137] The target weight corresponding to each meta-feature is obtained by minimizing the mean square error between the real label of the labeled sample and the ensemble label.

[0138] Based on the prediction label output by each target prediction model and the target weight corresponding thereto, an ensemble learning prediction result is obtained.

[0139] The present application uses a plurality of prediction models including a random forest model, an elastic network, a support vector machine, a gradient boosting regression tree and a Gaussian process regression model to perform heterogeneous model learning, and based on an ensemble learning strategy, uses a RidgeCV meta-learner to perform weighted fusion on the prediction results of the plurality of prediction models to obtain a more accurate ensemble learning prediction result.

[0140] Based on the above embodiment, in the present embodiment, the lithium ion battery early life prediction method provided by the present application is used to realize life prediction based on the charging and discharging data of the first 70 cycles of the lithium ion battery, the 10th cycle is selected as the starting cycle, the 70th cycle is selected as the end cycle, and the prediction is performed. Figure 3 As shown in the step flowchart for model verification, the specific steps include:

[0141] S201: Preprocess battery charge and discharge data;

[0142] ① A low-pass filter with a cutoff frequency of 0.1Hz is used to remove high-frequency noise from the data, and outliers are removed by interquartile range (IQR) method.

[0143] Noise is typically caused by sensor errors, changes in the external environment (such as temperature and humidity variations), or short-term fluctuations in the battery itself. The presence of noise can interfere with the model training process, thus affecting the accuracy of prediction results. To address this issue, this embodiment employs low-pass filtering technology, using a low-pass filter with a cutoff frequency of 0.1Hz to remove high-frequency noise from the signal while retaining the main low-frequency signals indicating the battery's health status.

[0144] After noise processing, the next step is to identify and remove outliers. Outliers are extreme values ​​that deviate significantly from the overall trend of the data. These outliers may be caused by factors such as equipment malfunction, sensor reading errors, and data entry problems. The presence of outliers can lead to model overfitting or bias, thus affecting the reliability of predictions. To effectively identify and remove these outliers, this invention uses IQR for outlier processing. The IQR method calculates the first quartile (Q1) and the third quartile (Q3) of the dataset, and uses these two values ​​to calculate the interquartile range (IQR = Q3 - Q1). Based on this, outliers are defined as values ​​below Q1 - 1.5 × IQR or above Q3 + 1.5 × IQR. In this way, outliers can be effectively identified and removed, effectively avoiding the adverse effects of extreme data on the model training process, thereby improving data quality.

[0145] S202: Extract multidimensional degradation features from preprocessed data;

[0146] Based on the publicly available battery degradation dataset (MIT dataset), this embodiment only uses the data from the first 70 cycles. By comparing the differences between the 10th cycle (performance stabilization period) and the 70th cycle (nonlinear degradation initiation period), a capacity-voltage (QV) curve, a temperature-voltage (TV) curve, an incremental capacity (IC) curve, impedance characteristics, and other features based on the differences between the 10th and 70th cycles are constructed to extract 14-dimensional degradation features.

[0147] ① Based on the charge-discharge data from the 10th and 70th cycles, the QV curves of the battery were plotted, and the features of the QV curves were extracted, including the capacity difference and its variance, root mean square value, area difference, standard deviation of the first derivative, and standard deviation of the second derivative between the 10th and 70th cycles, specifically:

[0148] Feature F13: The variance of the discharge capacity difference between the 70th and 10th cycles, reflecting the degree of fluctuation in capacity change;

[0149] Feature F2: Root mean square of discharge capacity difference between the 70th cycle and the 10th cycle, reflecting the overall level of capacity change;

[0150] Feature F3: Standard deviation of discharge capacity difference between the 70th cycle and the 10th cycle, reflecting the dispersion degree of capacity change;

[0151] Feature F4: Standard deviation of the first derivative of discharge capacity difference between the 70th cycle and the 10th cycle, reflecting the fluctuation of capacity change rate;

[0152] Feature F5: Standard deviation of the second derivative of discharge capacity difference between the 70th cycle and the 10th cycle, reflecting the fluctuation of capacity change acceleration;

[0153] ②T-V curve feature, extracting the temperature difference ΔT between the 10th cycle and the 70th cycle, and its first derivative standard deviation and second derivative standard deviation, expressed as:

[0154] Feature F6: Standard deviation of the first derivative of temperature difference between the 70th cycle and the 10th cycle, reflecting the fluctuation of temperature change rate;

