Method for estimating state of health of battery

Abstract

A method of estimating the state of health of a rechargeable battery (5). The method comprises: obtaining input data of a predetermined set of battery features collectively indicative of the state of health of the battery (5); applying a plurality of machine learning algorithms to perform battery (5) state-of-health estimation, wherein each machine learning algorithm calculates a battery (5) state-of-health estimate and a quantitative estimate of a confidence interval / value of the battery (5) state-of-health estimate based on the obtained input data of the battery characteristics; and applying a Kalman filter-based fusion algorithm to combine the health state estimates from all of the plurality of machine learning algorithms to provide a fused health state estimate.

Description

Technical Field

[0001] The present disclosure relates to a method for estimating the state of health (SoH) of a rechargeable battery and a corresponding system.

[0002] Furthermore, even though the method and system according to the present disclosure will be primarily described with respect to an automobile, the method and system are not limited to this particular vehicle, but may also be installed or implemented in another type of vehicle, such as a truck, a bus, a rail vehicle, an aircraft, a ship, an off-road vehicle, a mining vehicle, an agricultural vehicle, or a work vehicle, etc. In fact, the method and system according to the present disclosure may alternatively be used to estimate the SoH of a stationary battery, such as a large-scale grid battery power storage or a home battery power storage, etc. Background Art

[0003] Electrification has become an irreversible trend in the automotive industry, driven by environmental concerns, regulatory pressure, and technological advancements. Accurately estimating the battery's state of health (SoH) under various usage scenarios is essential for timely maintenance, optimized energy management, and customer satisfaction. Dynamic operating conditions, random user behavior, and cell-to-cell variations make battery SoH estimation challenging for practical applications.

[0004] Therefore, there is a need for an improved method of estimating battery SoH. Summary of the Invention

[0005] The object of the present disclosure is to provide an improved method and system for battery SoH estimation. This object is achieved at least in part by the features of the independent claims. The dependent claims contain further developments of the method and system.

[0006] According to a first aspect of the present disclosure, a method for estimating the state of health of a rechargeable battery is provided, the method comprising: obtaining input data of a predetermined set of battery features that collectively indicate the state of health of the battery; applying multiple machine learning algorithms to perform battery state of health estimation, wherein each machine learning algorithm calculates a battery state of health estimate and a confidence interval / value quantitative estimate of the battery state of health estimate based on the obtained input data of the battery features; and applying a Kalman filter (KF)-based fusion algorithm to combine the state of health estimates from all of the multiple machine learning algorithms to provide a fused state of health estimate.

[0007] According to a second aspect of the present disclosure, a system for estimating the state of health of a rechargeable battery is provided, the system comprising: a rechargeable battery; a group of sensors configured to sense a set of battery features about the rechargeable battery, wherein the set of battery features collectively indicates the state of health of the battery; and an electronic control unit connected to the group of sensors, wherein the electronic control unit is configured to: obtain input data related to a predetermined set of battery features; apply multiple machine learning algorithms to perform battery state of health estimation, wherein each machine learning algorithm calculates a battery state of health estimate and a quantitative estimate of a confidence interval / value of the battery state of health estimate based on the input data of the obtained battery features; and apply a Kalman filter (KF)-based fusion algorithm to combine the state of health estimates from all of the multiple machine learning algorithms to provide a fused state of health estimate.

[0008] In this way, a method is provided to more accurately and reliably estimate the battery's state of health (SOH) even when the battery is subjected to arbitrary usage conditions. After classifying various operating conditions, a suitable feature set is extracted to indicate the battery's state of health for each category. Multiple machine learning algorithms are deployed to utilize the obtained features for online SoH estimation. Based on a recently developed SoH prediction model using machine learning and usage-related histogram data, a Kalman filter is designed to optimally and in real time fuse all estimates. The efficacy and practicality of the developed method have been verified using experimental data on batteries of different types and under different operating conditions.

[0009] Further advantages are achieved by implementing one or several of the features of the dependent claims.

[0010] The term “input data” as used herein primarily refers to streaming data or online data, i.e., data measured online, for example by a battery management system (BMS), and used to feed into a machine learning pipeline for estimation.

[0011] In some exemplary embodiments, the input data is obtained in conjunction with a battery charging phase. Compared to the unpredictable and random discharge process, charging conditions are more predictable, such as for the constant current-constant voltage (CC-CV) scheme widely adopted in commercial BMSs. Therefore, since the charging event is relatively predetermined and well-defined in terms of the required time period, charging current, charging conditions, etc., the battery SoH estimate generated based on the battery charging phase is generally a more accurate estimate.

[0012] In some exemplary embodiments, which may be combined with any one or more of the above embodiments, the method further comprises a setup phase performed before the step of obtaining input data, wherein the setup phase comprises training a machine learning algorithm.

