Method for predicting remaining service life of lithium ion battery

Abstract

The invention provides a method for predicting the residual service life of a lithium ion battery, and relates to the technical field of prediction of the residual service life of the lithium ion battery, and the method comprises the steps: S1, employing an intermittent current interruption method to charge and discharge the battery, and recording a voltage signal; s2, extracting voltage characteristics before interruption, during interruption and during current recovery after charging and discharging reach a fixed capacity; s3, calculating the capacity of the battery by using an ampere-hour integral method, and constructing a data set of the characteristics and the capacity; s4, dividing a data set and training and testing the L1 regularization model; s5, screening features, and training and testing the model again; and S6, after multiple batteries are verified to be feasible, calculating and predicting the battery capacity by using the screening training model, and calculating the residual life by combining the rated capacity, the attenuation rate and the scrap threshold. Through intermittent voltage measurement, battery capacity calculation through an ampere-hour integral method, L1 model optimization and multi-battery verification, the residual service life of the lithium ion battery is accurately predicted, and the reliability and generalization of the method are improved.

Description

Technical Field

[0001] This invention relates to the field of predicting the remaining lifespan of lithium-ion batteries, and more specifically, to a method for predicting the remaining lifespan of lithium-ion batteries. Background Technology

[0002] With the rapid development of the new energy industry, lithium-ion batteries are widely used in electric vehicles, energy storage systems, and other fields due to their advantages such as high energy density and long cycle life. However, batteries experience capacity decay during use due to charge-discharge cycles and changes in ambient temperature. The remaining lifespan directly affects the operational safety and efficiency of equipment, making accurate prediction crucial for ensuring reliable battery application. Furthermore, the performance of lithium-ion batteries from different manufacturers and using different cathode materials varies significantly, placing higher demands on the generalizability of prediction methods and driving the industry to explore more accurate and universally applicable remaining lifespan prediction technologies.

[0003] Traditional methods for predicting the remaining lifespan of lithium-ion batteries often rely on single charge / discharge parameters, failing to fully capture the dynamic voltage characteristics during current interruption and recovery phases, resulting in incomplete feature extraction. Furthermore, some methods do not consider the impact of ambient temperature on test parameters; when the temperature deviates from the normal range, data acquisition errors can easily reduce the accuracy of capacity calculations. In addition, traditional model training does not perform feature weighting, leading to redundant features that increase model complexity and may introduce interference, resulting in larger prediction errors. Moreover, the lack of multi-type battery validation steps results in weak generalization ability and difficulty in adapting to different lithium-ion battery specifications.

[0004] Therefore, it is necessary to design a method to predict the remaining lifespan of lithium-ion batteries, in order to solve the problems of incomplete feature extraction, failure to consider the impact of temperature on testing, large errors caused by redundant features in the model, weak generalization ability, and inability to accurately adapt to different types of lithium-ion batteries in the existing technology. Summary of the Invention

[0005] In view of this, the present invention proposes a method for predicting the remaining service life of lithium-ion batteries, aiming to solve the problems of incomplete feature extraction, failure to consider the influence of temperature, large model error and weak generalization in the existing technology, which makes it difficult to accurately adapt to different types of lithium-ion batteries.

[0006] In one aspect, the present invention provides a method for predicting the remaining lifespan of a lithium-ion battery, comprising: S1, The lithium-ion battery is subjected to charge-discharge cycle test using an intermittent current interruption method, and the voltage signal during the test is recorded at preset time intervals; S2, after each charge-discharge cycle reaches a fixed capacity, extract voltage signal features from the voltage signal. The voltage signal features include: voltage signal features before interruption, voltage signal features during interruption, and voltage signal features when current is restored. S3, calculate the battery capacity using the ampere-hour integration method, and construct a dataset of the voltage signal characteristics and target values, wherein the target value is the calculated battery capacity; S4. Divide the dataset into a training set and a test set. Train an L1 regularized linear regression model using the training set to obtain a training model. Test the training model using the test set to obtain test data. Calculate an evaluation index using the test data and evaluate the predictive performance of the training model using the evaluation index. The evaluation index includes: mean absolute error, mean square error, root mean square error, and mean absolute percentage error. S5. Analyze the weight values ​​of each voltage signal feature in the training model, and select the selected voltage signal features whose weight values ​​are greater than or equal to the weight threshold. Replace the voltage signal features in the training set and the test set with the selected voltage signal features to obtain a selected training set and a selected test set. Train an L1 regularized linear regression model using the selected training set to obtain a selected training model. Test the selected training model using the selected test set to obtain selected test data. Calculate the evaluation index using the selected test data and evaluate the predictive performance of the selected training model using the evaluation index. S6. Repeat S1-S5 using data from multiple lithium-ion batteries to verify the feasibility of the method. When the method is feasible, use the screening and training model to obtain the predicted battery capacity, and use the predicted battery capacity, the rated capacity of the lithium-ion battery, the lithium-ion battery degradation rate, and the capacity degradation to the scrap threshold to calculate the remaining service life of the lithium-ion battery.

[0007] Furthermore, when conducting charge-discharge cycle tests on lithium-ion batteries using an intermittent current interruption method and recording voltage signals during the test at preset time intervals, the process includes: Based on the preset constant current charging and discharging duration and the preset current interruption duration, the constant current charging and discharging operation and the current interruption operation are performed alternately, and the voltage signal during the constant current charging and discharging process and the current interruption process is continuously recorded at preset time intervals. When the ambient temperature is within the preset temperature range, the test will be performed according to the preset constant current charging and discharging duration, preset current interruption duration, and preset time interval. When the ambient temperature of the test environment is greater than the upper limit of the preset temperature range, the constant current charging and discharging time, the current interruption time and the recording time interval are shortened by a preset ratio. When the ambient temperature is lower than the lower limit of the preset temperature range, the constant current charging and discharging time, current interruption time, and recording time interval will be extended by a preset ratio.

[0008] Further, extracting voltage signal features from the voltage signal includes: Based on the voltage signal characteristics before the interruption, one voltage signal corresponding to the last voltage signal recording time before the current interruption operation begins is selected as the voltage signal characteristics before the interruption. Based on the voltage signal characteristics during the interruption, all voltage signals recorded at preset time intervals during the entire current interruption duration are selected as the voltage signal characteristics during the interruption. Based on the voltage signal characteristics during the recovery current, the total duration of the recovery current stage is equally divided into a preset number of continuous time periods, and the same preset number of voltage signals are extracted at equal intervals within each time period. The voltage signals extracted from each time period together constitute the voltage signal characteristics during the recovery current.

[0009] Furthermore, when calculating battery capacity using the ampere-hour integration method, the following steps are included: When calculating battery capacity, instantaneous current is collected at the time points when the voltage signal is recorded in step S1, and then... - The battery capacity is obtained by dividing the instantaneous current into several small time intervals and accumulating the product of the instantaneous current and the small time intervals corresponding to the instantaneous current. The minute time interval corresponding to the instantaneous current is the time interval between the previous instantaneous current acquisition time point and the current instantaneous current acquisition time point; When the fluctuation amplitude of the instantaneous current during charging and discharging is less than the preset fluctuation threshold, the battery capacity is calculated using the ampere-hour integration method. When the fluctuation amplitude of the instantaneous current during charging and discharging is greater than or equal to the preset fluctuation threshold, the length of the tiny time interval is shortened by a preset reduction ratio before the battery capacity is calculated by the ampere-hour integration method.

[0010] Furthermore, when dividing the dataset into training and test sets, the following steps are included: Obtain the total number of charge-discharge cycles for the lithium-ion battery. The total number of charge-discharge cycles is divided according to a preset training ratio; Multiply the total number of charge-discharge cycles by the preset training ratio to obtain the number of training charge-discharge cycles; The voltage signal features and the target value of several charge-discharge cycles in the dataset are used as the training set; The number of preset training ratios is the preset number of training sessions, and each preset training ratio is different. The voltage signal features and target values ​​of the remaining charge and discharge cycles after removing the current training set from the dataset are used as the test set corresponding to the current training set; The training set and its corresponding test set together form a complete dataset.