[0155] Feature F7: Standard deviation of the second derivative of temperature difference between the 70th cycle and the 10th cycle, reflecting the fluctuation of temperature change acceleration;

[0156] ③IC curve feature, extracting the ΔIC difference between the 10th cycle and the 70th cycle, and its minimum value, extreme point voltage offset, root mean square value, expressed as:

[0157] Feature F8: Minimum value of ΔIC difference between the 70th cycle and the 10th cycle, reflecting the change of electrochemical characteristics of the battery at different cycles;

[0158] Feature F9: Extreme point voltage offset of ΔIC between the 70th cycle and the 10th cycle, reflecting the voltage change of the battery at different cycles;

[0159] Feature F10: Root mean square value of ΔIC between the 70th cycle and the 10th cycle, reflecting the overall level of incremental capacity change of the battery at different cycles;

[0160] ⑤Impedance feature extraction:

[0161] The dynamic correlation between the current response and the capacity evolution of the 10th cycle and the 70th cycle is analyzed, and a capacity-aligned Q index sequence Vc(Q), Vd(Q), Ic(Q) and Id(Q) is generated by linear interpolation, where Q varies from 0 to 1 with a preset step size, V represents voltage, I represents current, and d and c are subscripts representing discharge and charge stages respectively. The variance of the charging / discharging polarization voltage difference and the variance of the dynamic internal resistance coefficient of the difference between the 70th cycle and the 10th cycle are extracted, and the specific formula is as follows:

[0162] Feature F11: Var( , ; and represent the voltage-normalized capacity curves of the charging stage and the discharging stage respectively;

[0163] Feature F12: Var( , ; and represent the current absolute value-normalized capacity curves of the charging stage and the discharging stage respectively;

[0164] 6. Other feature extraction:

[0165] Feature F13: By analyzing the difference in charge and discharge characteristics between the 10th cycle and the 70th cycle, degradation features are extracted from the time and space dimensions, reflecting the change in charge and discharge time of the battery at different cycles;

[0166] Feature F14: The standard deviation of the discharge time difference and the charge curve slope difference between the 70th cycle and the 10th cycle is extracted, reflecting the fluctuation of the charging rate.

[0167] S203: Mixed feature selection framework filters key features:

[0168] In this embodiment, Elastic Net (EN) regression and Recursive Feature Elimination (RFE) are used to remove redundant features, combined with Random Forest (RF) for non-linear feature importance evaluation, to filter out the most influential features for early life prediction. EN regression combines L1 and L2 regularization, which can effectively filter out features that are linearly related to the target variable (early life).

[0169] In this embodiment, RFE is used to repeatedly train the Elastic Net regression model, evaluate and remove the least important features until a preset number of features remain. This embodiment not only takes advantage of the Elastic Net regression in dealing with non-linear features and high-dimensional data, but also provides a quantitative evaluation of the influence of each feature on the model prediction through feature importance.

[0170] Meanwhile, RF makes predictions by building multiple decision trees and combining their results, and evaluates the importance of features. RF is robust to outliers and noisy data, and also provides a quantitative evaluation of feature importance.

[0171] To further optimize the feature selection process, the embodiment combines the results of RFE and RF by weighting. By combining the feature ranking of RFE with the feature importance score of RF, a comprehensive score is assigned to each feature. Specifically, the smaller the ranking of RFE, the more important the feature, so the score of RFE is inversely proportional to the ranking, while RF directly uses the importance score of the feature. Therefore, the comprehensive score of each feature is , expressed as:

[0172] ;

[0173] wherein, represents the proportional score assigned to feature f by RFE, represents the importance score evaluated by RF for feature f; and represent the weighting coefficients of the proportional score and the importance score, which are in the range of 0 to 1; the embodiment selects , to calculate the comprehensive score, effectively reducing the feature dimension, reducing the computational complexity of model training, and improving the robustness and generalization ability of the model.

[0174] In this embodiment, according to the comprehensive score, 12 most important features are selected as model inputs to improve prediction accuracy. At the same time, Pearson correlation coefficient is used to analyze the correlation between the extracted features and the early life of the battery. Pearson Correlation Coefficient is a statistical indicator that measures the strength and direction of linear correlation between two variables. Its value ranges from -1 to 1, where 1 indicates complete positive correlation, -1 indicates complete negative correlation, and 0 indicates no linear correlation. By calculating the Pearson correlation coefficient between each feature and the early life, the linear correlation between the feature and the early life can be evaluated.

