Method for estimating a remaining useful life of a battery, non-transitory computer-readable medium, and electronic estimation device
The method improves battery lifespan prediction accuracy by preprocessing and selecting features from charge-discharge cycles using a machine learning model, addressing inaccuracies and reducing test cycles.
Patent Information
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- TOTALENERGIES EP BRASIL LTDA
- Filing Date
- 2025-12-26
- Publication Date
- 2026-07-02
AI Technical Summary
Existing methods for predicting the remaining service life of batteries suffer from inaccuracies due to variations among batteries from the same manufacturing batch and require a large number of test cycles, making them impractical.
A method using a machine learning model that processes selected features from battery charge-discharge cycles, applying outlier detection and data smoothing treatments, and selects features based on correlation to improve prediction accuracy and reduce the number of required test cycles.
The method achieves more accurate predictions of battery lifespan with fewer cycles by preprocessing and selecting features that are generalizable, enhancing the model's ability to predict the lifespan of unseen batteries.
Smart Images
Figure BR2025050608_02072026_PF_FP_ABST
Abstract
Description
[0001] "METHOD FOR ESTIMATING THE REMAINING USEFUL LIFE OF A BATTERY, COMPUTER-READABLE NON-TRANSIENTIAL MEANS AND ELECTRONIC ESTIMATION DEVICE"
[0002]
[0001] The present invention relates to a method for estimating the remaining service life of a battery.
[0003]
[0002] The present invention also relates to a non-transient computer-readable means, including a computer program comprising software instructions which, when executed by a computer, implement such a method. It also relates to an electronic estimating device for estimating the remaining service life of a battery.
[0004]
[0003] Given the current scenario of increasing demand for batteries in various sectors of the economy due to the need for electrification in the energy transition, it is necessary to understand batteries in order to guide their application, ensuring purchase and sale, appropriate use and manufacture of new chemical products.
[0005]
[0004] There are known methods that allow for better diagnosis of a battery, in particular for estimating its remaining service life. This parameter is specifically relevant for the diagnosis of second-life batteries and for defining battery charging and usage parameters to optimize their capabilities depending on the remaining service life. For example, documents KR 20230166196 A and CN 113109715 A describe methods for predicting the remaining service life of a battery using machine learning techniques.
[0006]
[0005] However, the accuracy of these predictions depends heavily on the data used to train the machine learning model. In particular, batteries, even from the same manufacturing batch, have differences between them, which means that test results obtained from one battery are not necessarily applicable to others. Furthermore, the trade-off between the accuracy in estimating the remaining lifespan of batteries and the number of test cycles required still tends to involve a large number of test cycles, which is sometimes not feasible, as the battery may not degrade enough for testing purposes.
[0007]
[0006] Therefore, it is an objective of the invention to provide a method for estimating the remaining service life of a battery using a machine learning model, which reduces prediction errors and the number of cycles required when testing a battery that has not been seen by the model before.
[0008]
[0007] To this end, the object of the invention is a method for estimating the remaining useful life of a battery, the method being implemented by an electronic estimation device and comprising extracting a set of selected features from at least one charge-discharge cycle of the battery and estimating the remaining useful life of the battery by applying a machine learning model to the set of selected features, the machine learning model having as input variables the extracted selected features and as an output variable an estimate of said remaining useful life, the machine learning model having been trained during a preliminary phase, the preliminary phase comprising training the machine learning model with a training set of selected features extracted from several charge-discharge cycles of each reference battery from a set of reference batteries whose remaining useful lives are known,the preliminary phase, also including, before the aforementioned training:
[0009] Extract a set of primary characteristics from the charge-discharge cycles of each reference battery in the reference battery set, each primary characteristic assuming a respective value in each charge-discharge cycle, said respective value being a statistical metric computed on a variable measured during said charge-discharge cycle, each primary characteristic corresponding to a respective metric and variable; process the set of primary characteristics by applying, for each primary characteristic and each reference battery in the reference battery set, an outlier detection treatment and / or a data smoothing treatment on a curve of said primary characteristic versus charge-discharge cycles; and
[0010] Select the features from among the processed primary features using correlation between the set of primary features and the known remaining service lives.
[0011]
[0008] Thanks to the invention, features are pre-processed and selected so that the machine learning model learns properties that are better generalizable. More precisely, outlier detection and / or data smoothing treatment allows the unusable behavior of a reference battery during the preliminary phase to be ignored by the model during training, while feature selection thanks to correlation allows selecting the features that are most related to the remaining lifespan across the entire set of reference batteries. Therefore, predicting the remaining lifespan of a battery never seen by the trained model is more accurate and requires fewer charge-discharge cycles.