[0011] Furthermore, when dividing the dataset into training and test sets, the method also includes: When the total number of charge-discharge cycles is greater than the preset charge-discharge cycle threshold, the training set and the test set are divided according to the preset training ratio. When the total number of charge-discharge cycles is less than or equal to a preset charge-discharge cycle threshold, the percentage of the training set in the total number of charge-discharge cycles is increased by a preset increase ratio before the training set and the test set are divided.

[0012] Further, the process of training an L1 regularized linear regression model using the training set to obtain a trained model, testing the trained model using the test set to obtain test data, calculating an evaluation index using the test data, and evaluating the predictive performance of the trained model using the evaluation index includes: When the values ​​of each of the evaluation indicators are less than the corresponding preset indicator threshold, step S5 is executed. When the value of any of the evaluation indicators is greater than or equal to the corresponding preset indicator threshold, the regularization strength is adjusted according to the preset adjustment step size, and the L1 regularized linear regression model is retrained using the training set. The evaluation indicators are then retested and calculated using the test set until the values ​​of all the evaluation indicators are less than the corresponding preset indicator threshold.

[0013] Further, when analyzing the weight values ​​of each voltage signal feature in the training model and selecting voltage signal features with weight values ​​greater than or equal to a weight threshold, the process includes: When the number of selected voltage signal features exceeds the upper limit of the preset key feature number range, the weight threshold is increased by a preset increment and the selection is repeated until the number of selected voltage signal features is within the preset key feature number range. When the number of selected voltage signal features is less than the lower limit of the preset key feature number range, the weight threshold is reduced by a preset reduction and the selection is repeated until the number of selected voltage signal features is within the preset key feature number range.

[0014] Further, the voltage signal features in the training set and the test set are replaced with the selected signal voltage features to obtain a selected training set and a selected test set. An L1 regularized linear regression model is trained using the selected training set to obtain a selected training model. The selected training model is tested using the selected test set to obtain selected test data. An evaluation index is calculated using the selected test data, and the predictive performance of the selected training model is evaluated using the evaluation index, including: When the difference between each evaluation index of the selected training model and the evaluation index corresponding to the training model is less than the preset difference threshold, step S6 is executed. When the difference between any evaluation index of the selected training model and the evaluation index corresponding to the training model is greater than or equal to a preset difference threshold, the regularization strength of the selected training model is adjusted by a preset adjustment step size, and the L1 regularized linear regression model is retrained using the selected training set. The evaluation index is then tested and calculated again using the selected test set until the difference between each evaluation index of the selected training model and the evaluation index corresponding to the training model is less than the preset difference threshold.

[0015] Furthermore, when repeating S1-S5 using data from multiple lithium-ion batteries and verifying the feasibility of the method, the following steps are taken: When the percentage of batteries whose values ​​are all less than the corresponding preset threshold values ​​of the selection training model is greater than the preset percentage threshold, the determination method is feasible. When the percentage of batteries whose values ​​of each evaluation index of the screening training model are less than the corresponding preset index threshold is less than or equal to the preset percentage threshold, the preset number of non-compliant lithium-ion batteries is adjusted until the percentage of batteries whose values ​​of each evaluation index of the screening training model are less than the corresponding preset index threshold is greater than the preset percentage threshold. The adjustment rule for the preset number is to increase by 1 or decrease by 1.

[0016] Compared with existing technologies, the beneficial effects of this invention are as follows: The method for predicting the remaining lifespan of lithium-ion batteries records voltage signals through intermittent current interruption, comprehensively extracting voltage features before, during, and after current recovery, avoiding the problem of single features in traditional methods; it combines the ampere-hour integration method to accurately calculate battery capacity, and adjusts the calculation interval according to current fluctuations to improve capacity calculation accuracy; it optimizes the model through L1 regularization model training and feature selection to reduce redundant feature interference and lower prediction errors; it also considers the influence of temperature on test parameters, adjusting the test duration and interval according to temperature to adapt to different environments; finally, multi-cell verification ensures the generalization of the method, enabling accurate prediction of the remaining lifespan of lithium-ion batteries and providing a guarantee for reliable battery applications. Attached Figure Description

[0017] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1A flowchart of a method for predicting the remaining lifespan of a lithium-ion battery provided in an embodiment of the present invention; Figure 2 Lasso weight graph for a machine learning model applying 202 voltage features; Figure 3 A graph showing the prediction performance of a machine learning model that applies 202 voltage features; Figure 4 A graph showing the prediction performance of a machine learning model that applies three voltage features; Figure 5 The prediction performance of a machine learning model applying three voltage features to three batteries is shown in the graph.

[0018] Figure 6 This is a performance evaluation table for predicting subsequent test sets using 202 voltage features and 3 voltage features in this method. Detailed Implementation

[0019] Exemplary embodiments of the present invention will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present invention are shown in the drawings, it should be understood that the present invention can be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present invention and to fully convey the scope of the invention to those skilled in the art. It should be noted that, without conflict, the embodiments and features described herein can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0020] Reference Figure 1 As shown in some embodiments of this application, a method for predicting the remaining lifespan of a lithium-ion battery includes: S1, The lithium-ion battery is subjected to charge-discharge cycle test using an intermittent current interruption method, and the voltage signal during the test is recorded at preset time intervals; S2, after each charge-discharge cycle reaches a fixed capacity, extract voltage signal features from the voltage signal. The voltage signal features include: voltage signal features before interruption, voltage signal features during interruption, and voltage signal features when current is restored. S3, calculate the battery capacity using the ampere-hour integration method, and construct a dataset of the voltage signal characteristics and target values, wherein the target value is the calculated battery capacity; S4. Divide the dataset into a training set and a test set. Train an L1 regularized linear regression model using the training set to obtain a training model. Test the training model using the test set to obtain test data. Calculate an evaluation index using the test data and evaluate the predictive performance of the training model using the evaluation index. The evaluation index includes: mean absolute error, mean square error, root mean square error, and mean absolute percentage error. S5. Analyze the weight values ​​of each voltage signal feature in the training model, and select the selected voltage signal features whose weight values ​​are greater than or equal to the weight threshold. Replace the voltage signal features in the training set and the test set with the selected voltage signal features to obtain a selected training set and a selected test set. Train an L1 regularized linear regression model using the selected training set to obtain a selected training model. Test the selected training model using the selected test set to obtain selected test data. Calculate the evaluation index using the selected test data and evaluate the predictive performance of the selected training model using the evaluation index. S6. Repeat S1-S5 using data from multiple lithium-ion batteries to verify the feasibility of the method. When the method is feasible, use the screening and training model to obtain the predicted battery capacity, and use the predicted battery capacity, the rated capacity of the lithium-ion battery, the lithium-ion battery degradation rate, and the capacity degradation to the scrap threshold to calculate the remaining service life of the lithium-ion battery.

[0021] Specifically, the training model and the selection training model are essentially the same, only the training and test sets are different: the training model in step S4 and the selection training model in step S5 are essentially L1 regularized linear regression models, only the selection training set and selection test set of the selection training model are composed of selection voltage signal features with weight values ​​greater than or equal to the weight threshold extracted from the training set and test set of the training model based on the weight threshold.

[0022] Specifically, "fixed capacity" refers to a specific capacity value determined to ensure consistent voltage signal feature extraction conditions in each charge-discharge cycle. This is achieved by testing the rated capacity of the lithium-ion battery, considering the capacity decay pattern during charge-discharge cycles, and verifying through pre-testing (confirming that the voltage signal features at this capacity stably reflect the battery's capacity state). When the battery reaches this capacity value during each charge-discharge cycle, the current charge-discharge operation is stopped, and an intermittent current interruption (ICI) test is performed. This allows for the extraction of voltage signal features before, during, and after current recovery from the recorded voltage signals, providing consistent feature extraction nodes for subsequent construction of a dataset of voltage signal features and battery capacity. The specific process for determining the fixed capacity is as follows: first, the rated capacity of the target type of lithium-ion battery is tested, and then multiple cells of this type are tested... The battery undergoes multiple charge-discharge cycle tests, and the actual capacity data for each cycle is recorded. Capacity decay curves are plotted to analyze its decay pattern (such as decay rate and capacity range in the stable decay stage). From this stable decay stage, multiple capacity values ​​within a reasonable proportion of the rated capacity are selected as fixed capacity candidate values. Subsequently, for each candidate value, in the charge-discharge cycles of multiple batteries of this type, intermittent current interruption (ICI) tests are triggered when the charge-discharge reaches the candidate value. The corresponding voltage signal features are extracted. By calculating the correlation between the voltage signal features and the actual battery capacity under each candidate value, the consistency of features between different batteries, and the stability of features changing with the number of cycles, the candidate value with the highest correlation, the most stable features, and the ability to continuously reflect the battery capacity state is selected. Finally, this candidate value is determined as the fixed capacity used uniformly in the entire prediction method.