[0175] Specifically, the working principle of the hybrid feature selection framework includes:

[0176] S203-1: Use recursive feature elimination method combined with elastic net regression to select features;

[0177] S203-2: Use random forest model to evaluate feature importance to select features;

[0178] S203-3: Weighted fusion of recursive feature elimination and random forest model feature selection results;

[0179] S203-4: Determine the final feature subset.

[0180] S204: Construct a semi-supervised learning framework:

[0181] By improved K-means++ clustering algorithm to dynamically divide the battery life, and combined with improved KNN label smoothing technology, generate high quality pseudo label; through the confidence driven self-training strategy, gradually expand the training set, optimize the generalization ability of the model.

[0182] ① After standardizing the target variable and early life to eliminate the influence of dimension, and through K-means++ clustering, the early life is divided into multiple clusters;

[0183] In order to determine the best cluster number, this embodiment introduces a dynamic threshold adjustment mechanism based on the traditional elbow rule to obtain an improved elbow rule. By calculating the WSS (sum of squared errors) under different cluster numbers and monitoring the WSS change rate, the cluster number is accurately determined, and the different stages of battery life are accurately divided;

[0184] The error square sum when the cluster number is is represented as:

[0185] ;

[0186] Wherein, represents the total number of samples, represents the cluster number, is an indicator function, when the first lithium ion battery sample belongs to the first cluster, its value is 1, otherwise it is 0; represents the candidate feature set of the first lithium ion battery sample in the initial training set, represents the centroid of the first cluster.

[0187] ② Heterogeneous model generates pseudo label:

[0188] Use 10% labeled data to train heterogeneous models including RF, EN, SVM, GBRT, GPR, etc. to predict the unlabeled data and generate initial pseudo label.

[0189] ​For each pseudo label, the prediction value probability density function, the inter-tree prediction result variance, and the "prediction bias + external consistency" mechanism are used to evaluate the pseudo label confidence for different model characteristics, providing a basis for subsequent screening of high-quality pseudo labels.

[0190] ③Improved KNN label smoothing algorithm is used to generate pseudo labels. With the improved KNN label smoothing algorithm, the smoothing coefficient is dynamically adjusted according to the data density within the cluster and the similarity between samples, further optimizing the pseudo labels and improving the quality.

[0191] ④Confidence-driven self-training mechanism, based on improved KNN label smoothing algorithm and dynamic threshold adjustment strategy, uses dynamic confidence screening mechanism, with initial confidence threshold set to 0.9, adjusted by 0.9 times / 1.1 times dynamically, ranging from 10% to 50%.

[0192] By monitoring the average absolute percentage error (MAPE) of the validation set in real time, the threshold and sampling ratio are dynamically adjusted, and high-confidence pseudo label data is gradually added to the training set to optimize the model generalization ability.

[0193] Specifically, referring to Figure 4 , it is a working principle diagram of the semi-supervised learning framework, and the specific process includes:

[0194] S204-1: Mixed feature selection;

[0195] S204-2: Dynamic clustering and pseudo label generation:

[0196] Based on the improved elbow rule, the number of clusters k is dynamically determined, the WSS change rate is monitored, and the battery life stage clusters are divided; 10% labeled data is used to train RF, GPR, SVM, EN, and GBRT models, and initial pseudo labels are generated for unlabeled data;

[0197] For GPR, the confidence is calculated based on the prediction variance, which is the pseudo label , represented as: ; represents the prediction variance, represents the GPR prediction value, represents the mean of the labeled training data;

[0198] For RF, it is based on multiple decision trees, and the consistency is measured by the variance of the prediction results between trees, which is the pseudo label , represented as: ; represents the total number of trees in the RF model, represents the prediction value of the th tree; represents the overall prediction value of the random forest, the more consistent the prediction is, the smaller the variance is, and the higher the confidence is;

[0199] For models without uncertainty output such as SVM, EN, GBRT, a multi-model consensus mechanism is adopted to generate pseudo labels; let its current prediction be , and the consensus prediction mean composed of other reference models (such as RF, GPR, etc.) is: ; M represents the number of reference models, is the prediction function of the mth reference model. At the same time, let its prediction stability on labeled data be , then the confidence can be expressed as: ; when the model prediction value is close to the reference consensus prediction, tends to 0, indicating that the model judgment is consistent with the multi-model result, and the reliability is improved; when the model performs stably on labeled data, that is, is small, indicating that its generalization ability is strong, and the confidence is also improved.