[0012]
[0009] According to other aspects of the invention, which are advantageous but not mandatory, the method may incorporate one or more of the following features:
[0013]
[0010] - The variable corresponding to a given primary characteristic belongs to the group of variables that includes:
[0014] • a reference battery voltage;
[0015] • a reference battery current;
[0016] • a reference battery capacity; and
[0017] • a reference battery temperature;
[0011] - The statistical metric corresponding to a given primary characteristic belongs to the group of metrics that includes:
[0018] • an average of the variable through a charge-discharge cycle;
[0019] • a median of the variable across a charge-discharge cycle;
[0020] • a standard deviation of the variable through a charge-discharge cycle;
[0021] • a sum of the variable through a charge-discharge cycle;
[0022] • an interquartile range of the variable across a charge-discharge cycle; and
[0023] • a kurtosis of the variable through a charge-discharge cycle;
[0024]
[0012] - The set of primary characteristics comprises all primary characteristics corresponding to all combinations between the group of metrics and the group of variables;
[0025]
[0013] - A battery is a battery cell, a set of battery cells, or a battery pack;
[0026]
[0014] - The machine learning model produces an estimate of the remaining battery life from a set of selected features extracted based on a single charge-discharge cycle of the battery;
[0027]
[0015] - The outlier detection treatment of a given primary characteristic curve versus charge-discharge cycles comprises:
[0028] • Cut the curve into several parts, each part corresponding to a respective interval of charge-discharge cycles;
[0029] • For each part, calculate a first quartile and a third quartile of the primary characteristic values;
[0030] • For each part, identify values that are less than the first quartile minus an interquartile range multiplied by a given margin, or greater than the third quartile plus the interquartile range multiplied by the given margin, where the interquartile range is the result of subtracting the first quartile from the third quartile; such values are called outliers; and
[0031] • For each part, exclude each outlier or replace each outlier with a representative value;
[0032]
[0016] - The smoothing treatment of the curve of a given primary characteristic versus charge-discharge cycles involves applying a moving average to the primary characteristic through charge-discharge cycles;
[0033]
[0017] - The reference battery set comprises different groups of reference batteries and a charge / discharge protocol applied to the reference batteries differs from one group of reference batteries to another;
[0034]
[0018] - The preliminary phase comprises calculating, for each reference battery in the reference battery set and each primary characteristic, a target correlation coefficient of said primary characteristic extracted from said reference battery with the remaining useful life of said reference battery; the preliminary phase comprises calculating, for each primary characteristic, an average value of the target correlation coefficient and a standard deviation of the target correlation coefficient across all reference batteries in the reference battery set; and the selected characteristics are the primary characteristics that maximize the average value of the target correlation coefficient and minimize the standard deviation of the target correlation coefficient;
[0019] - The preliminary phase comprises calculating, for each reference battery in the set of reference batteries and each primary characteristic, a characteristic correlation coefficient, defined as an average value across all other primary characteristics of a respective correlation coefficient of said primary characteristic extracted from said reference battery with said other primary characteristic extracted from said reference battery; the preliminary phase comprises calculating, for each primary characteristic, an average value of the characteristic correlation coefficient and a standard deviation of the characteristic correlation coefficient across all reference batteries in the set of reference batteries; and the selected characteristics are the primary characteristics that minimize the average value of the characteristic correlation coefficient and the standard deviation of the characteristic correlation coefficient.
[0035]
[0020] The invention also relates to a non-transient computer-readable means, including a computer program comprising software instructions that, when executed by a computer, implement a method as defined above.
[0036]
[0021] The invention also relates to an electronic estimating device for estimating the remaining service life of a battery, comprising:
[0037] an acquisition module, configured to receive at least one measured variable during at least one battery charge-discharge cycle;
[0038] a feature extraction module, configured to extract a set of selected features from at least one variable; and
[0039] An estimation module, configured to provide an estimate of the remaining battery life using a trained machine learning model, the machine learning model having as input variables the selected extracted features and as an output variable an estimate of said remaining life, the machine learning model having been trained during a preliminary phase, the preliminary phase comprising training the machine learning model with a training set of selected features extracted from several charge-discharge cycles of each reference battery from a set of reference batteries whose remaining lifespans are known, the preliminary phase further comprising, before said training:
[0040] + extract a set of primary characteristics from the charge-discharge cycles of each reference battery from a set of reference batteries whose remaining service lives are known, each primary characteristic being a statistical metric computed on a variable measured during the said charge-discharge cycle;
[0041] + process the set of primary characteristics by applying, for each primary characteristic and each reference battery in the set of reference batteries, an outlier detection treatment and / or a data smoothing treatment on a curve of said primary characteristic versus charge-discharge cycles; and
[0042] + select features from among the processed primary features using correlation between the set of primary features and known remaining lifespans.