[0023] Specifically, the process of constructing the dataset of voltage signal features and target values ​​is as follows: First, after completing the intermittent current interruption (ICI) test for each charge-discharge cycle, voltage signal features are extracted from the test data of that cycle: One voltage signal corresponding to the last voltage signal recording moment before the start of the current interruption operation is selected as the voltage signal feature before the interruption; all voltage signals recorded at preset time intervals throughout the entire current interruption duration are selected as the voltage signal features during the interruption; the total duration of the current recovery phase is divided into a preset number of continuous time periods (each time period has an equal duration); within each time period, the same set number of voltage signals are extracted at equal intervals as the voltage signal features during current recovery; and these three types of voltage signal features are integrated into the voltage signal feature vector corresponding to that charge-discharge cycle; simultaneously, the ampere-hour integral method formula is used... (in For battery capacity, This is the start time of the charge / discharge cycle. This refers to the charging / discharging termination time. for The instantaneous current at each moment is used to calculate the battery capacity corresponding to the charge-discharge cycle, and this capacity is used as the target value. Then, the voltage signal feature vector of each charge-discharge cycle is associated with the target value (battery capacity) calculated for that cycle to form a single sample of "voltage signal feature vector - target value". Finally, the "voltage signal feature vector - target value" samples corresponding to all completed charge-discharge cycles are collected and organized according to a unified data structure (such as a table, where each row corresponds to one sample and the columns correspond to each dimension of the voltage signal feature vector and the target value, respectively), thereby constructing a dataset of voltage signal features and target values.

[0024] Specifically, the intermittent current interruption (ICI) method involves charging / discharging at a constant current (3250mA, 1C rate) for 150 seconds, followed by a 10-second interruption; then charging / discharging at a constant current for another 150 seconds, followed by a 10-second interruption. The charge / discharge cycle test is conducted within a voltage range of 2.7V-4.35V, with one voltage signal recorded every 0.1 seconds.

[0025] Specifically, capacity decay to the scrap threshold refers to the critical value at which the actual capacity of a lithium-ion battery decays to the point where it can no longer meet preset usage requirements or safety standards during charge-discharge cycles. It is usually expressed as a percentage of the battery's rated capacity (e.g., 80% of the rated capacity). When the actual capacity of the battery falls below this threshold, it is determined that the battery has reached the scrap state and can no longer be used normally. Obtaining this threshold requires a multi-step process: First, refer to the capacity specifications for battery life termination in lithium-ion battery industry standards (such as IEC and GB standards) to determine the basic threshold range; second, combine the usage requirements of actual battery application scenarios (such as energy storage and electric vehicles), test whether the discharge performance and safety performance of batteries at different capacities meet the scenario requirements, and adjust the threshold value accordingly; finally, conduct long-term charge-discharge cycle experiments on the same type of battery, record the actual capacity when the battery capacity decays to the point where it can no longer meet usage requirements, and determine the final capacity decay to scrap threshold after statistical analysis. For example, if the capacity decay to the scrap threshold is 80% of the rated capacity, and the predicted battery capacity obtained by the screening and training model is 90% of the rated capacity, with an average annual decay rate of 5% of the rated capacity, then the remaining service life of the lithium-ion battery is 2 years.

[0026] Understandably, by recording voltage signals through intermittent current interruption, the voltage characteristics before, during, and after current recovery are comprehensively extracted, avoiding the problem of single-feature analysis in traditional methods. Combined with the ampere-hour integration method, battery capacity is accurately calculated, and the calculation interval is adjusted according to current fluctuations to improve calculation accuracy. L1 regularization model training and feature selection optimize the model, reducing redundant feature interference and lowering prediction errors. The impact of temperature on test parameters is also considered, with test duration and interval adjusted according to temperature to adapt to different environments. Finally, multi-cell validation ensures the method's generalizability, enabling accurate prediction of the remaining lifespan of lithium-ion batteries and providing assurance for reliable battery applications.

[0027] In some embodiments of this application, when using an intermittent current interruption method to perform charge-discharge cycle tests on lithium-ion batteries and recording voltage signals during the test at preset time intervals, the method includes: Based on the preset constant current charging and discharging duration and the preset current interruption duration, the constant current charging and discharging operation and the current interruption operation are performed alternately, and the voltage signal during the constant current charging and discharging process and the current interruption process is continuously recorded at preset time intervals. When the ambient temperature is within the preset temperature range, the test will be performed according to the preset constant current charging and discharging duration, preset current interruption duration, and preset time interval. When the ambient temperature of the test environment is greater than the upper limit of the preset temperature range, the constant current charging and discharging time, the current interruption time and the recording time interval are shortened by a preset ratio. When the ambient temperature is lower than the lower limit of the preset temperature range, the constant current charging and discharging time, current interruption time, and recording time interval will be extended by a preset ratio.

[0028] Specifically, the preset constant current charge / discharge duration refers to the duration of each constant current charge or discharge operation during intermittent current interruption testing. It is obtained through pre-testing: charge / discharge experiments are conducted under different candidate durations, and the dynamic changes in the voltage signal are monitored in real time. When the voltage fluctuation amplitude stabilizes within 0.01V for three consecutive preset time intervals (i.e., the voltage change tends to stabilize), the shortest candidate duration at this point is taken as the preset value. The preset current interruption duration refers to the duration of each pause in charging / discharging and current interruption, obtained through pre-testing: voltage recovery curves are recorded under different candidate interruption durations, and the voltage recovery rate (the amount of voltage change per unit time) is calculated. When the recovery rate drops below 5% of the initial recovery rate (i.e., the voltage recovery tends to stabilize), this duration is taken as the preset value. The preset time interval... This refers to the time interval for recording voltage signals, obtained through preliminary testing: voltage data is collected at different candidate intervals, and the data integrity (such as whether the voltage abrupt change point during current switching is fully captured) and data redundancy (such as the proportion of repetitive feature data) are compared. The interval that can fully reflect the dynamic characteristics of voltage and has a redundant data ratio of less than 10% is selected as the preset value. The preset ratio is a proportional variable determined by statistically analyzing the linear correspondence between the temperature deviation value and the parameter adjustment amount based on the preliminary test data under different temperature deviation values. The value ranges from 5% to 20% and is positively correlated with the temperature deviation value. When the ambient temperature exceeds the preset range, the constant current charging and discharging duration, current interruption duration, and preset time interval are adjusted according to this ratio to adapt to the impact of temperature changes on battery performance and maintain the reliability and consistency of test data.

[0029] Specifically, assuming the preset temperature range is 25℃-35℃, if the ambient temperature is 38℃, exceeding the upper limit by 3℃, the temperature deviation is 3℃, corresponding to a preset ratio of 10% (the lower value in the ratio range is used because the deviation is small). Therefore, the original preset constant current charging / discharging time (e.g., 150s) and preset current interruption time (e.g., 10s) are both shortened by 10%, becoming 135s and 9s respectively. Simultaneously, the original preset time interval (e.g., 0.1s) is also shortened by 10% to 0.09s. If the ambient temperature is 20℃, below the lower limit by 5℃, the temperature deviation is 5℃, corresponding to a preset ratio of 18% (the higher value in the ratio range is used because the deviation is large). Therefore, the original preset constant current charging / discharging time and preset current interruption time are both extended by 18%, becoming 177s and 11.8s respectively. Simultaneously, the original preset time interval is extended by 18% to 0.118s. This demonstrates the positive correlation between the preset ratio and the temperature deviation value, and its specific application.