[0200] S204-3: Improved KNN label smoothing algorithm:

[0201] Through the improved KNN label smoothing algorithm, pseudo labels are dynamically generated, and the smoothing coefficient of the pseudo label is adjusted according to the data density within the cluster and the similarity between samples. The specific formula is:

[0202] ; is the predicted label, is the nearest neighbor sample label, is the adaptive smoothing coefficient, is the coefficient that controls the steepness of the smoothing coefficient; is the Mahalanobis distance between the sample and its ith nearest neighbor, which is expressed as: ;

[0203] Fusion of nearest neighbor label and model prediction value, get the optimized pseudo label, can effectively suppress the interference of outliers.

[0204] S204-4: Dynamic threshold and sampling control:

[0205] The initial confidence threshold is 0.9, which ensures that only pseudo label data with high confidence is selected for training, which helps to avoid introducing noise data when the model is not stable; at the same time, the confidence threshold is dynamically adjusted according to the MAPE change of the validation set; when the model performance improves (MAPE value decreases), reduce the threshold by 0.9 times, allow more pseudo label data to enter the training set, thereby expanding the training set size; when the model performance decreases (MAPE value increases), increase the threshold by 1.1 times, reduce the introduction of pseudo label data, prevent the influence of noise data on the model from being too large, and the threshold range is 0.7~0.9.

[0206] The initial sampling ratio is 10%, which is dynamically adjusted according to the MAPE of the validation set. When the performance improves, the sampling ratio is increased by 1.2 times (upper limit 50%), and when the performance decreases, the sampling ratio is reduced by 0.9 times (lower limit 10%). The sampling ratio is dynamically adjusted to ensure the stability and safety of model training.

[0207] Based on the improved KNN label smoothing algorithm and dynamic threshold adjustment strategy, a dynamic confidence screening mechanism is adopted to gradually add high-confidence pseudo-label data to the training set. The threshold and pseudo-label sampling ratio are dynamically adjusted by monitoring the average absolute percentage error (MAPE) of the validation set in real time. The threshold dynamic adjustment formula is:

[0208] ;

[0209] wherein, represents the threshold at time t, , , is the adjustment coefficient, , controls the upper and lower bounds of the threshold; represents the change of the average absolute percentage error of the current training set and the average absolute percentage error of the training set before updating, represents the loss deterioration threshold, represents the loss improvement threshold. This mechanism can dynamically tighten or relax the pseudo-label screening criteria according to the performance of the validation set, especially when the performance degrades, it can improve the screening threshold, thereby reducing the introduction of unreliable samples.

[0210] At the same time, the pseudo-label sampling ratio is dynamically adjusted according to the performance of the validation set , and the updating method of the sampling ratio is:

[0211] ;

[0212] wherein, and represent the growth and decay coefficients, and control the upper and lower bounds of the sampling ratio. This mechanism can speed up the introduction of pseudo-labels when the model performs well, and can shrink in time when the performance fluctuates, thereby improving the stability and safety of model training.

[0213] S204-5: If the MAPE reduction amplitude is <1e-4 for 5 consecutive iterations, terminate training and save the prediction model parameters. That is, after each iteration, calculate the MAPE of the model on the validation set. If the MAPE of the validation set does not significantly improve (i.e., the MAPE value does not further decrease) for 5 consecutive iterations, the early stopping mechanism is triggered, the training is terminated, the overfitting is prevented, the model generalization ability is ensured, and the waste of invalid training resources is avoided.

[0214] Specifically, the embodiment initializes the confidence threshold, sets the initial confidence threshold to 0.9, uses the preliminary model to predict unlabeled data, calculates the confidence, filters high-confidence samples, expands the training set, re-trains the model, evaluates the performance of the validation set, dynamically adjusts the confidence threshold, and checks the early stopping condition.

[0215] In the self-training process, the model is trained based on the current training set and predicts the unlabeled data to generate pseudo labels. The generation of pseudo labels is based on the prediction results of the model and dynamically adjusts the confidence threshold combined with the K-means++ clustering results to ensure the addition of high-confidence samples. In each round of self-training, first, use the current model to predict the unlabeled data to generate pseudo labels. For RF and GPR models, directly calculate the confidence of the predicted value; for EN, GBRT and SVM, use multi-model consensus and cross-validation prediction variance of labeled data to estimate the uncertainty of the prediction and as a measure of confidence; based on the confidence threshold, filter out the pseudo label data with high confidence to ensure the high quality of these data and thus enhance the effectiveness of the training set. Only when the confidence of the pseudo label is higher than the set threshold, it will be added to the training set; the filtered pseudo label data is combined with the original labeled data to update the training set. At the same time, remove the selected samples from the unlabeled pool to prevent repeated use.