[0043]
[0022] According to an additional aspect of the invention, which is advantageous but not mandatory, the electronic estimation device further comprises a processing module configured to apply an outlier detection treatment and / or a data smoothing treatment to a primary characteristic curve versus charge-discharge cycles.
[0044]
[0023] The invention will be better understood based on the following description, which is given in correspondence with the accompanying Figures and as an illustrative example, without restricting the object of the invention. In the accompanying Figures:
[0045]
[0024] Figure 1 is a diagram of a battery, a battery management system and an electronic estimation device according to the invention;
[0046]
[0025] Figure 2 represents example curves of different variables versus time during a respective charge-discharge cycle of a reference battery, each inset (A) to (D) corresponding to a respective variable;
[0047]
[0026] Figure 3 represents example curves of a primary characteristic versus charge-discharge cycles of a reference battery, before outlier detection treatment on insertion (A) and after outlier detection on insertion (B);
[0048]
[0027] Figure 4 is a diagram of a preliminary phase of a method for estimating the remaining service life of a battery according to the invention; and
[0049]
[0028] Figure 5 is a diagram of a method for estimating the remaining service life of a battery according to the invention.
[0050]
[0029] Figure 1 represents a battery (1), a battery management system (BMS) (3) and an electronic estimating device (5).
[0051]
[0030] The battery (1) is advantageously a battery cell. In a variant, the battery (1) is a set of battery cells or a battery pack. For example, the battery (1) is a lithium-ion battery. The battery (1) is characterized by a remaining service life (RUL).
[0052]
[0031] The remaining useful life of battery (1) is the number of cycles that the battery (1) can still operate effectively before degrading to a point where it can no longer perform its intended function.
[0032] Knowing the remaining useful life of battery (1) can be useful for characterizing, validating, monitoring, and developing battery (1) purposes. For example, battery (1) is a second-life battery whose remaining useful life is unknown, and determines an application of the second-life battery (1).
[0053]
[0033] The battery (1) comprises two terminals (1A and 1B), the terminals (1A and 1B) being advantageously connected to an electrical load (7).
[0054]
[0034] The battery management system (3) is an electronic device configured to manage and monitor the battery (1). For example, the battery management system (3) is configured to manage the charge and discharge cycles of the battery (1) and to acquire a plurality of variables during said charge and discharge cycles.
[0055]
[0035] Advantageously, the plurality of variables comprises a battery voltage (1), a battery current (1), a battery capacity (1) and / or a battery temperature (1). The battery voltage (1) is a voltage between both terminals (1A and 1B). The battery current (1) is a current flowing between both terminals (1A and 1B). The battery capacity (1) represents the amount of electrical energy that can be supplied by the battery (1) before it needs to be recharged, and is given in ampere-hours. The battery temperature (1) is, for example, measured in a battery casing (1).
[0056]
[0036] The electronic estimating device (5) is connected to the battery management system (3) and configured to estimate the remaining battery life (1). In a variant not shown, the electronic estimating device (5) belongs to the battery management system (3).
[0057]
[0037] The electronic estimation device (5) comprises an acquisition module (9), a feature extraction module (11) and an estimation module (13). As an optional addition, the electronic estimation device (5) further comprises a data processing module (15).
[0058]
[0038] The acquisition module (9) is configured to receive at least one of the variables measured by the battery management system (3) during at least one battery charge-discharge cycle (1). Advantageously, the acquisition module (9) receives the voltage, current, capacity and battery temperature (1) measured by the battery management system (3).
[0059]
[0039] A charge-discharge cycle is a cycle during which the battery (1) is charged and discharged, in any order. In other words, the charge-discharge cycle comprises a battery charge (1) followed by a battery discharge (1), or a battery discharge (1) followed by a battery charge (1). Alternatively, a charge-discharge cycle comprises only battery charging or battery discharging (1).
[0060]
[0040] Figure 2 shows examples of variables versus time, as measured by the battery management system and acquired by the acquisition module (9), during a respective charge-discharge cycle. Examples of battery current (I) (1) versus time are shown in inset (A), examples of battery capacity (Q) (1) versus time in inset (B), examples of battery temperature (T) (1) versus time in inset (C), and examples of battery voltage (U) (1) versus time in inset (D). Each of the insets (A) to (D) shows the evolution of the said variable during a period corresponding to a charge-discharge cycle. In each inset, each curve, represented in a respective pattern, represents this evolution through a different cycle. For example, the dashed curve l(t)i corresponds to the evolution of battery current (1) during the 1 obattery charge-discharge cycle (1), while the solid line curve I(t)e29 corresponds to the 629th battery charge-discharge cycle (1).
