A battery health status estimation method and system based on time series characteristics

Abstract

The present invention provides a method and system for estimating the state of health of a battery based on timing characteristics in the field of battery detection technology. The method includes the following steps: step S10, obtaining a large amount of battery charging data and the corresponding SOH true value; step S20, preprocessing each of the charging data; step S30, using a timing feature extraction tool to extract features from each of the preprocessed charging data to obtain timing features; step S40, sorting and screening each of the timing features by importance to obtain a feature data set; step S50, creating a health state estimation model, using the feature data set and the SOH true value to train the health state estimation model, and using the trained health state estimation model to estimate the health state of the battery. The advantages of the present invention are: greatly improving the accuracy of battery health state estimation.

Description

Technical Field

[0001] The present invention relates to the technical field of battery detection, and in particular to a method and system for estimating a battery health state based on time series characteristics. Background Art

[0002] The state of health (SOH) of a battery is defined as the percentage between the current available capacity of the battery and its initial capacity. During the charge and discharge cycle, some irreversible chemical reactions will occur inside the battery, which will lead to battery aging and inevitably lead to some safety hazards. Therefore, the health of the battery is a key indicator.

[0003] Traditionally, there are three methods for estimating the health status of batteries: 1. Estimation method based on electrochemical mechanism model. This method uses many model parameters and usually requires intrusion into the battery to understand the internal reaction mechanism. It is also easily disturbed by the external environment, resulting in poor accuracy. 2. Estimation method based on equivalent circuit model. This method is an estimation of the physical meaning, but its generalization is too poor, resulting in poor accuracy. 3. Estimation method based on big data and artificial intelligence. This method can transform the problem of estimating the health status of batteries into a regression problem of multivariate time series input, and is particularly critical for feature extraction of multivariate time series such as voltage and current. However, traditionally, feature extraction relies on manual labor, and the features of different batteries or batteries in different health states cannot be accurately extracted manually, or even cannot be obtained or calculated, so the accuracy of battery health status estimation cannot be guaranteed.

[0004] Therefore, how to provide a battery health state estimation method and system based on time series characteristics to improve the accuracy of battery health state estimation has become a technical problem that needs to be solved urgently. Summary of the invention

[0005] The technical problem to be solved by the present invention is to provide a method and system for estimating the health status of a battery based on time series characteristics, so as to improve the accuracy of the estimation of the health status of the battery.

[0006] In a first aspect, the present invention provides a method for estimating a battery health state based on time series characteristics, comprising the following steps:

[0007] Step S10, obtaining a large amount of battery charging data and the corresponding SOH real value;

[0008] Step S20, pre-processing each of the charging data;

[0009] Step S30, using a time series feature extraction tool to extract features from each of the pre-processed charging data to obtain time series features;

[0010] Step S40: sorting and screening the time series features by importance to obtain a feature data set;

[0011] Step S50: create a health state estimation model, use the feature data set and the true value of SOH to train the health state estimation model, and use the trained health state estimation model to estimate the health state of the battery.

[0012] Furthermore, in the step S10, the charging data is voltage timing data and current timing data of a battery charging from a preset first SOC to a preset second SOC under several battery types, models, test temperatures, and charging rates.

[0013] Furthermore, the step S20 is specifically as follows:

[0014] A deviation ratio threshold is set, and data in the charging data whose deviation from the mean exceeds the deviation ratio threshold is eliminated, and missing data is filled based on time adjacent values;

[0015] In the step S30, the time series feature extraction tool is tsfresh.

[0016] Furthermore, the step S40 specifically includes:

[0017] Step S41, after arranging the time series features in random order, merge them with the initial time series features;

[0018] Step S42, calculating the importance of each of the time series features by using a random forest algorithm, and determining whether the number of repetitions is greater than a preset threshold value, if so, proceeding to step S43; if not, proceeding to step S41;

[0019] Step S43: construct a feature data set based on the time series features whose importance has changed.

[0020] Furthermore, in the step S50, the health status estimation model is created based on the LightGBM algorithm.