[0216] S205: Model collaborative training

[0217] A heterogeneous model pool composed of random forest (RF), elastic network (EN), support vector machine (SVM), gradient boosting regression tree (GBRT), Gaussian process regression (GPR) and other machine learning models is constructed to improve the performance of the model through 10-fold cross-validation and hyperparameter optimization, and the prediction results of each base model are integrated through ensemble learning (EL) to further improve the prediction accuracy and stability.

[0218] ① Construct a heterogeneous model pool:

[0219] Based on the selected feature subset, use the labeled data to train the preliminary model; construct a heterogeneous model pool containing RF, EN, SVM, GBRT, GPR and other machine learning models.

[0220] Among them, RF as a tree-based ensemble method can effectively capture the nonlinear relationship between features and has strong robustness to noise in the data, can process high-dimensional features, and is suitable for most data types; EN combines L1 and L2 regularization, can perform feature selection and handle multicollinearity problems, especially suitable for processing sparse data with a large number of features; SVM is suitable for high-dimensional space, can effectively perform nonlinear classification and regression, and through the selection of appropriate kernel functions, SVM can capture complex patterns in the data, suitable for processing nonlinear problems; GBRT is a powerful ensemble learning method, which can improve the prediction accuracy by gradually correcting the residual of the previous model, it has low sensitivity to outliers and can effectively capture the nonlinear structure of the data; GPR is a regression model based on Bayesian method, which can provide uncertainty estimation for each prediction result, which is very valuable in predicting the early life of the life.

[0221] ②Improve model performance:

[0222] 10-fold cross-validation and hyperparameter optimization methods are used to train each model, and the Bayesian optimizer is used to fine-tune the hyperparameters to obtain stable and efficient model configurations, to tap the potential of the model and improve the performance of the model.

[0223] ③Self-training iterative retraining: In the confidence-driven self-training process, after introducing a batch of high-confidence pseudo-labels, the expanded training set is used to retrain all base models to ensure that the model continues to adapt to new samples;

[0224] ④Ensemble learning fusion: Design an ensemble learning method to fuse the prediction results of various base models to further enhance the accuracy and stability of the prediction and fully utilize the advantages of each model.

[0225] The prediction results of each base model on the validation set or out-of-bag data are used as meta-features, and RidgeCV is used to build a secondary learner to weight and fuse the meta-features to generate the final prediction model;

[0226] Specifically, the base model prediction results are used as meta-features, RidgeCV (regularization parameter α=1.0) is used to train the secondary model, and the weights are dynamically assigned based on the validation set RMSE to build the meta-learner; input the test set data, output the battery life prediction; evaluate the model performance through RMSE, MAPE, MAE, etc. Referring to Figures 5 to 8 , Figure 5 is the box plot of the RMSE index of multiple prediction models, Figure 6 is the box plot of the MAE index of multiple prediction models, Figure 7 is the box plot of the MAPE index of multiple prediction models, Figure 8 is the Whisker plot of indicators.

[0227] ⑤Consistency screening and weight adjustment: Based on the performance of each base model on the validation set (such as MAPE, RMSE, etc.), the weight of each base model in the meta-learner is dynamically adjusted, and the models with poor performance or high bias are removed to further improve the accuracy and stability of the integrated prediction.

[0228] S206: Model optimization:

[0229] ① Hyperparameter optimization:

[0230] For RF models, the TPE (Tree-structured Parzen Estimator) algorithm is used for sequential hyperparameter optimization;

[0231] For EN, SVM, GBRT and GPR models, the Bayesian optimizer based on Gaussian process regression model is used for hyperparameter space search to efficiently find the optimal hyperparameter combination and improve the model performance;

[0232] The optimization strategy of the present embodiment can realize global optimal search of hyperparameter space in a small number of iterations by constructing a probability proxy model of the objective function and combining an acquisition function to dynamically guide the parameter search direction, which has higher parameter search efficiency and better convergence performance compared with traditional grid search and random search.

[0233] S207: Model performance evaluation and result output;

[0234] The trained model is used to predict the unlabeled data, and the model performance is comprehensively evaluated by RMSE, MAPE, R 2 These indicators can objectively reflect the prediction accuracy and reliability of the model, providing strong decision basis for practical application. The final output of the early life prediction result of the battery provides a strong basis for battery management, maintenance cost control, and system reliability improvement.