[0061]
[0041] In the context of predicting remaining battery life (1), the variables are not measured and acquired in all cycles shown in Figure 2. For example, the variables are measured and acquired in a single battery charge-discharge cycle (1).
[0062]
[0042] As can be seen by comparing the different curves in each insertion (A) to (D) of Figure 2, the evolution of a given variable versus time during a charge-discharge cycle depends on the charge-discharge cycle considered. This shows that using such variables to predict the remaining battery life (1) is relevant.
[0063]
[0043] The feature extraction module (11) is configured to extract a set of selected features from the variables acquired by the acquisition module (9). The features are selected from a set of primary features. A primary feature is a quantity that assumes a respective value in each load-discharge cycle, said respective value being a statistical metric computed on one of the variables during said load-discharge cycle. Each primary feature corresponds to a respective metric and variable.For example, the statistical metric is an average of the variable across a charge-discharge cycle, a median of the variable across a charge-discharge cycle, a standard deviation of the variable across a charge-discharge cycle, a sum of the variable across a charge-discharge cycle, an interquartile range of the variable across a charge-discharge cycle, or a kurtosis of the variable across a charge-discharge cycle. In other words, a primary characteristic is, for example, the average battery voltage (1) during a charge-discharge cycle. This primary characteristic is assuming a respective value for each charge-discharge cycle.
[0064]
[0044] When computed over multiple charge-discharge cycles, the set of primary characteristics therefore comprises a plurality of values for each primary characteristic. More precisely, the set of primary characteristics comprises one value per charge-discharge cycle and per primary characteristic. Therefore, each primary characteristic forms a curve (F(C)A) of said primary characteristic values versus charge-discharge cycles. An example of such a curve (F(C)A) of a primary characteristic (FA) versus charge-discharge cycles (C) is shown in Figure 3. As discussed above, when predicting the remaining battery life (1), the number of charge-discharge cycles is generally much smaller than that shown in Figure 3. In fact, Figure 3 shows a curve acquired during a preliminary phase, as explained further in the descriptive report.
[0065]
[0045] Advantageously, the set of primary features comprises at least two primary features corresponding to two different statistical metrics applied to the same variable. For example, the set of primary features comprises a first primary feature, which is the mean battery current (1), and a second primary feature, which is the standard deviation of the battery current (1).
[0066]
[0046] Advantageously, the set of primary characteristics comprises all primary characteristics corresponding to all combinations between the aforementioned metrics and the aforementioned variables.
[0067]
[0047] Thanks to this plurality of primary characteristics, the chances that the set of primary characteristics will include characteristics relevant to predicting the remaining useful life are increased.
[0068]
[0048] Advantageously, features are selected from the set of primary features using rules that are determined during a preliminary phase, as described further in the descriptive report.
[0049] The estimation module (13) is configured to provide an estimate of the remaining battery life (1) using a machine learning model. The machine learning model takes the selected extracted features as input variables and provides an estimate of said remaining life as an output variable. The machine learning model was trained during a preliminary phase (100), illustrated in Figure 4 and described further in the descriptive report.The machine learning model is, for example, a meta-estimator that fits a series of random decision trees (also known as extra-trees) on various subsamples of a dataset and uses the average to improve predictive accuracy and control overfitting, called Extra TreesRegressor.
[0069]
[0050] The data processing module (15) is configured to apply outlier detection treatment and / or data smoothing treatment to the primary characteristic (FA) versus charge-discharge cycles (C) curve (F(C)A).
[0070]
[0051] Advantageously, outlier detection involves cutting the curve (F(C)A) into several parts, for example, four parts (P1, P2, P3 and P4), as represented in inset (A) of Figure 3. Each part (P1, P2; P3 or P4) corresponds to a respective charge-discharge cycle interval (R1, R2, R3 or R4).
[0071]
[0052] Outlier detection also involves, for each part (P1, P2, P3 or P4), computing a first quartile and a third quartile of the primary characteristic values. The first quartile of a part is the value such that 25% of the primary characteristic values belonging to that part are below that value. The third quartile of a part is the value such that 75% of the primary characteristic values belonging to that part are below that value.
[0053] Then, outlier detection treatment involves identifying values that are less than the first quartile minus an interquartile range multiplied by a given margin or greater than the third quartile plus the interquartile range multiplied by the given margin, where the interquartile range is the result of subtracting the first quartile from the third quartile. Such values are called outliers (25). The given margin is, for example, 1.5.In the example in Figure 3, part (P1) includes an outlier (25), which is less than the first quartile of said part (P1) minus the interquartile range of that part multiplied by 1.5. Parts (P2, P3 and P4) do not have any outliers (25) in this example.