[0021] In a second aspect, the present invention provides a battery health status estimation system based on time series characteristics, comprising the following modules:

[0022] Data acquisition module, used to obtain a large amount of battery charging data and the corresponding SOH real value;

[0023] A data preprocessing module, used for preprocessing each of the charging data;

[0024] A feature extraction module, used to extract features from each of the pre-processed charging data using a time series feature extraction tool to obtain a time series feature;

[0025] A feature data set construction module is used to sort and filter the time series features by importance to obtain a feature data set;

[0026] The health state estimation module is used to create a health state estimation model, train the health state estimation model using the feature data set and the true value of SOH, and use the trained health state estimation model to estimate the health state of the battery.

[0027] Furthermore, in the data acquisition module, the charging data is voltage timing data and current timing data of a battery charging from a preset first SOC to a preset second SOC under several battery types, models, test temperatures, and charging rates.

[0028] Furthermore, the data preprocessing module is specifically used for:

[0029] A deviation ratio threshold is set, and data whose deviation from the mean value exceeds the deviation ratio threshold is eliminated from the charging data, and missing data is filled based on time adjacent values;

[0030] In the feature extraction module, the time series feature extraction tool is tsfresh.

[0031] Furthermore, the feature data set construction module specifically includes:

[0032] A feature reorganization unit, used for arranging the time series features in random order and merging them with the initial time series features;

[0033] An importance calculation unit, used to calculate the importance of each of the time series features through a random forest algorithm, and determine whether the number of repetitions is greater than a preset threshold value. If so, enter the feature screening unit; if not, enter the feature recombination unit;

[0034] The feature screening unit is used to construct a feature data set based on the time series features with changed importance.

[0035] Furthermore, in the health status estimation module, the health status estimation model is created based on the LightGBM algorithm.

[0036] The advantages of the present invention are:

[0037] The timing feature extraction tool is used to extract features from the preprocessed charging data to obtain timing features, and the timing features are sorted and screened by importance to obtain a feature data set. The feature data set and the true SOH value are then used to train the health state estimation model. It is best to use the trained health state estimation model to estimate the health state of the battery, that is, to automatically extract timing features from the charging data using a machine learning method. Compared with traditional manual feature extraction, the richness and scale of feature extraction are better, so that the health state estimation model has better generalization, and does not need to invade the battery, and is not easily affected by the external environment. Sorting and screening the importance of timing features can not only improve the efficiency of the health state estimation model prediction, but also improve the robustness, and ultimately greatly improve the accuracy of the battery health state estimation. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] The present invention will be further described below in conjunction with embodiments with reference to the accompanying drawings.

[0039] Figure 1 It is a flow chart of a battery health status estimation method based on time series characteristics of the present invention.

[0040] Figure 2 It is a structural schematic diagram of a battery health status estimation system based on time series characteristics of the present invention. DETAILED DESCRIPTION

[0041] The technical solution in the embodiments of the present application has the following overall idea: automatically extracting timing features from charging data through machine learning methods instead of manually extracting features, thereby improving the richness and scale of feature extraction, allowing the health status estimation model to have better generalization, without the need to invade the interior of the battery, and not easily affected by external environmental interference; by sorting and screening the timing features by importance to improve the efficiency and robustness of the health status estimation model prediction, thereby improving the accuracy of battery health status estimation.

[0042] Please refer to Figure 1 to Figure 2 As shown, a preferred embodiment of a battery health status estimation method based on time series characteristics of the present invention includes the following steps:

[0043] Step S10, obtaining a large amount of battery charging data and the corresponding SOH real value;

[0044] Step S20, pre-processing each of the charging data;

[0045] Step S30, using a time series feature extraction tool to automatically extract features from each of the pre-processed charging data to obtain a time series feature;

[0046] Step S40: sorting and screening the time series features by importance to obtain a feature data set;

[0047] Step S50: create a health state estimation model, use the feature data set and the true value of SOH to train the health state estimation model, and use the trained health state estimation model to estimate the health state of the battery.

[0048] In step S10, the charging data is voltage timing data and current timing data of the battery charging from a preset first SOC to a preset second SOC under several battery types, models, test temperatures, and charging rates; the value of the first SOC is 0, and the value of the second SOC is 100.