[0235] Based on the above embodiments, based on two groups of charge and discharge data of lithium ion batteries in the first 70 cycle periods, the models including RF, EN, GBRT, SVM, GPR and EL are used to verify the early life prediction method of lithium ion batteries proposed in the present application, and the RMSE, MAE, MAPE and R 2The same index; this embodiment selects data including the first cycle period and the 70th cycle period, the 10th cycle period and the 70th cycle period, the 30th cycle period and the 70th cycle period, and the 60th cycle period and the 70th cycle period for verification, each verification contains 20% labeled data, and the specific verification model index comparison is shown in Table 1:

[0236] Table 1 Model evaluation index

[0237]

[0238] As can be seen from the table data, for the random forest RF, the RMSE, MAE, MAPE and R 2 All perform best when the selected period is the 30th cycle period and the 70th cycle period; for EN, the RMSE, MAE, MAPE and R 2 All perform best when the selected period is the 30th cycle period and the 70th cycle period; for GBRT, the RMSE, MAE, MAPE and R 2 All perform best when the selected period is the 10th cycle period and the 70th cycle period; for SVM, the RMSE, MAE, MAPE and R 2 All perform best when the selected period is the 1st cycle period and the 70th cycle period; for GPR, the RMSE, MAE, MAPE and R 2 All perform best when the selected period is the 30th cycle period and the 70th cycle period, but perform very poorly in other periods; for EL, the RMSE, MAE, MAPE and R 2 All perform best when the selected period is the 30th cycle period and the 70th cycle period.

[0239] Therefore, based on the index comparison of various models at different cycle intervals, the present application concludes that for long cycle intervals, such as the 1st cycle and the 70th cycle, or the 10th cycle and the 70th cycle, the weights of GBRT and SVM models in ensemble learning can be increased; for medium cycle intervals, such as the 30th cycle and the 70th cycle, the weights of GPR, RF and EN models in ensemble learning can be increased; for short cycle intervals, although prediction can also be achieved, the prediction result accuracy is relatively poor, therefore, this embodiment preferentially selects long cycle interval cycle data for lithium ion battery early life prediction; in particular, the prediction effect of the 30th cycle and the 70th cycle is better.

[0240] The application significantly reduces the dependence on labeled data, can achieve good prediction performance in the case of data scarcity, adopts a hybrid feature selection framework, combines RFE and RF importance evaluation method, effectively reduces the feature dimension, and reduces the calculation complexity of model training, the proposed semi-supervised self-training algorithm gradually adds the pseudo-label data with high confidence to the training set by dynamically adjusting the confidence threshold and sampling ratio, effectively avoids the noise accumulation and sample selection bias problem, and ensures that the model is continuously optimized in the iteration process, the prediction results of multiple base models are integrated through ensemble learning, the advantages of different models are integrated, and the overall prediction performance is further improved, the application can achieve excellent prediction effect, provides reliable technical support for battery health state monitoring and management, can help optimize the battery management strategy, prolong the service life of the battery, reduce the system operation cost, improve the reliability and safety of the overall system, has important practical application value and wide popularization prospect.

[0241] Those skilled in the art will appreciate that embodiments of the application can be provided as methods, systems, or computer program products. Accordingly, the application can be embodied in the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the application can be embodied in the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk memory, CD-ROMs, optical memory, etc.) having computer usable program code embodied thereon.

[0242] The application is described with reference to flowcharts and / or block diagrams according to the methods, devices (systems), and computer program products of the embodiments of the application. It should be understood that each flow and / or block in the flowcharts and / or block diagrams, and the combination of flows and / or blocks in the flowcharts and / or block diagrams can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, a special-purpose computer, an embedded processor, or other programmable data processing apparatus to produce a machine, so that the instructions executed by the processor of the computer or other programmable data processing apparatus produce a device that implements the functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in one or more flows and / or blocks Figure 1 The functions specified in one or more flows and / or blocks

[0243] These computer program instructions can also be stored in a computer-readable memory that can direct the computer or other programmable data processing apparatus to work in a specific manner, so that the instructions stored in the computer-readable memory produce a product including instruction devices that implement the functions specified in the flowcharts and / or block diagrams. Figure 1 The functions specified in one or more flows and / or blocks Figure 1 The functions specified in one or more flows and / or blocks

[0244] These computer program instructions can also be loaded into a computer or other programmable data processing devices, so that a series of operational steps are performed on the computer or other programmable data processing devices to generate a computer implemented process, so that the instructions executed on the computer or other programmable data processing devices provide steps for implementing the function specified in the flowchart Figure 1 one flow or multiple flows and / or blocks Figure 1 one block or multiple blocks.