[0072]
[0054] Finally, the outlier detection treatment comprises, for each part, the exclusion of each outlier (25) or the replacement of each outlier (25) with a representative value. The representative value is advantageously a value of the said primary characteristic in the previous cycle, since the difference between the values of a given primary characteristic in two successive cycles is minimal. The result is a corrected curve (F(C)B) of the primary characteristic (F) through charge-discharge cycles (C), as shown in inset (B) of Figure 3.
[0073]
[0055] Implementing outlier treatment in different parts (P1, P2, P3 and P4) of the curve (F(C)A) allows avoiding the identification of a relevant part of the curve (F(C)A), where the value of the characteristic (F) decreases, for example, as an outlier.
[0074]
[0056] The smoothing treatment of the curve (F(C)A) of a given primary characteristic (F) versus charge-discharge cycles (C) comprises applying a moving average to the primary characteristic (F) through charge-discharge cycles (C). More precisely, an average is applied successively to a set of charge-discharge cycle intervals (C), each interval of the set of intervals comprising some cycles (C), for example 20 cycles (C), and starting at a different cycle (C), for example shifted 10 cycles (C).
[0075]
[0057] Outlier detection treatment aims to remove singularities in the selected feature set. Data smoothing treatment aims to reduce noise. Therefore, the data processing module (15) participates in reducing the prediction error by the machine learning model.
[0076]
[0058] In the example in Figure 1, the electronic estimation device (5) comprises an information processing unit (17), for example, composed of a memory (19) and a processor (21) associated with the memory (19).
[0077]
[0059] In the example of Figure 1, the acquisition module (9), the feature extraction module (11) and the estimation module (13), as well as, by optional addition, the data processing module (15), are each made in the form of software, or a software component, executable by the processor (21). The memory (19) of the electronic estimation device (5) is then capable of storing an acquisition software, a feature extraction software and an estimation software, as well as, by optional addition, a data processing software. The processor (21) is then capable of executing each of the software applications among the acquisition software, the feature extraction software and the estimation software, as well as, by optional addition, the data processing software.
[0078]
[0060] In a variant not shown, the acquisition module (9), the feature extraction module (11) and the estimation module (13), as well as, by optional addition, the data processing module (15), are each made in the form of a programmable logic component, such as an FPGA (Field Programmable Gate Array), or in the form of a dedicated integrated circuit, such as an ASIC (Application Specific Integrated Circuit).
[0079]
[0061] When the electronic estimation device (5) is made in the form of one or more software programs, i.e., in the form of a computer program, it is still capable of being stored on a computer-readable medium, not shown. The computer-readable medium is, for example, a medium suitable for storing electronic instructions and capable of being coupled to a computer system bus. As an example, the readable medium is an optical disc, a magnetic-optical disc, a ROM memory, a RAM memory, any type of non-volatile memory (e.g., EPROM, EEPROM, FLASH, NVRAM), a magnetic card or an optical card. A computer program including software instructions is then stored on the readable medium.
[0080]
[0062] Advantageously, the electronic estimating device (5) further comprises a communication interface (23). The communication interface (23) is, for example, a display screen and is configured to display the remaining service life to a user.
[0081]
[0063] Before estimating the remaining battery life (1), the machine learning model is trained during a preliminary phase (100), represented in Figure 4 and described below.
[0082]
[0064] The preliminary phase (100) involves an unrepresented set of reference batteries whose remaining lifespans are known. Several charge-discharge cycles of each reference battery in the reference battery set are used to train the machine learning model. The reference batteries are, advantageously, battery cells of the same type as battery (1), for example, lithium-ion. The Battery Management System (3) controls a plurality of charge-discharge cycles for each reference battery in the reference battery set. During each charge-discharge cycle, the aforementioned variables (I, Q, T, U) are acquired.
[0083]
[0065] Advantageously, the reference battery set comprises different groups of reference batteries. A group of reference batteries is defined by a respective charge-discharge protocol. In other words, the charge / discharge protocol applied to the reference batteries differs from one group of reference batteries to another. For example, charge-discharge protocols differ from each other by a capacity rate, also called the C-rate, which represents the measurement of the current at which a battery is charged relative to its nominal capacity.
[0084]
[0066] The preliminary phase comprises a feature extraction stage (110), a primary feature processing stage (120), a feature selection stage (130) and a training stage (140).
[0085]
[0067] The feature extraction step (110) consists of extracting a set of primary features from the charge-discharge cycles of each reference battery in the reference battery set. In other words, the respective values of each primary feature are computed by applying the aforementioned statistical metric to the variables (I, Q, T, U) measured during each charge-discharge cycle.
[0086]
[0068] The primary characteristics are advantageously computed over a relatively large number of charge-discharge cycles, for example, at least 400 charge-discharge cycles.