[0013] In some exemplary embodiments that may be combined with any one or more of the above embodiments, the predetermined battery characteristics include one or more of the following battery characteristics based on a recent battery charging event: a voltage profile; a current profile; a time interval between predetermined voltage windows; a signal strength over time, which is calculated as Where s(t) is the signal; the area under the current curve; the area under the voltage curve; the slope of the voltage curve; the slope of the current curve; the initial state of charge (SoC); the final SoC, a charge temperature-dependent characteristic; the peak of the capacity increase curve; the voltage level of the capacity increase curve at the peak; the final total battery output voltage; the final individual battery cell voltage; and the differential voltage curve. These battery characteristics are believed to provide a relatively strong indication of the battery's SoH.

[0014] In some exemplary embodiments that may be combined with any one or more of the above embodiments, the method further comprises a setup phase, the setup phase comprising: selecting a set of unique charging scenarios (S1-S5), each unique charging scenario having a unique charging start location and / or charging end location; and for each of the selected charging scenarios S1-S5, training a machine learning algorithm based on a data set corresponding to the selected charging scenario S1-S5, wherein the step of obtaining input data comprises: determining to which of the unique charging scenarios (S1-S5) the obtained input data corresponds, and wherein the step of applying a plurality of machine learning algorithms to perform battery state of health estimation comprises: for each machine learning algorithm, applying a machine learning algorithm trained based on data associated with the determined charging scenarios (S1-S5) to calculate the battery state of health estimate and a confidence interval / value quantitative estimate of the battery state of health estimate. Due to random user behavior, the state of charge (SoC) range may frequently change from one charging cycle to another in actual EV (electric vehicle) use, thereby causing the execution of the machine learning algorithm to be less accurate and reliable. However, by providing a machine learning algorithm trained on a separate set of predetermined partial charging curves, the SoH estimation output of the machine learning algorithm can be significantly improved by selecting the machine learning (ML) algorithm trained on the charging scenario most similar to the most recent real charging scenario.

[0015] In some exemplary embodiments, which may be combined with any one or more of the above embodiments, a predetermined set of battery characteristics indicative of a battery state of health is selected for each charging scenario (S1-S5) in the set of unique charging scenarios (S1-S5), and wherein the predetermined battery characteristics for at least one charging scenario (S1-S5) in the set of unique charging scenarios (S1-S5) are different from the predetermined battery characteristics for another charging scenario (S1-S5) in the set of unique charging scenarios (S1-S5). The most appropriate battery characteristics may be the same for different charging scenarios in some examples, but generally, a unique set of battery characteristics is applied within a charging scenario, and particularly within each charging scenario, to provide an improved battery state of health estimate.

[0016] In some exemplary embodiments that may be combined with any one or more of the above embodiments, each of the predetermined battery feature sets indicative of a battery state of health is determined by first identifying a preliminary set of battery features that collectively indicate a battery state of health, and then performing a correlation analysis on the preliminary battery features. The purpose of the correlation analysis is to eliminate redundant features, thereby improving estimation quality and / or computational efficiency.

[0017] In some exemplary embodiments that may be combined with any one or more of the above embodiments, the method further comprises: calculating a battery SoH prediction value and a quantitative estimate of a confidence interval / value of the battery SoH prediction value by a machine learning prediction model based on histogram data; and applying the Kalman filter (KF)-based fusion algorithm to combine the battery state of health estimation values from all of the multiple machine learning algorithms and the battery SoH prediction value from the machine learning prediction model based on the histogram data to provide a fused battery state of health estimation value.

[0018] For a given battery charging situation, the predefined scenarios S1-S5 may not be suitable for various reasons. For example, a given charging scenario may correspond to any of the predefined scenarios S1-S5, or a given battery charging situation may include data corruption, communication delays / failures, etc. In this case, it is impossible to extract the selected features using any of the charging scenarios S1-S5, making all the ML models discussed above infeasible. In this case, battery SoH estimation can be accurately performed using a model based on histogram data for capacity prediction, as this prediction model is less dependent on the given charging scenario and is therefore more robust.

[0019] In some exemplary embodiments that may be combined with any one or more of the above embodiments, the step of calculating a battery SoH prediction value by using a machine learning prediction model based on histogram data includes a setup phase, the setup phase including: obtaining historical battery usage data, converting the battery usage data into a 1D histogram and extracting statistical characteristics from the 1D histogram, determining battery characteristics based on the extracted statistical characteristics, providing a global model by selecting and offline training a machine learning algorithm based on the obtained historical battery usage data, and wherein during online use of the battery, the step of calculating a battery SoH prediction value by using a machine learning prediction model based on the histogram data includes: calculating a global battery SoH prediction value based on the global model; and online adjusting the global battery SoH prediction value based on a measured historical battery capacity estimate of the current battery to provide a final battery SoH prediction value.