[0245] Obviously, the above embodiments are only examples for clearly illustrating the present application, and are not intended to limit the embodiments. Based on the above description, other different forms of changes or variations can be made by those skilled in the art. Here, all the embodiments are not required to be exhausted, and the obvious changes or variations derived therefrom are still within the protection scope of the present application.

Claims

1. A method for predicting the early lifespan of lithium-ion batteries in scenarios where labeled data is scarce, characterized in that, include: Using the battery life of lithium-ion batteries as a label, we obtained labeled charge-discharge data for the first 70 cycles to construct an initial training set. Obtain unlabeled charge-discharge data of lithium-ion batteries in the first 70 cycles and construct a candidate training set; Based on the difference in charge and discharge data between any two cycles in the first 70 cycles of the training samples, a set of candidate features of a preset dimension is extracted. Based on the candidate feature set, all unlabeled samples are clustered into multiple clusters, and a corresponding cluster label is assigned to each cluster. Multiple prediction models are trained using the initial training set, and the trained prediction models are used to predict unlabeled samples to obtain pseudo-labels for each unlabeled sample. Based on the pseudo-labels and cluster labels of each unlabeled sample, a confidence score is constructed for each unlabeled sample. In unlabeled samples with confidence scores greater than a preset confidence threshold, a stratified sampling method is used to select multiple unlabeled samples and their corresponding pseudo-labels based on a preset sampling ratio and add them to the initial training set to update the training set. Calculate the change between the mean absolute percentage error of the updated training set and the mean absolute percentage error of the training set before the update, dynamically adjust the preset confidence threshold and preset sampling ratio, update the training set, until the mean absolute percentage error of the updated training set converges, and obtain multiple target prediction models. The candidate feature set of the lithium-ion battery to be predicted is obtained and input into multiple target prediction models. The corresponding prediction results are then weighted and fused to obtain the early life prediction result of the lithium-ion battery to be predicted. The dynamic adjustment of the preset reliability threshold is expressed as follows: ； in, Indicates the first The preset confidence threshold is used when updating the training set. and The preset adjustment coefficient; This represents the change in the mean absolute percentage error of the current training set compared to the mean absolute percentage error of the training set before the update. Indicates the threshold for loss deterioration. Indicates the threshold for loss improvement; The preset sampling ratio is dynamically adjusted, as shown below: ； in, Indicates the first The preset sampling ratio when updating the training set. The growth factor is greater than 1. This indicates the upper limit of the preset sampling ratio. For attenuation coefficients with values ​​less than 1, This indicates the lower limit of the preset sampling ratio.

2. The method for predicting the early lifespan of lithium-ion batteries in scenarios with scarce labeled data as described in claim 1, characterized in that, Based on the differences in charge-discharge data between any two cycles in the first 70 cycles of the training samples, degradation features of a preset dimension are extracted to form a candidate feature set, including: Based on the difference in charge and discharge data between any two cycles in the first 70 cycles of the training samples, multidimensional degradation features are extracted. Recursive feature elimination is used to train an elastic network regression model to obtain the ranking score of each degenerate feature. Using a random forest model, the importance score of each degenerate feature is predicted; The ranking score and importance score of each degenerate feature are weighted and fused to obtain the comprehensive score of each degenerate feature. Select the predefined degenerate feature with the highest comprehensive score to construct a candidate feature set for each sample.

3. The method for predicting the early lifespan of lithium-ion batteries in scenarios with scarce labeled data according to claim 2, characterized in that, Based on the differences in charge-discharge data between any two cycles in the first 70 cycles of the training samples, multidimensional degradation features are extracted, including: Two cycles are randomly selected from the first 70 cycles of the training samples as the start cycle and the end cycle; Based on the differences in charge and discharge data between the start and end cycles, capacity-voltage curves, temperature-voltage curves, and incremental capacity curves are constructed. Based on the capacity-voltage curve, calculate the capacity difference between the start and end cycles, and calculate the variance, root mean square value, area difference, standard deviation of the first derivative, and standard deviation of the second derivative of the capacity difference. Based on the temperature-voltage curve, the temperature difference between the start and end periods is calculated, and the standard deviation of the first derivative and the standard deviation of the second derivative of the temperature difference are calculated. Based on the incremental capacity curve, calculate the incremental capacity difference between the start and end periods, and calculate the minimum value, extreme point voltage offset, and root mean square value of the incremental capacity difference. Calculate the variance of the charge / discharge polarization voltage difference between the start and end of the cycle. , is represented as: ; and The voltage-normalized capacity curves for the charging and discharging phases are shown respectively. Calculate the variance of the dynamic internal resistance coefficient between the start and end periods. , is represented as: ; and These represent the absolute current-normalized capacity curves for the charging and discharging phases, respectively. Calculate the standard deviation of the difference in discharge time and the difference in the slope of the charging curve between the start and end cycles.