[0030] Specifically, refer to Figure 2 As shown, the characteristics of the voltage signal one day before the interruption ( Figure 2 The diagram shows the INV feature area and 101 voltage signal characteristics during interruption. Figure 2The figure shows the ICI characteristic region and the 100 voltage signal characteristics during recovery current. Figure 2 (This is denoted as the CC feature region in the diagram). The fixed capacity is 948.5mAh. Since one voltage signal is recorded every 0.1s, all 101 voltage signals during the 10s interruption are used as features. When the current is restored, there are a total of 1501 voltage signals in 150s. The feature extraction method is as follows: 50 features are extracted at equal intervals from 1 to 100; 35 features are extracted at equal intervals from 101 to 500; and 15 features are extracted at equal intervals from 501 to 1501.

[0031] Understandably, by adjusting test parameters (constant current charging / discharging duration, current interruption duration, and recording interval) in conjunction with ambient temperature, the impact of different temperature environments on the accuracy of test data is resolved. When the temperature is within the preset range, tests are performed according to standard parameters; when the temperature exceeds the limit, adjustments are made flexibly to ensure reliable voltage signals can still be acquired under complex temperature conditions. This improves the stability of the test data, thereby providing high-quality data support for subsequent feature extraction and model training, and enhancing the applicability of the method in different environmental scenarios.

[0032] In some embodiments of this application, extracting voltage signal features from the voltage signal includes: Based on the voltage signal characteristics before the interruption, one voltage signal corresponding to the last voltage signal recording time before the current interruption operation begins is selected as the voltage signal characteristics before the interruption. Based on the voltage signal characteristics during the interruption, all voltage signals recorded at preset time intervals during the entire current interruption duration are selected as the voltage signal characteristics during the interruption. Based on the voltage signal characteristics during the recovery current, the total duration of the recovery current stage is equally divided into a preset number of continuous time periods, and the same preset number of voltage signals are extracted at equal intervals within each time period. The voltage signals extracted from each time period together constitute the voltage signal characteristics during the recovery current.

[0033] Specifically, the "preset number" refers to the specific number of consecutive time periods into which the total duration of the recovery current phase is divided. This is obtained through prior analysis of the voltage signal variation patterns during the recovery current phase. The specific methods are as follows: collecting voltage curves of the recovery current process in different cycle stages, identifying the dynamic response characteristics of voltage from sudden changes to stability, counting the number of key change nodes, and determining the optimal number of time periods through multiple comparative experiments (testing the correlation between features and capacity under different number of divisions) in combination with the requirement that feature extraction needs to cover the complete dynamic process. The "set number" refers to the number of voltage signals extracted in each time period. This is obtained through the following specific methods: analyzing the temporal distribution density of voltage signals in each time period, testing the feature effectiveness corresponding to different extraction numbers (such as 3, 5, 7, etc.) (calculating the Pearson correlation coefficient and mutual information value between features and capacity), and selecting the number that can retain key dynamic information while avoiding data redundancy, while ensuring that the extraction number is the same in each time period to maintain the consistency of feature dimensions.

[0034] Specifically, the "equal interval method" means that within each continuous time period, voltage signals are extracted at the same time interval from the start time to the end time of that time period, that is, the time difference between two adjacent extraction points is exactly the same. For example, if the total duration of the current recovery stage is 60 seconds, it is divided into 3 time periods (each time period is 20 seconds), and 3 voltage signals are set to be extracted in each time period, then the extraction time points in each time period are the 5th second, 10th second, and 15th second of that time period (the interval between adjacent extraction points is 5 seconds). A total of 9 voltage signals are extracted in the 3 time periods, which together form the voltage signal characteristics during current recovery.

[0035] Understandably, separate extraction methods were developed for the voltage signals before, during, and after the interruption: the last signal was taken before the interruption, all signals were taken during the interruption, and the signals were extracted at equal intervals during the current recovery. This ensured the standardization and consistency of feature extraction, avoiding differences in feature quality caused by arbitrary extraction methods, and comprehensively captured the key information of the voltage signals at different stages. This ensured that the extracted voltage features could effectively reflect the battery state, providing a guarantee for building an accurate dataset and helping to improve the prediction performance of subsequent models.

[0036] In some embodiments of this application, calculating battery capacity using the ampere-hour integration method includes: When calculating battery capacity, instantaneous current is collected at the time points when the voltage signal is recorded in step S1, and then... - The battery capacity is obtained by dividing the instantaneous current into several small time intervals and accumulating the product of the instantaneous current and the small time intervals corresponding to the instantaneous current. The minute time interval corresponding to the instantaneous current is the time interval between the previous instantaneous current acquisition time point and the current instantaneous current acquisition time point; When the fluctuation amplitude of the instantaneous current during charging and discharging is less than the preset fluctuation threshold, the battery capacity is calculated using the ampere-hour integration method. When the fluctuation amplitude of the instantaneous current during charging and discharging is greater than or equal to the preset fluctuation threshold, the length of the tiny time interval is shortened by a preset reduction ratio before the battery capacity is calculated by the ampere-hour integration method.

[0037] Specifically, the battery capacity is calculated using the ampere-hour integration method according to the following formula: in For battery capacity, This refers to the start time of the charge / discharge cycle. This refers to the charging / discharging termination time. for Instantaneous current at a given moment; The essence of calculating battery capacity using the ampere-hour integration method is to analyze the instantaneous current. During the charge / discharge time interval [ , The continuous time interval is continuously integrated to accumulate the total capacity; however, in actual operation, it is impossible to collect and calculate the instantaneous current at every infinitesimal point in time. Therefore, it is necessary to divide the continuous time interval into multiple tiny time intervals, which is an essential engineering method to transform continuous integration into discrete calculation. Simultaneously, when the fluctuation amplitude of the instantaneous current during charging and discharging is less than a preset fluctuation threshold, the battery capacity is calculated using the ampere-hour integration method; when the fluctuation amplitude of the instantaneous current during charging and discharging is greater than or equal to the preset fluctuation threshold, the length of the tiny time interval is shortened by a preset reduction ratio before calculating the battery capacity using the ampere-hour integration method, ensuring that the battery capacity is calculated regardless of current fluctuations. The instantaneous current change within a small time interval is always within the preset error allowable range. At this time, the instantaneous current at any moment within the interval (such as the current at the start, middle or end of the interval) can be used to approximately represent the average current within the interval. Its product with the small time interval can be equivalent to the amount of electricity within the interval. By accumulating the product of all intervals, it is possible to achieve an accurate approximation of continuous integration while taking into account the computational efficiency in engineering. At the same time, the accuracy is improved by controlling the "preset fluctuation threshold" and "preset reduction ratio". Thus, a closed loop is formed in terms of principle, engineering implementation and error control, making the calculation method reasonable and feasible.

[0038] Specifically, the minute time interval is the time interval between two instantaneous current acquisition points; the instantaneous current refers to the real-time current value measured at the current acquisition moment corresponding to the voltage signal recording time point in step S1, which is directly acquired by a current sensor synchronized with the voltage acquisition within each preset time interval; the fluctuation amplitude of the instantaneous current refers to the maximum value of the absolute value of the instantaneous current difference between two adjacent current acquisition time points, which is obtained by calculating the current difference of all adjacent acquisition points and taking the maximum value; the preset fluctuation threshold is obtained by conducting multiple cycle tests on lithium-ion batteries under standard constant current charge and discharge conditions, collecting instantaneous current data at different cycle stages and under different temperature environments, statistically analyzing the current fluctuation distribution range under normal operating conditions, and combining it with capacity calculation. The allowable error range (e.g., setting the capacity calculation error to no more than 0.5%) is a critical value. When the current fluctuation is less than this value, the current under the existing time interval can approximately meet the calculation accuracy. The preset reduction ratio is determined by conducting comparative tests under different current fluctuation amplitudes (including cases exceeding the preset fluctuation threshold). That is, the capacity is calculated using the original time interval and time intervals with different reduction ratios for the same charging and discharging process. The calculation results are compared with the actual capacity measured by a high-precision fuel meter. The optimal time interval reduction ratio that can bring the capacity calculation error back to the allowable range under different fluctuation amplitudes is statistically analyzed. Then, the correspondence between fluctuation amplitude and reduction ratio is determined (e.g., for every 10% increase in fluctuation amplitude, the reduction ratio increases by 5%), and finally, the preset reduction ratio parameters are formed.