[0049] Since the discharge condition of the battery in actual application is often more complicated than the charging condition, and the charging data of the charging condition is also easier to obtain (for example, obtaining the charging condition data when charging a new energy vehicle), the present invention performs health status detection based on the charging condition of the battery.

[0050] The actual SOH value is the discharge capacity of the battery after the battery has been left at room temperature for a period of time, charged to a full charge state using constant current and constant voltage, and then discharged to a cut-off voltage using constant current. In actual measurement, the actual SOH value is generally calibrated after a fixed number of cycles. Therefore, it is necessary to perform curve fitting based on the discrete actual SOH value to obtain the actual SOH value under each cycle. The double exponential empirical capacity decay model can be used for calculation:

[0051] y=a·exp(b·k)+c·exp(d·k);

[0052] Where a, b, c, and d represent fitting parameters; k represents the number of cycles; and y represents the true value of SOH.

[0053] The network loss function uses the square root error between the true SOH value and the predicted SOH value, and the formula is as follows:

[0054]

[0055] Among them, y prodict represents the predicted value of SOH; y truth represents the true value of SOH; w i represents the weight of different charging cycles; m represents the total number of charging cycles; n represents the fitting parameter.

[0056] The step S20 is specifically as follows:

[0057] A deviation ratio threshold is set, and data in the charging data whose deviation from the mean exceeds the deviation ratio threshold is eliminated, and missing data is filled based on time adjacent values;

[0058] In step S30, the time series feature extraction tool is tsfresh; tsfresh is a Python time series data feature extraction module that can automatically calculate a large number of time series features, including a variety of statistical and transformation-based time series extraction methods, such as absolute energy value, absolute sum of first-order differences, aggregated statistical characteristics of autocorrelation coefficients of various orders, approximate entropy, autoregressive coefficient, ADF test, l ag-order autocorrelation, etc.

[0059] The step S40 specifically includes:

[0060] Step S41, after arranging the time series features in random order, merge them with the initial time series features;

[0061] Step S42, calculating the importance of each of the time series features by using a random forest algorithm, and determining whether the number of repetitions is greater than a preset number threshold, if so, proceeding to step S43; if not, proceeding to step S41; the number threshold is preferably 100 times;

[0062] Step S43: construct a feature data set based on the time series features whose importance has changed.

[0063] That is, the extracted time series features are sorted and screened by importance through the Boruta method; since the number of output tsfresh time series features is large and the amount of calculation is large, in order to improve the algorithm operation efficiency and avoid extracting irrelevant features, it is necessary to sort the calculated time series features by importance. The general method uses the Pearson correlation coefficient to sort the feature importance. The closer the obtained correlation coefficient is to 1, the higher the importance; however, the feature importance sorting based on the Pearson correlation coefficient can only calculate the correlation of a single feature, and cannot mine the correlation between multiple features, and each feature needs to be calculated separately, which has calculation redundancy. Therefore, the present invention adopts a feature importance selection strategy based on the Boruta method.

[0064] The goal of Boruta is to select all feature sets related to the dependent variable, rather than selecting a feature set that minimizes the model cost function for a specific model. The significance of the Boruta algorithm is that it can more comprehensively understand the influencing factors of the dependent variable, so as to perform feature selection better and more efficiently.

[0065] In the step S50, the health status estimation model is created based on the LightGBM algorithm; the LightGBM algorithm is a framework for implementing the GBDT algorithm, supports efficient parallel training, and has faster training speed, lower memory consumption, better accuracy, and supports distributed processing to quickly process massive data.

[0066] A preferred embodiment of a battery health status estimation system based on time series characteristics of the present invention includes the following modules:

[0067] The data acquisition module is used to obtain a large amount of battery charging data and the corresponding SOH real value;

[0068] A data preprocessing module, used for preprocessing each of the charging data;

[0069] A feature extraction module, used to automatically extract features from each of the pre-processed charging data using a time series feature extraction tool to obtain a time series feature;

[0070] A feature data set construction module is used to sort and filter the time series features by importance to obtain a feature data set;

[0071] The health state estimation module is used to create a health state estimation model, train the health state estimation model using the feature data set and the true value of SOH, and use the trained health state estimation model to estimate the health state of the battery.