4. The method for predicting the early lifespan of lithium-ion batteries in scenarios with scarce labeled data according to claim 1, characterized in that, Based on the candidate feature set, all unlabeled samples are clustered into multiple clusters. This process also includes using the elbow rule to determine the optimal number of clusters, including: Based on the elbow rule, the number of clusters is calculated as follows: At that time, the sum of squared errors of all samples to their respective cluster centers , is represented as: ; Sort the clusters by number of clusters from smallest to largest, and calculate the rate of change of the sum of squared errors for two adjacent cluster numbers, expressed as: ; Until the sum of squared errors of two adjacent cluster numbers converges, the latter cluster number is taken as the optimal cluster number. in, This represents the total number of samples in the initial training set; For indicator functions, when the first The lithium-ion battery sample belongs to the first When there are 1 clusters, its value is 1; otherwise, it is 0. Indicates the first number in the initial training set. A candidate feature set for a lithium-ion battery sample. Indicates the first The centroid of each cluster.

5. The method for predicting the early lifespan of lithium-ion batteries in scenarios with scarce labeled data as described in claim 1, characterized in that, After obtaining the pseudo-label for each unlabeled sample, the algorithm also includes smoothing optimization of the pseudo-label using an improved KNN algorithm, including: For each unlabeled sample, the unlabeled samples belonging to the same cluster are taken as its nearest neighbors. Calculate each unlabeled sample Its nearest neighbor samples Mahalanobis distance , is represented as: ； Using the KNN label smoothing algorithm, smoothed and optimized pseudo-labels are obtained based on the data density within clusters and the similarity between samples. , is represented as: ； in, Indicates unlabeled samples The total number of nearest neighbor samples, These are pseudo-labels for nearest neighbor samples. For adaptive smoothing coefficients, A preset coefficient used to control the steepness of the change in the smoothing coefficient.

6. The method for predicting the early lifespan of lithium-ion batteries in scenarios with scarce labeled data according to claim 1, characterized in that, Multiple prediction models are available, including random forest, elastic network, support vector machine, gradient boosting regression tree and Gaussian process regression model.

7. The method for predicting the early lifespan of lithium-ion batteries in scenarios with scarce labeled data as described in claim 6, characterized in that, Based on the pseudo-label and cluster label of each unlabeled sample, a confidence score is constructed for each unlabeled sample, including: For the random forest model, the variance of the prediction results between trees is used to measure consistency and is used as the confidence score. ; This represents the total number of trees in the random forest model. Indicates the first The predicted value of each tree; This represents the overall prediction value of the random forest; For Elastic Networks, Support Vector Machines, and Gradient Boosting Regression Trees, a multi-model consensus mechanism is employed to generate confidence scores. ; This represents the predicted value from the current prediction model. The expression represents the mean prediction of all reference models for the current model. M represents the total number of reference models. This is the predicted value of the m-th reference model; Indicates the stability of the model's predictions; For Gaussian process regression models, confidence scores are calculated based on prediction variance. ; Indicates the prediction variance. This represents the predicted value from the Gaussian process regression model. This represents the mean of the labeled training data.

8. The method for predicting the early lifespan of lithium-ion batteries in scenarios with scarce labeled data as described in claim 6, characterized in that, After obtaining multiple target prediction models, the process also includes optimizing each target prediction model separately, including: The TPE algorithm was used to optimize the sequence hyperparameters of the random forest model. We used a Bayesian optimizer based on a Gaussian process regression model to perform hyperparameter space search on elastic networks, support vector machines, gradient boosting regression trees, and Gaussian process regression models, respectively.

9. The method for predicting the early lifespan of lithium-ion batteries in scenarios with scarce labeled data according to claim 1, characterized in that, After acquiring multiple target prediction models, the process also includes ensemble learning of these models, including: For each labeled sample, obtain its predicted label as a meta-feature, which is output by multiple target prediction models. A ridge regression model with cross-validation is used as a meta-learner to perform weighted fusion of multiple meta-features to obtain an integrated label; Minimize the mean squared error between the real labels of labeled samples and the ensemble labels to obtain the target weight corresponding to each meta-feature; Based on the prediction model for each target, the predicted label and its corresponding target weight are output to obtain the ensemble learning prediction result.

Images