[0039] Understandably, by collecting instantaneous current at the time points recorded by voltage signals, calculating capacity by accumulating the data over small time intervals, and adjusting the length of these small time intervals according to the amplitude of current fluctuations, the problem of current fluctuations affecting the accuracy of capacity calculation is solved. When current fluctuations are small, calculations are performed at regular intervals; when fluctuations are large, the intervals are shortened. This ensures that battery capacity can be accurately calculated regardless of whether the current is stable or not, making the target value (capacity) when building the dataset more accurate. This provides reliable labeled data for subsequent model training and improves the basic accuracy of model predictions.

[0040] In some embodiments of this application, dividing the dataset into a training set and a test set includes: Obtain the total number of charge-discharge cycles for the lithium-ion battery. The total number of charge-discharge cycles is divided according to a preset training ratio; Multiply the total number of charge-discharge cycles by the preset training ratio to obtain the number of training charge-discharge cycles; The voltage signal features and the target value of several charge-discharge cycles in the dataset are used as the training set; The number of preset training ratios is the preset number of training sessions, and each preset training ratio is different. The voltage signal features and target values ​​of the remaining charge and discharge cycles after removing the current training set from the dataset are used as the test set corresponding to the current training set; The training set and its corresponding test set together form a complete dataset.

[0041] Specifically, the total number of charge-discharge cycles of a lithium-ion battery is obtained by first conducting a cycle life test on the lithium-ion battery: continuous charge-discharge cycle operation is performed using an intermittent current interruption method. Each completed charge-discharge cycle (i.e., one complete charge-discharge operation) is counted once. At the same time, the battery capacity is calculated periodically using the ampere-hour integration method during the cycle, and the capacity decay is monitored in real time. When the calculated battery capacity decays to a preset scrap threshold (usually 80% of the battery's rated capacity, which is determined in conjunction with the battery application scenario and industry standards), the cycle life test is stopped. The total number of complete charge-discharge cycles completed from the start to the end of the test is counted, and this number is the total number of charge-discharge cycles of the lithium-ion battery.

[0042] Specifically, the dataset is divided into training and testing sets: the first 20%, 40%, 60%, and 80% of the total battery cycles count; and the last 80%, 60%, 40%, and 20% of the total battery cycles count. A "cycle" refers to the unit of a lithium-ion battery completing one full charge-discharge cycle. Figure 3 As shown, machine learning models using the L1 regularized linear regression (Lasso) model are trained and tested on the above datasets respectively.

[0043] Understandably, by obtaining the total number of charge-discharge cycles, dividing the training set (first N% of cycles) according to different preset training ratios, and using the remaining cycles after removing the training set as the test set, while clarifying multiple different training ratio division methods, the logic and repeatability of the dataset division are ensured. At the same time, by trying multiple training ratios, the impact of different data volumes on the model training effect can be explored, providing the possibility of selecting the optimal training set ratio, thereby improving the flexibility and targeting of model training, and ensuring that the model can make full use of data information.

[0044] In some embodiments of this application, when dividing the dataset into a training set and a test set, the method further includes: When the total number of charge-discharge cycles is greater than the preset charge-discharge cycle threshold, the training set and the test set are divided according to the preset training ratio. When the total number of charge-discharge cycles is less than or equal to a preset charge-discharge cycle threshold, the percentage of the training set in the total number of charge-discharge cycles is increased by a preset increase ratio before the training set and the test set are divided.

[0045] Specifically, the preset training ratio is a proportion used to divide the training set (e.g., the top 20%, 40%, 60%, or 80% of the total charge-discharge cycles) to achieve gradient validation of the model's predictive performance under different data volumes. Through pre-testing, multiple different ratios (e.g., 10%, 20%,...90%) are selected to divide the training and test sets respectively. After training an L1-regularized linear regression model, the evaluation metrics (MAE, MSE, etc.) of the model under each ratio are compared to select the ratio that covers the data volume gradient and has stable model performance as the final preset training ratio. The preset charge-discharge cycle threshold is a critical value for the total charge-discharge cycles used to determine whether the training set ratio needs to be adjusted. Through statistical analysis, lithium-ion battery data with different total charge-discharge cycles (e.g., 50 cycles, 100 cycles, 200 cycles, etc.) are collected, the dataset is divided according to the preset training ratio, and the model is trained. The prediction accuracy of the model under different total cycles is compared to determine when the total cycle count falls below a certain value. The model accuracy drops significantly when the total number of charge-discharge cycles falls below the threshold. The preset boost ratio is the adjustment ratio (e.g., 10%-30%) used to increase the proportion of the training set when the total number of charge-discharge cycles is below the threshold. It is achieved through pre-testing adjustment methods, increasing the proportion of the training set by different boost rates (e.g., 5%, 10%...40%) for battery data with a total number of cycles below the threshold, retraining the model, and comparing the evaluation metrics of the adjusted model. The boost rate that can bring the model accuracy back to a qualified range is selected as the preset boost ratio. The preset training quantity is the number of preset training ratios (4 in this case). It is achieved through demand matching methods, based on the needs of model gradient verification (e.g., needing to verify the performance changes from small to large amounts of data), testing the impact of different numbers of preset training ratios (e.g., 2, 3, 4, 5) on the completeness of the verification results. It is determined that 4 ratios can fully cover the data volume gradient without reducing verification efficiency due to excessive quantity. Therefore, the preset training quantity is set to 4.

[0046] Understandably, by flexibly adjusting the proportion of the training set, the effectiveness of the test set is ensured when there is sufficient data, and the training set is ensured to have enough data to support model learning when there is insufficient data. This balances the rationality of dataset partitioning under different data volume scenarios and improves the adaptability of the method to battery data with different sample sizes.

[0047] In some embodiments of this application, when training an L1 regularized linear regression model using the training set to obtain a trained model, testing the trained model using the test set to obtain test data, calculating an evaluation index using the test data, and evaluating the predictive performance of the trained model using the evaluation index, the process includes: When the values ​​of each of the evaluation indicators are less than the corresponding preset indicator threshold, step S5 is executed. When the value of any of the evaluation indicators is greater than or equal to the corresponding preset indicator threshold, the regularization strength is adjusted according to the preset adjustment step size, and the L1 regularized linear regression model is retrained using the training set. The evaluation indicators are then retested and calculated using the test set until the values ​​of all the evaluation indicators are less than the corresponding preset indicator threshold.

[0048] Specifically, the objective function of the L1 regularized linear regression model is: in, is the true battery capacity value in the training set, and w is the weight vector of the voltage signal features in the training set. is the regularization strength, and n is the number of samples in the training set. It is the voltage signal feature vector of the i-th sample in the training set. It is the L1 norm. It is the weight value corresponding to the j-th voltage signal feature in the training set, and p is the total number of voltage signal features; The formula for calculating the mean absolute error is: The formula for calculating the mean square error is: The formula for calculating the root mean square error is: The formula for calculating the mean absolute percentage error is: Where m is the number of samples in the test set. It is the true battery capacity value of the k-th sample in the test set. It is the predicted battery capacity value of the k-th sample in the test set; This is the actual battery capacity, calculated using the ampere-hour integration method. Each sample contains the voltage signal features extracted before the interruption, the voltage signal features during the interruption, and the voltage signal features during the recovery current (as input features) within that cycle, as well as the battery capacity Q for that cycle calculated by the ampere-hour integration method (as the target value). Specifically, the training set samples are constructed based on lithium-ion battery charge-discharge cycle data. Specifically, the total number of charge-discharge cycles of the battery is obtained first, the number of training charge-discharge cycles is calculated according to a preset training ratio, and then relevant data from the first few charge-discharge cycles in the dataset are selected as samples of a single training set. Each sample contains two core contents: first, the voltage signal features extracted within the charge-discharge cycle (voltage signal features before interruption, voltage signal features during interruption, and voltage signal features when current is restored); second, the battery capacity (i.e., the target value) calculated according to the ampere-hour integration method formula within the charge-discharge cycle.