[0072] In the data acquisition module, the charging data is voltage timing data and current timing data of the battery charging from a preset first SOC to a preset second SOC under several battery types, models, test temperatures, and charging rates; the value of the first SOC is 0, and the value of the second SOC is 100.

[0073] Since the discharge condition of the battery in actual application is often more complicated than the charging condition, and the charging data of the charging condition is also easier to obtain (for example, obtaining the charging condition data when charging a new energy vehicle), the present invention performs health status detection based on the charging condition of the battery.

[0074] The actual SOH value is the discharge capacity of the battery after the battery has been left at room temperature for a period of time, charged to a full charge state using constant current and constant voltage, and then discharged to a cut-off voltage using constant current. In actual measurement, the actual SOH value is generally calibrated after a fixed number of cycles. Therefore, it is necessary to perform curve fitting based on the discrete actual SOH value to obtain the actual SOH value under each cycle. The double exponential empirical capacity decay model can be used for calculation:

[0075] y=a·exp(b·k)+c·exp(d·k);

[0076] Where a, b, c, and d represent fitting parameters; k represents the number of cycles; and y represents the true value of SOH.

[0077] The network loss function uses the square root error between the true SOH value and the predicted SOH value, and the formula is as follows:

[0078]

[0079] Among them, y prodict represents the predicted value of SOH; y truth represents the true value of SOH; w i represents the weight of different charging cycles; m represents the total number of charging cycles; n represents the fitting parameter.

[0080] The data preprocessing module is specifically used for:

[0081] A deviation ratio threshold is set, and data in the charging data whose deviation from the mean exceeds the deviation ratio threshold is eliminated, and missing data is filled based on time adjacent values;

[0082] In the feature extraction module, the time series feature extraction tool is tsfresh; tsfresh is a Python time series data feature extraction module that can automatically calculate a large number of time series features, including a variety of statistical and transformation-based time series extraction methods, such as absolute energy value, absolute sum of first-order differences, aggregated statistical characteristics of autocorrelation coefficients of various orders, approximate entropy, autoregressive coefficient, ADF test, l ag-order autocorrelation, etc.

[0083] The feature data set construction module specifically includes:

[0084] A feature reorganization unit, used for arranging the time series features in random order and then merging them with the initial time series features;

[0085] An importance calculation unit is used to calculate the importance of each of the time series features through a random forest algorithm, and determine whether the number of repetitions is greater than a preset number threshold. If so, it enters a feature screening unit; if not, it enters a feature recombination unit; the number threshold is preferably 100 times;

[0086] The feature screening unit is used to construct a feature data set based on the time series features with changed importance.

[0087] That is, the extracted time series features are sorted and screened by importance through the Boruta method; since the number of output tsfresh time series features is large and the amount of calculation is large, in order to improve the algorithm operation efficiency and avoid extracting irrelevant features, it is necessary to sort the calculated time series features by importance. The general method uses the Pearson correlation coefficient to sort the feature importance. The closer the obtained correlation coefficient is to 1, the higher the importance; however, the feature importance sorting based on the Pearson correlation coefficient can only calculate the correlation of a single feature, and cannot mine the correlation between multiple features, and each feature needs to be calculated separately, which has calculation redundancy. Therefore, the present invention adopts a feature importance selection strategy based on the Boruta method.

[0088] The goal of Boruta is to select all feature sets related to the dependent variable, rather than selecting a feature set that can minimize the model cost function for a specific model. The significance of the Boruta algorithm is that it can more comprehensively understand the influencing factors of the dependent variable, so as to perform feature selection better and more efficiently.

[0089] In the health status estimation module, the health status estimation model is created based on the LightGBM algorithm; the LightGBM algorithm is a framework for implementing the GBDT algorithm, supports efficient parallel training, and has faster training speed, lower memory consumption, better accuracy, and supports distributed processing of massive data.