[0087]
[0069] The feature extraction step (100) therefore provides, for each primary feature of the primary feature set and each battery of the reference battery set, a curve (F(C)A) of said primary feature (F) versus charge-discharge cycles (C) similar to that illustrated in inset (A) of Figure 3.
[0070] Then, the primary feature processing step (120) consists of processing the primary feature set by applying, for each primary feature and each reference battery of the reference battery set, an outlier detection treatment and / or a data smoothing treatment to the curve (F(C)A) of said primary feature versus charge-discharge cycles. The outlier detection treatment and the data smoothing treatment are advantageously implemented as explained above.The primary feature processing step (120), therefore, prevents the machine learning model from learning unusable behaviors from a reference battery. In other words, the primary feature processing step (120) ensures that the machine learning model generalizes better to a battery not seen during training.
[0088]
[0071] The feature selection step (130) consists of selecting features from among the processed primary features using correlation between the set of primary features and the known remaining service lives.
[0089]
[0072] In one embodiment of the invention, all primary features are selected. This embodiment is equivalent to skipping the feature selection step (130), which is illustrated by the dashed arrow in Figure 4.
[0090]
[0073] In a more advantageous embodiment of the invention, not all primary features are selected. Therefore, the feature selection step (130) allows for a reduction in the size of the machine learning model by focusing on the most relevant features.
[0091]
[0074] Advantageously, the feature selection step (130) comprises calculating, for each reference battery in the reference battery set and each primary feature, a target correlation coefficient of said primary feature extracted from said reference battery with the remaining service life of said reference battery. The target correlation coefficient therefore represents how relevant the primary feature is for predicting the remaining service life of said reference battery.
[0092]
[0075] Then, an average value of the target correlation coefficient and a standard deviation of the target correlation coefficient are advantageously calculated across all reference batteries in the reference battery set. Finally, the selected features are the primary features that maximize the average value of the target correlation coefficient and minimize the standard deviation of the target correlation coefficient.
[0093]
[0076] For example, a minimum mean value and a maximum standard deviation value are defined before the feature selection step (130), and the selected features are the primary features whose target mean correlation coefficient value is above the minimum mean value and whose target standard deviation correlation coefficient is below the maximum standard deviation value. The minimum mean value is, for example, 0.8 and the maximum standard deviation value is, for example, 0.2.
[0094]
[0077] Advantageously, a similar correlation analysis is conducted to evaluate the relationship between each primary trait and all other primary traits. This analysis allows for reducing redundancy when selecting traits.
[0095]
[0078] To conduct this correlation analysis, the preliminary phase further comprises calculating, for each reference battery in the set of reference batteries and each primary characteristic, a characteristic correlation coefficient defined as an average value across all other primary characteristics of a respective correlation coefficient of said primary characteristic extracted from said reference battery with said other primary characteristic extracted from said reference battery. The characteristic correlation coefficient therefore represents how much the primary characteristic correlates with the other primary characteristics. Primary characteristics with a low characteristic correlation coefficient are considered unique and help to reduce redundancy within the set of selected characteristics.
[0096]
[0079] Then, an average value of the trait correlation coefficient and a standard deviation of the target correlation coefficient are advantageously calculated across all reference batteries in the reference battery set. Finally, the selected traits are the primary traits that minimize the average value of the trait correlation coefficient and the standard deviation of the trait correlation coefficient.
[0097]
[0080] For example, a maximum mean value and a second maximum standard deviation value are defined before the feature selection step (130), and the selected features are the primary features respecting the criteria mentioned above in the target correlation coefficient, whose mean value of the feature correlation coefficient is below the minimum mean value and whose standard deviation of the target correlation coefficient is below the second minimum standard deviation value. The minimum mean value is, for example, 0.75 and the second minimum standard deviation value is, for example, 0.2.
[0098]
[0081] Thanks to this implementation of the feature selection step (130), the selected features are the preliminary features that are most correlated with the remaining service life and that minimize redundancy for most reference batteries in the reference battery set.
[0099]
[0082] In particular, when the reference battery set comprises several reference battery groups, as described above, this method allows the selection of features that are relevant to each, or to most, of the different reference battery groups. In other words, this implementation avoids selecting features that may be highly correlated with the remaining lifespans of reference batteries belonging to some groups, but poorly correlated with the remaining lifespans of reference batteries belonging to other groups. Therefore, the feature selection step (130) ensures the robustness and generalization of the machine learning model trained for a wider range of operating conditions.
[0100]
[0083] The training step (140) consists of training the machine learning model with a training set of selected features, thus providing the trained machine learning model. During training, the machine learning model learns to predict the remaining lifespan of a given battery from the selected features during the charge-discharge cycle(s) of said battery. The training step (140) advantageously includes hyperparameter tuning and validation using, for example, a 10-fold K-fold cross-validation method.