[0049] Specifically, the process of training an L1 regularized linear regression model using a training set to obtain the trained model is as follows: First, determine the input data of the training set (i.e., the voltage signal feature vector of each sample in the training set). Let i = 1, 2, 3, ..., n, where n is the number of samples in the training set. It contains p voltage signal features (corresponding to feature numbers j=1,2,3...,p) and an output label (i.e., the true battery capacity value corresponding to each sample). , The battery capacity Q is calculated using the ampere-hour integration method; then, an optimization algorithm for the L1 regularized linear regression model (such as coordinate descent) is called, with the objective function of the model as the optimization objective, and the weight values ​​of each voltage signal feature are updated iteratively. (In each iteration, the weights of other features are fixed, and only the weights of the current feature are adjusted to reduce the objective function value); then, an iteration termination condition is set (such as the difference between the objective function values ​​of two adjacent iterations being less than a preset convergence threshold, or the number of iterations reaching a preset maximum number). The iteration stops when the termination condition is met; finally, the weight vector w at this point is output (containing the final weight values ​​corresponding to each voltage signal feature). This yields the trained L1 regularized linear regression model (trained model), which can calculate the predicted battery capacity based on the input voltage signal feature vector.

[0050] Specifically, the preset threshold values ​​corresponding to Mean Absolute Error (MAE), Mean Square Error (MSE), Root Mean Square Error (RMSE), and Mean Absolute Percentage Error (MAPE) refer to the maximum allowable average absolute deviation of MAE, maximum square deviation of MSE, maximum root mean square deviation of RMSE, and maximum percentage deviation of MAPE in the scenario of predicting the remaining lifespan of lithium-ion batteries. These threshold values ​​are obtained through multiple preliminary experiments: selecting lithium-ion batteries from different manufacturers and using different cathode materials, completing charge-discharge tests, feature extraction, capacity calculation, and dataset construction according to the steps in claims 1-4, training with multiple rounds of L1 regularized linear regression models (covering different training set proportions and feature counts), recording the actual values ​​of MAE, MSE, RMSE, and MAPE for all effective training models, and combining these with the actual application's requirements for prediction accuracy (e.g., in industrial scenarios where capacity prediction error ≤ 5% corresponds to a MAPE value). The threshold for each indicator is determined through statistical analysis (taking the 90th percentile of the error distribution after removing outliers, or by back-calculating based on the allowable error range). The preset adjustment step size is a fixed increment / decrement used to adjust the regularization strength α of the L1 regularized linear regression model. It needs to be obtained through a sensitivity experiment of α and error indicators: set an initial value range for α (e.g., 0.001-10), gradually adjust α and train the model according to different candidate step sizes (e.g., 0.001, 0.01, 0.1), record the convergence speed and stability of error indicators (MAE, MSE, RMSE, MAPE) at each step size, and screen out the step size that can make the error indicators converge quickly to the preset threshold and avoid overfitting and has high computational efficiency (e.g., when α=0.001, the error decreases significantly, so the step size is set to 0.001; when α>1, the error changes slowly, so the step size can be set to 0.1), or determine the optimal step size through optimization methods such as grid search and random search.

[0051] Specifically, the model performance is judged based on the calculation results of each evaluation index. The smaller the value of each index, the smaller the deviation between the model prediction result and the actual battery capacity, and the better the model's prediction performance.

[0052] Specifically, the method for adjusting the regularization strength α according to the preset adjustment step size is as follows: First, determine the adjustment direction based on the comparison between the evaluation index and the preset threshold and the model fitting status—if any evaluation index (such as MAE, MSE, RMSE, MAPE) is greater than or equal to the corresponding preset threshold, and the model shows a significant deviation between the predicted value and the true value (such as high MAE and MSE), it indicates that the model is underfitting. At this time, decrease α by the preset adjustment step size (this step size is determined by testing the impact of different α values ​​on the model training effect in the early stage, and the value is usually 0.005-0.02) to weaken the regularization constraint and allow the model to fit more effectively. The model learns more fully the relationship between voltage signal characteristics and battery capacity. If the model performs well in terms of evaluation metrics on the training set but poorly in terms of metrics on the test set (overfitting occurs), the step size is increased by α according to the same preset value to enhance the regularization constraint and suppress the model's overfitting to the noise of the training data. After each adjustment of α, the L1 regularized linear regression model needs to be retrained with the original training set, and then the model is tested with the original test set and the evaluation metrics are recalculated. The above adjustment, training, and testing process is repeated until all evaluation metric values ​​are less than the corresponding preset thresholds, and the step size remains fixed during the adjustment process to ensure the stability and repeatability of the adjustment.

[0053] Specifically, such as Figure 6 As shown, the evaluation metrics are mean absolute error (MAE), mean square error (MSE), root mean square error (RMSE), and mean absolute percentage error (MAPE). The closer MAE, MSE, RMSE, and MAPE are to 0, the better the predictive performance for the future capacity of the battery.

[0054] Understandably, by setting a threshold for the evaluation metric, the problem of poor model prediction performance caused by inappropriate initial regularization strength is addressed. This mechanism can proactively optimize model parameters, ensuring that the trained model meets preset performance requirements and avoiding the impact of inappropriate model parameters on prediction results.

[0055] In some embodiments of this application, analyzing the weight values ​​of each voltage signal feature in the training model and filtering out voltage signal features with weight values ​​greater than or equal to a weight threshold includes: When the number of selected voltage signal features exceeds the upper limit of the preset key feature number range, the weight threshold is increased by a preset increment and the selection is repeated until the number of selected voltage signal features is within the preset key feature number range. When the number of selected voltage signal features is less than the lower limit of the preset key feature number range, the weight threshold is reduced by a preset reduction and the selection is repeated until the number of selected voltage signal features is within the preset key feature number range.

[0056] Specifically, the weight threshold is a critical weight value used to select voltage signal features that contribute significantly to battery capacity prediction. It is obtained through the following steps: First, extract the weight values ​​of all voltage signal features from the trained L1 regularized linear regression model. Statistics of these The distribution characteristics (such as maximum, minimum, median, and standard deviation) are analyzed. Multiple control experiments are conducted, using different weight values ​​as candidate thresholds to filter features and train models. The prediction accuracy (e.g., MAE, MSE) of the models under each candidate threshold is compared. A critical value that retains key prediction information while avoiding feature redundancy is selected as the weight threshold. The preset range of the number of key features is a reasonable interval (including upper and lower limits) for the number of key voltage signal features after screening. This is obtained through the following specific operations: Using the controlled variable method, L1 regularized linear regression models are trained with different numbers (e.g., 2-10) of features. The prediction accuracy and training efficiency (e.g., training time) of each model are compared, and features are discarded. Except for the intervals where insufficient features lead to insufficient accuracy (evaluation metrics exceeding thresholds) and excessive features lead to low efficiency, the interval of feature quantity that meets the accuracy standard and has high efficiency is retained as the preset key feature quantity range. The preset increase is the fixed adjustment range when increasing the weight threshold, and the preset decrease is the fixed adjustment range when decreasing the weight threshold. The two are obtained through the following specific operations: First, set the initial weight threshold, and then gradually adjust the threshold by different ranges (such as 0.01, 0.05, 0.1). Record the change in the number of selected features and the corresponding model accuracy after each adjustment. Select the adjustment range that can make the number of features stably converge to the preset key feature quantity range and the model accuracy fluctuate less, and use them as the preset increase and preset decrease, respectively.

[0057] Specifically, the trained L1 regularized linear regression (Lasso) model is analyzed, and features with large weights (w) are extracted and then trained. The objective function of the L1 regularized linear regression (Lasso) model is based on ordinary least squares (OLS) with the addition of an L1 regularization term.

[0058] Understandably, by setting a preset range for the number of key features, the weight threshold is increased when the number of selected features exceeds the upper limit and decreased when it falls below the lower limit, until the number of features meets the requirements. This solves the problem of too many features after selection (increasing model complexity and introducing redundant information) or too few features (losing key information). This rule ensures that the number of selected voltage features is moderate and critical, which simplifies the model structure, reduces training complexity, and retains effective features, laying a good foundation for subsequent model training and selection, and balancing model complexity and prediction performance.