[0090] In summary, the advantages of the present invention are:

[0091] The timing feature extraction tool is used to extract features from the preprocessed charging data to obtain timing features, and the timing features are sorted and screened by importance to obtain a feature data set. The feature data set and the true SOH value are then used to train the health state estimation model. It is best to use the trained health state estimation model to estimate the health state of the battery, that is, to automatically extract timing features from the charging data using a machine learning method. Compared with traditional manual feature extraction, the richness and scale of feature extraction are better, so that the health state estimation model has better generalization, and does not need to invade the battery, and is not easily affected by the external environment. Sorting and screening the importance of timing features can not only improve the efficiency of the health state estimation model prediction, but also improve the robustness, and ultimately greatly improve the accuracy of the battery health state estimation.

[0092] Although the specific implementation modes of the present invention are described above, those skilled in the art should understand that the specific implementation modes described are only illustrative and are not intended to limit the scope of the present invention. Equivalent modifications and changes made by those skilled in the art in accordance with the spirit of the present invention should be included in the scope of protection of the claims of the present invention.

Claims

1. A battery health status estimation method based on time series characteristics, Features: The steps include: Step S10, obtaining a large amount of battery charging data and the corresponding SOH real value; Step S20, pre-processing each of the charging data; Step S30, using a time series feature extraction tool to extract features from each of the pre-processed charging data to obtain time series features; Step S40: sorting and screening the time series features by importance to obtain a feature data set; Step S50, creating a health state estimation model, using the feature data set and the true value of SOH to train the health state estimation model, and using the trained health state estimation model to estimate the health state of the battery; The step S40 specifically includes: Step S41, after arranging the time series features in random order, merge them with the initial time series features; Step S42, calculating the importance of each of the time series features by using a random forest algorithm, and determining whether the number of repetitions is greater than a preset threshold value, if so, proceeding to step S43; if not, proceeding to step S41; Step S43: construct a feature data set based on the time series features whose importance has changed.

2. A battery health status estimation method based on time series characteristics as claimed in claim 1, Features: In the step S10, the charging data is voltage timing data and current timing data of a battery charging from a preset first SOC to a preset second SOC under several battery types, models, test temperatures, and charging rates.

3. A battery health status estimation method based on time series characteristics as claimed in claim 1, Features: The step S20 is specifically as follows: A deviation ratio threshold is set, and data in the charging data whose deviation from the mean exceeds the deviation ratio threshold is eliminated, and missing data is filled based on time adjacent values; In the step S30, the time series feature extraction tool is tsfresh.

4. A method for estimating a battery health state based on time series characteristics as claimed in claim 1, Features: In step S50, the health status estimation model is created based on the LightGBM algorithm.

5. A battery health status estimation system based on time series characteristics, Features: Includes the following modules: Data acquisition module, used to obtain a large amount of battery charging data and the corresponding SOH real value; A data preprocessing module, used for preprocessing each of the charging data; A feature extraction module, used to extract features from each of the pre-processed charging data using a time series feature extraction tool to obtain a time series feature; A feature data set construction module is used to sort and filter the time series features by importance to obtain a feature data set; A health state estimation module is used to create a health state estimation model, train the health state estimation model using the feature data set and the true value of SOH, and estimate the health state of the battery using the trained health state estimation model; The feature data set construction module specifically includes: A feature reorganization unit, used for arranging the time series features in random order and merging them with the initial time series features; An importance calculation unit, used to calculate the importance of each of the time series features through a random forest algorithm, and determine whether the number of repetitions is greater than a preset threshold value. If so, enter the feature screening unit; if not, enter the feature recombination unit; The feature screening unit is used to construct a feature data set based on the time series features with changed importance.

6. A battery health status estimation system based on time series characteristics as claimed in claim 5, Features: In the data acquisition module, the charging data are voltage timing data and current timing data of a battery charging from a preset first SOC to a preset second SOC under several battery types, models, test temperatures, and charging rates.

7. A battery health status estimation system based on time series characteristics as claimed in claim 5, Features: The data preprocessing module is specifically used for: A deviation ratio threshold is set, and data whose deviation from the mean value exceeds the deviation ratio threshold is eliminated from the charging data, and missing data is filled based on time adjacent values; In the feature extraction module, the time series feature extraction tool is tsfresh.

8. A battery health status estimation system based on time series characteristics as claimed in claim 5, Features: In the health status estimation module, the health status estimation model is created based on the LightGBM algorithm.

Images