[0101]
[0084] Therefore, at the end of the preliminary phase (100), the machine learning model is trained to be able to predict the remaining battery life (1) using the selected feature set extracted from data from at least one battery charge-discharge cycle (1). Advantageously, the machine learning model predicts the remaining battery life (1) using the selected feature set extracted from a single battery charge-discharge cycle (1).
[0102]
[0085] A method (200) for estimating the remaining useful battery life (1) is represented in Figure 5 and described below.
[0103]
[0086] The method (200) comprises an extraction step (210) and an estimation step (230). Advantageously, the method (200) also comprises a processing step (220) and a display step (240).
[0104]
[0087] The extraction step (210) is implemented by the feature extraction module (11) and consists of extracting the set of selected features from the variables (I, Q, T and U) measured during at least one battery charge-discharge cycle (1).
[0105]
[0088] The processing step (220) is implemented by the data processing module (15) and consists of applying an outlier detection treatment and / or a data smoothing treatment to a curve (F(C)A) of the primary characteristic (F) versus battery charge-discharge cycles (C) (1). The processing step (220) is optional, but allows the machine learning model to ignore unusual battery behavior (1) during at least one charge-discharge cycle. In particular, the processing step (220) is useless if the at least one charge-discharge cycle comprises a single charge-discharge cycle.
[0106]
[0089] The estimation step (230) is implemented by the estimation module (13) and consists of providing an estimate of the remaining battery life (1) using the trained machine learning model. More precisely, the estimation step (230) consists of running the trained machine learning model with the selected set of features extracted from at least one battery charge-discharge cycle (1) as inputs, which returns the estimated remaining life as output.
[0107]
[0090] The display stage (240) is implemented by the communication interface (23) and consists of displaying to the user the remaining service life predicted by the electronic estimation device (5).
[0108] The respective characteristics of the electronic estimation device (5) and of the method (200) and embodiments considered in this description can be combined.
Claims
CLAIMS 1. METHOD (200) FOR ESTIMATING THE REMAINING USEFUL LIFE OF A BATTERY (1), characterized in that the method (200) is implemented by an electronic estimation device (5) and comprises extracting (210) a set of selected features from at least one charge-discharge cycle of the battery (1) and estimating (230) the remaining useful life of the battery (1) by applying a machine learning model to the set of selected features, the machine learning model having as input variables the extracted selected features and as an output variable an estimate of the remaining useful life, the machine learning model having been trained during a preliminary phase (100),the preliminary phase (100) comprising training the machine learning model (140) with a training set of selected features extracted from several charge-discharge cycles (C) of each reference battery from a set of reference batteries whose remaining service lives are known, wherein the preliminary phase (100) further comprises, before training (140):, extract (110) a set of primary characteristics of the charge-discharge cycles (C) of each reference battery in the set of reference batteries, each primary characteristic (F) assuming a respective value in each charge-discharge cycle, the respective value being a statistical metric computed on a variable (I, Q, T, U) measured during the charge-discharge cycle, each primary characteristic (F) corresponding to a respective metric and variable; process (120) the set of primary features by applying, for each primary feature (F) and each reference battery in the set of reference batteries, an outlier detection treatment (25) and / or a data smoothing treatment on a curve (F(C)A) of the primary feature (F) versus charge-discharge cycles (C); and select (130) the features among the processed primary features using correlation between the set of primary features and known remaining service lives.
2. METHOD (200), according to claim 1, characterized in that the variable (I, Q, T, U) corresponding to a given primary characteristic (F) belongs to the group of variables comprising: a reference battery voltage (U); a reference battery current (I); a reference battery capacity (Q); and a reference battery temperature (T).
3. METHOD (200), according to any one of claims 1 to 2, characterized in that the statistical metric corresponding to a given primary characteristic (F) belongs to the group of metrics comprising: an average of the variable (I, Q, T, U) through a charge-discharge cycle; a median of the variable (I, Q, T, U) through a charge-discharge cycle; a standard deviation of the variable (I, Q, T, U) through a charge-discharge cycle; a sum of the variable (I, Q, T, U) through a charge-discharge cycle; an interquartile range of the variable (I, Q, T, U) through a charge-discharge cycle; and a kurtosis of the variable (I, Q, T, U) through a charge-discharge cycle.
4. METHOD (200), according to any one of claims 2 to 3, characterized in that the set of primary characteristics comprises all the primary characteristics corresponding to all combinations between the group of metrics and the group of variables.
5. METHOD (200), according to any one of claims 1 to 4, characterized in that the battery (1) is a battery cell, a set of battery cells or a pack of batteries.