[0059] In some embodiments of this application, the voltage signal features in the training set and the test set are replaced with the selected signal voltage features to obtain a selected training set and a selected test set. An L1 regularized linear regression model is trained using the selected training set to obtain a selected training model. The selected training model is tested using the selected test set to obtain selected test data. An evaluation index is calculated using the selected test data, and the predictive performance of the selected training model is evaluated using the evaluation index. The process includes: When the difference between each evaluation index of the selected training model and the evaluation index corresponding to the training model is less than the preset difference threshold, step S6 is executed. When the difference between any evaluation index of the selected training model and the evaluation index corresponding to the training model is greater than or equal to a preset difference threshold, the regularization strength of the selected training model is adjusted by a preset adjustment step size, and the L1 regularized linear regression model is retrained using the selected training set. The evaluation index is then tested and calculated again using the selected test set until the difference between each evaluation index of the selected training model and the evaluation index corresponding to the training model is less than the preset difference threshold.

[0060] Specifically, the preset difference threshold is a critical value used to determine whether the difference in prediction performance between the selected training model (the model trained based on key voltage signal features) and the original training model (the model trained based on all voltage signal features) is within an acceptable range. Its core function is to ensure that the selected training model, while simplifying the number of features and reducing computational complexity, does not significantly lose prediction accuracy (e.g., avoiding a significant increase in the mean absolute error of the selected model compared to the original model, leading to a prediction deviation of the battery's remaining lifespan exceeding the tolerance range for practical applications). The specific method for obtaining this threshold is as follows: First, collect multiple sets of charge-discharge cycle test data, voltage signal feature data, and corresponding model training (original model and selected model) results from different types of lithium-ion batteries (covering different manufacturers, different cathode materials, and different nominal capacities). First, a historical database is constructed. Second, based on the database statistics, the actual difference distribution of various evaluation indicators (mean absolute error, mean square error, etc.) between the screening model and the original model under different feature screening scenarios is determined to determine the reasonable fluctuation range of the difference. Then, combined with the actual application requirements for the prediction accuracy of battery remaining life (e.g., in industrial scenarios where the prediction error is required to be no more than 5%, the difference threshold is derived to be less than 10%-20% of the error, i.e., 0.5%-1%), and the threshold range is initially set. Finally, through cross-validation (randomly dividing historical data into multiple training and validation sets to test the stability of the model under different initial thresholds) and iterative optimization (adjusting the threshold according to the validation results to ensure that in more than 90% of the test scenarios, the screening model meets both the accuracy requirements and achieves feature simplification), the preset difference threshold is finally determined.

[0061] Specifically, the objective function and the formulas for each evaluation index of the selected training model and the L1 regularized linear regression model of the training model are the same. The difference lies in the distinction between the selected training set and the selected test set and the training set and test set.

[0062] Specifically, such as Figure 2 As shown in the figure, based on the information, the first feature of each voltage feature region (INV, ICI, and CC) is taken as the target that has a significant impact on the Lasso model. Training is then performed using these three voltage features. Figure 4 As shown, the performance evaluation is as follows Figure 6 As shown, the comparison revealed that, compared with 202 features, the three voltage features did not have a significant impact on the performance of battery capacity prediction.

[0063] Understandably, by comparing the difference in evaluation metrics between the selected model and the original trained model, the regularization strength of the selected model is adjusted and retrained when the difference exceeds the standard, until the difference reaches the standard, thus solving the problem of model performance degradation after feature selection. This mechanism ensures that the selected model, while simplifying its structure (reducing the number of features), still maintains predictive performance similar to the original model, achieving the effect of "reducing complexity without reducing performance." This improves model training efficiency and ensures the reliability of model prediction, providing performance assurance for ultimately using the selected model to predict remaining lifetime.

[0064] In some embodiments of this application, when repeating S1-S5 using data from multiple lithium-ion batteries to verify the feasibility of the method, the following steps are included: When the percentage of batteries whose values ​​are all less than the corresponding preset threshold values ​​of the selection training model is greater than the preset percentage threshold, the determination method is feasible. When the percentage of batteries whose values ​​of each evaluation index of the screening training model are less than the corresponding preset index threshold is less than or equal to the preset percentage threshold, the preset number of non-compliant lithium-ion batteries is adjusted until the percentage of batteries whose values ​​of each evaluation index of the screening training model are less than the corresponding preset index threshold is greater than the preset percentage threshold. The adjustment rule for the preset number is to increase by 1 or decrease by 1.

[0065] Specifically, the preset percentage threshold refers to the minimum proportion of lithium-ion batteries in the total number of lithium-ion batteries tested, where all evaluation indicators of the screening and training model are less than the corresponding preset threshold, when verifying the feasibility of the method in S6. Its core function is to measure the generalization ability and applicability of the prediction method across different battery types (different manufacturers, different cathode materials). Obtaining this threshold requires multiple steps: First, conduct preliminary experiments, selecting a small number of lithium-ion batteries covering mainstream market types (e.g., at least 3 manufacturers, 2 cathode materials, 2-3 battery samples per type), and complete the tests according to the S1-S5 process, recording whether each battery screening and training model meets the standard. Second, statistically analyze the actual percentage of batteries that meet the standard in the preliminary experiments, combining this with the method's expected generalization goal (e.g., aiming for applicability to most common battery types, rather than being limited to specific batteries). First, determine the threshold range; then, refer to the proportion requirements of similar verification steps in relevant technical literature or industry standards in the field of lithium-ion battery life prediction to ensure that the threshold setting conforms to the general logic of the industry; finally, through multiple rounds of cross-validation (such as adjusting the number of pre-experiment battery samples and fine-tuning the preset threshold of evaluation indicators), compare the rationality of the method feasibility judgment results under different proportion thresholds (if the threshold is too high, it is easy for the method to be misjudged as infeasible due to a few batteries not meeting the standard; if the threshold is too low, it will reduce the reliability of the method), and finally determine a specific value that takes into account both reliability and practicality (for example, determined to be 80% through the above methods, that is, when the proportion of qualified batteries exceeds 80%, the method is deemed feasible). R² measures the degree of fit between the capacity value predicted by the model and the actual capacity value of the battery. When R² is closer to 1, the MSE is closer to 0, indicating that the accuracy of the prediction model is better, that is, the prediction performance is better.

[0066] Specifically, training tests and evaluations were conducted on 18650 lithium-ion batteries from three different manufacturers using different cathode materials, such as... Figure 5 As shown, the Lasso method accurately predicts the capacity of subsequent batteries during the use of these three batteries by using three voltage characteristics. This demonstrates the feasibility of the Lasso method.

[0067] Understandably, by setting a preset percentage threshold, the method is deemed feasible when the percentage of compliant batteries exceeds the threshold. If the percentage fails to meet the threshold, the preset number of recovery current stages is adjusted (increase or decrease by 1) until the target is met. This solves the problem of questionable feasibility in multi-battery verification due to some batteries failing to meet the target. This measure ensures the generalizability of the method through multi-battery verification and optimizes parameters for non-compliant batteries, improving the method's adaptability to different battery types. Ultimately, it ensures that the method can stably and reliably predict the remaining lifespan of lithium-ion batteries in practical applications.

[0068] It should be noted that: Numerous specific details are set forth in the specification provided herein. However, it will be understood that embodiments of this application may be practiced without these specific details. In some instances, well-known structures and techniques have not been shown in detail so as not to obscure the understanding of this specification.

[0069] Furthermore, those skilled in the art will understand that although some embodiments described herein include certain features included in other embodiments but not others, combinations of features from different embodiments are meant to be within the scope of this application and form different embodiments.