6. METHOD (200), according to any one of claims 1 to 5, characterized in that the machine learning model produces the estimate of the remaining useful life of the battery (1) from the set of selected features extracted based on a single charge-discharge cycle of the battery (1).
7. METHOD (200), according to any one of claims 1 to 6, characterized by the outlier detection treatment (25) of the curve (F(C)A) of a given primary characteristic (F) versus charge-discharge cycles (C) comprising: cut the curve (F(C)A) into several parts (P1, P2, P3, P4), each part corresponding to a respective interval (R1, R2, R3, R4) of charge-discharge cycles (C); For each part (P1, P2, P3, P4), compute a first quartile and a third quartile of the primary characteristic (F) values; for each part (P1, P2, P3, P4), identify values that are less than the first quartile minus an interquartile range multiplied by a given margin or greater than the third quartile plus the interquartile range multiplied by the given margin, where the interquartile range is the result of subtracting the first quartile from the third quartile, such values being called outliers (25); and for each part (P1, P2, P3, P4), exclude each outlier (25) or replace each outlier (25) with a representative value.
8. METHOD (200), according to any one of claims 1 to 7, characterized by the smoothing treatment of curve data (F(C)A) of a given primary characteristic (F) versus charge-discharge cycles (C) comprising applying a moving average to the primary characteristic (F) through charge-discharge cycles (C).
9. METHOD (200), according to any one of claims 1 to 8, characterized by the set of reference batteries comprising different groups of reference batteries and wherein a charge / discharge protocol applied to the reference batteries differs from one group of reference batteries to another.
10. METHOD (200), according to any one of claims 1 to 9, characterized by: the preliminary phase (100) comprises calculating, for each reference battery in the reference battery set and each primary characteristic (F), a target correlation coefficient of the primary characteristic (F) extracted from the reference battery with the remaining service life of the reference battery; the preliminary phase (100) shall comprise calculating, for each primary characteristic (F), an average value of the target correlation coefficient and a standard deviation of the target correlation coefficient across all reference batteries in the set of reference batteries; and The selected characteristics are the primary characteristics that maximize the average value of the target correlation coefficient and minimize the standard deviation of the target correlation coefficient.
11. METHOD (200), according to claim 10, characterized by: the preliminary phase (100) shall comprise calculating, for each reference battery of the set of reference batteries and each primary characteristic (F), a characteristic correlation coefficient, defined as an average value across all other primary characteristics of a respective correlation coefficient of the primary characteristic (F) extracted from the reference battery with the other primary characteristic (F) extracted from the reference battery; the preliminary phase (100) shall comprise calculating, for each primary characteristic (F), an average value of the characteristic correlation coefficient and a standard deviation of the characteristic correlation coefficient across all reference batteries in the set of reference batteries; and the selected characteristics (100) are the primary characteristics that minimize the average value of the characteristic correlation coefficient and the standard deviation of the characteristic correlation coefficient.
12. NON-TRANSIENT COMPUTER-READABLE MEDIUM, characterized by including a set of instructions that, when executed by a computer, implement a method as defined in any of claims 1 to 11.
13. ELECTRONIC ESTIMATION DEVICE (5) for estimating the remaining service life of a battery (1), characterized by the electronic estimation device (5) comprising: an acquisition module (9), configured to receive at least one variable (I, C, T, U) measured during at least one battery charge-discharge cycle (1); a feature extraction module (11), configured to extract a set of selected features from at least one variable (I, C, T, U); and an estimation module (13), configured to provide an estimate of the remaining battery life (1) using a trained machine learning model, the machine learning model having as input variables the selected extracted features and as an output variable an estimate of the remaining life, the machine learning model having been trained during a preliminary phase (100), the preliminary phase (100) comprising the training (140) of the machine learning model with a training set of selected features extracted from several charge-discharge cycles (C) of each reference battery from a set of reference batteries whose remaining lifespans are known; wherein the preliminary phase (100) also includes, before training (140): + extract (110) a set of primary characteristics of the charge-discharge cycles (C) of each reference battery from the set of reference batteries whose remaining useful lives are known, each primary characteristic (F) being a statistical metric computed on a variable measured during the charge-discharge cycle (C); + process (120) the set of primary characteristics by applying, for each primary characteristic (F) and each reference battery in the set of reference batteries, an outlier detection treatment (25) and / or a data smoothing treatment in a curve (F(C)A) of the primary characteristic (F) versus charge-discharge cycles (C); and + select (130) features from among the processed primary features using correlation between the set of primary features and known remaining useful lives.
14. ELECTRONIC ESTIMATION DEVICE (5), according to claim 13, characterized by further comprising a processing module (15) configured to apply outlier detection treatment (25) and / or data smoothing treatment on the primary characteristic (F) versus charge-discharge cycles (C) curve (F(C)A).