[0070] The above description is merely a preferred embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for predicting the remaining lifespan of a lithium-ion battery, characterized in that, include: S1, The lithium-ion battery is subjected to charge-discharge cycle test using an intermittent current interruption method, and the voltage signal during the test is recorded at preset time intervals; S2, after each charge-discharge cycle reaches a fixed capacity, extract voltage signal features from the voltage signal. The voltage signal features include: voltage signal features before interruption, voltage signal features during interruption, and voltage signal features when current is restored. S3, calculate the battery capacity using the ampere-hour integration method, and construct a dataset of the voltage signal characteristics and target values, wherein the target value is the calculated battery capacity; S4. Divide the dataset into a training set and a test set. Train an L1 regularized linear regression model using the training set to obtain a training model. Test the training model using the test set to obtain test data. Calculate an evaluation index using the test data and evaluate the predictive performance of the training model using the evaluation index. The evaluation index includes: mean absolute error, mean square error, root mean square error, and mean absolute percentage error. S5. Analyze the weight values ​​of each voltage signal feature in the training model, and select the selected voltage signal features whose weight values ​​are greater than or equal to the weight threshold. Replace the voltage signal features in the training set and the test set with the selected voltage signal features to obtain a selected training set and a selected test set. Train an L1 regularized linear regression model using the selected training set to obtain a selected training model. Test the selected training model using the selected test set to obtain selected test data. Calculate the evaluation index using the selected test data and evaluate the predictive performance of the selected training model using the evaluation index. S6. Repeat S1-S5 using data from multiple lithium-ion batteries to verify the feasibility of the method. When the method is feasible, use the screening and training model to obtain the predicted battery capacity, and use the predicted battery capacity, the rated capacity of the lithium-ion battery, the lithium-ion battery degradation rate, and the capacity degradation to the scrap threshold to calculate the remaining service life of the lithium-ion battery.

2. The method for predicting the remaining lifespan of a lithium-ion battery according to claim 1, characterized in that, When conducting charge-discharge cycle tests on lithium-ion batteries using an intermittent current interruption method and recording the voltage signals during the test at preset time intervals, the following steps are included: Based on the preset constant current charging and discharging duration and the preset current interruption duration, the constant current charging and discharging operation and the current interruption operation are performed alternately, and the voltage signal during the constant current charging and discharging process and the current interruption process is continuously recorded at preset time intervals. When the ambient temperature is within the preset temperature range, the test will be performed according to the preset constant current charging and discharging duration, preset current interruption duration, and preset time interval. When the ambient temperature of the test environment is greater than the upper limit of the preset temperature range, the constant current charging and discharging time, the current interruption time and the recording time interval are shortened by a preset ratio. When the ambient temperature is lower than the lower limit of the preset temperature range, the constant current charging and discharging time, current interruption time, and recording time interval will be extended by a preset ratio.

3. The method for predicting the remaining lifespan of a lithium-ion battery according to claim 2, characterized in that, Extracting voltage signal features from the voltage signal includes: Based on the voltage signal characteristics before the interruption, one voltage signal corresponding to the last voltage signal recording time before the current interruption operation begins is selected as the voltage signal characteristics before the interruption. Based on the voltage signal characteristics during the interruption, all voltage signals recorded at preset time intervals during the entire current interruption duration are selected as the voltage signal characteristics during the interruption. Based on the voltage signal characteristics during the recovery current, the total duration of the recovery current stage is equally divided into a preset number of continuous time periods, and the same preset number of voltage signals are extracted at equal intervals within each time period. The voltage signals extracted from each time period together constitute the voltage signal characteristics during the recovery current.

4. The method for predicting the remaining lifespan of a lithium-ion battery according to claim 3, characterized in that, When calculating battery capacity using the ampere-hour integration method, the following steps are included: When calculating battery capacity, instantaneous current is collected at the time points when the voltage signal is recorded in step S1, and then... - The battery capacity is obtained by dividing the instantaneous current into several small time intervals and accumulating the product of the instantaneous current and the small time intervals corresponding to the instantaneous current. The minute time interval corresponding to the instantaneous current is the time interval between the previous instantaneous current acquisition time point and the current instantaneous current acquisition time point; When the fluctuation amplitude of the instantaneous current during charging and discharging is less than the preset fluctuation threshold, the battery capacity is calculated using the ampere-hour integration method. When the fluctuation amplitude of the instantaneous current during charging and discharging is greater than or equal to the preset fluctuation threshold, the length of the tiny time interval is shortened by a preset reduction ratio before the battery capacity is calculated by the ampere-hour integration method.

5. The method for predicting the remaining lifespan of a lithium-ion battery according to claim 4, characterized in that, When dividing the dataset into training and test sets, the following is included: Obtain the total number of charge-discharge cycles for the lithium-ion battery. The total number of charge-discharge cycles is divided according to a preset training ratio; Multiply the total number of charge-discharge cycles by the preset training ratio to obtain the number of training charge-discharge cycles; The voltage signal features and the target value of several charge-discharge cycles in the dataset are used as the training set; The number of preset training ratios is the preset number of training sessions, and each preset training ratio is different. The voltage signal features and target values ​​of the remaining charge and discharge cycles after removing the current training set from the dataset are used as the test set corresponding to the current training set; The training set and its corresponding test set together form a complete dataset.

6. The method for predicting the remaining lifespan of a lithium-ion battery according to claim 5, characterized in that, When dividing the dataset into training and test sets, the following is also included: When the total number of charge-discharge cycles is greater than the preset charge-discharge cycle threshold, the training set and the test set are divided according to the preset training ratio. When the total number of charge-discharge cycles is less than or equal to a preset charge-discharge cycle threshold, the percentage of the training set in the total number of charge-discharge cycles is increased by a preset increase ratio before the training set and the test set are divided.

7. The method for predicting the remaining lifespan of a lithium-ion battery according to claim 6, characterized in that, The process of training an L1-regularized linear regression model using the training set to obtain a training model, testing the training model using the test set to obtain test data, calculating an evaluation index using the test data, and evaluating the predictive performance of the training model using the evaluation index includes: When the values ​​of each of the evaluation indicators are less than the corresponding preset indicator threshold, step S5 is executed. When the value of any of the evaluation indicators is greater than or equal to the corresponding preset indicator threshold, the regularization strength is adjusted according to the preset adjustment step size, and the L1 regularized linear regression model is retrained using the training set. The evaluation indicators are then retested and calculated using the test set until the values ​​of all the evaluation indicators are less than the corresponding preset indicator threshold.

8. The method for predicting the remaining lifespan of a lithium-ion battery according to claim 7, characterized in that, Analyzing the weight values ​​of each voltage signal feature in the training model and selecting voltage signal features with weight values ​​greater than or equal to a weight threshold includes: When the number of selected voltage signal features exceeds the upper limit of the preset key feature number range, the weight threshold is increased by a preset increment and the selection is repeated until the number of selected voltage signal features is within the preset key feature number range. When the number of selected voltage signal features is less than the lower limit of the preset key feature number range, the weight threshold is reduced by a preset reduction and the selection is repeated until the number of selected voltage signal features is within the preset key feature number range.

9. A method for predicting the remaining lifespan of a lithium-ion battery according to claim 8, characterized in that, The voltage signal features in the training set and the test set are replaced with the selected signal voltage features to obtain a selected training set and a selected test set. An L1 regularized linear regression model is trained using the selected training set to obtain a selected training model. The selected training model is tested using the selected test set to obtain selected test data. An evaluation index is calculated using the selected test data, and the predictive performance of the selected training model is evaluated using the evaluation index, including: When the difference between each evaluation index of the selected training model and the evaluation index corresponding to the training model is less than the preset difference threshold, step S6 is executed. When the difference between any evaluation index of the selected training model and the evaluation index corresponding to the training model is greater than or equal to a preset difference threshold, the regularization strength of the selected training model is adjusted by a preset adjustment step size, and the L1 regularized linear regression model is retrained using the selected training set. The evaluation index is then tested and calculated again using the selected test set until the difference between each evaluation index of the selected training model and the evaluation index corresponding to the training model is less than the preset difference threshold.

10. A method for predicting the remaining lifespan of a lithium-ion battery according to claim 9, characterized in that, When repeating S1-S5 using data from multiple lithium-ion batteries and verifying the feasibility of the method, the following steps are included: When the percentage of batteries whose values ​​are all less than the corresponding preset threshold values ​​of the selection training model is greater than the preset percentage threshold, the determination method is feasible. When the percentage of batteries whose values ​​of each evaluation index of the screening training model are less than the corresponding preset index threshold is less than or equal to the preset percentage threshold, the preset number of non-compliant lithium-ion batteries is adjusted until the percentage of batteries whose values ​​of each evaluation index of the screening training model are less than the corresponding preset index threshold is greater than the preset percentage threshold. The adjustment rule for the preset number is to increase by 1 or decrease by 1.

Images