[0020] In some exemplary embodiments that may be combined with any one or more of the above embodiments, the set of unique charging scenarios (S1-S5) includes one or more of the following battery charging scenarios: full constant current (CC)-constant voltage (CV) charging; partial CC-CV charging including starting after the peak of the increasing capacity (IC) curve and ending with a complete constant voltage (CV) phase; partial constant current (CC)-constant voltage (CV) charging when starting after the peak of the increasing capacity (IC) curve and ending without a constant voltage (CV) phase; partial constant current (CC)-constant voltage (CV) charging when starting before the peak of the increasing capacity (IC) curve and ending with a complete constant voltage (CV) phase; partial constant current (CC)-constant voltage (CV) charging when starting before the peak of the IC curve and ending without a constant voltage (CV) phase.

[0021] In some exemplary embodiments that may be combined with any one or more of the above embodiments, the method further comprises the step of: when the obtained input data does not correspond to any of the set of unique charging scenarios (S1-S5), setting the battery state of health estimate to be equal to the battery SoH prediction value derived by the machine learning prediction model based on the histogram data. In other words, during those charging events that do not correspond to any of the trained charging scenarios, the Kalman filter can be bypassed, and only the machine learning prediction model based on the histogram data provides a single reasonable SoH estimate, as fusing a single SoH prediction value does not make sense.

[0022] The present disclosure also relates to a vehicle comprising the above-mentioned system.

[0023] The present disclosure also relates to a data processing control unit, which includes a processor configured to execute the steps in the above method.

[0024] The present disclosure also relates to a computer program, which includes instructions. When the program is executed by a computer, the instructions cause the computer to execute the steps in the above method.

[0025] Other features and advantages of the present invention will become apparent when studying the appended claims and the following description. Those skilled in the art realize that different features of this disclosure may be combined to create other embodiments besides those explicitly described above and below, without departing from the scope of the present disclosure. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] The method and system according to the present disclosure will be described in detail below with reference to the accompanying drawings, in which:

[0027] Figure 1 schematically illustrates a side view of a battery electric vehicle having the system of the present disclosure,

[0028] Figure 2 The basic steps of the method according to the present disclosure are shown,

[0029] Figures 3A to 3E Various graphs showing battery charge and discharge conditions,

[0030] Figures 4A to 4C various graphs showing various battery capacity scenarios over time,

[0031] Figure 5 shows a flow chart of an exemplary embodiment of a method according to the present disclosure,

[0032] Figure 6 shows a flow chart of another exemplary embodiment of a method according to the present disclosure,

[0033] Figure 7 and Figure 8 Two exemplary embodiments of flowcharts describing how to identify suitable battery characteristics for an ML algorithm according to the present disclosure are shown.

[0034] Figure 9 and Figure 10 Two different diagrams showing the structure of a model for calculating a battery SoH prediction value by a machine learning prediction model based on histogram data according to the present disclosure,

[0035] Figure 11 shows a flow chart of another exemplary embodiment of a method according to the present disclosure,

[0036] 12A to 12E shows the final battery characteristics selected for five different exemplary charging scenarios according to the present disclosure,

[0037] Figure 13 The capacity estimation results for a certain battery type under different charging scenarios are shown.

[0038] 14A to 14C shows the estimation results of randomly selected battery cells under a certain charging scenario,

[0039] Figure 15 The standard deviation of the results from the best model, the worst model, and the model ensemble is shown for randomly selected battery cells.

[0040] 16A to 16D shows the capacity estimation results of randomly selected battery cells under actual charging cycles,

[0041] 17A to 17E Various exemplary flow charts are shown that describe the main steps of the method according to the present disclosure. DETAILED DESCRIPTION

[0042] Various aspects of the disclosure will be described below in conjunction with the accompanying drawings to illustrate rather than limit the disclosure, wherein like reference numerals represent like elements, and variations of the described aspects are not limited to the specifically illustrated embodiments but are applicable to other variations of the disclosure.

[0043] Those skilled in the art will understand that the steps, services, and functions described herein can be implemented using separate hardware circuits, using software running in conjunction with a programmed microprocessor or general-purpose computer, using one or more application-specific integrated circuits (ASICs), and / or using one or more digital signal processors (DSPs). It will also be understood that when the present disclosure is described in terms of methods, it can also be embodied in one or more processors and one or more memories coupled to the one or more processors, wherein the one or more memories store one or more programs that, when executed by the one or more processors, perform the steps, services, and functions disclosed herein.

[0044] The present method and system for estimating the state of health of a rechargeable battery can be implemented in various types of battery-powered applications, such as in electric vehicle applications, but is also applicable to, for example, stationary battery applications such as large-scale grid battery power storage or home battery power storage. This disclosure will describe a method and system for estimating the state of health of a rechargeable battery primarily for battery-powered electric vehicles, such as Figure 1 As shown schematically in FIG.

[0045] Specifically, Figure 1A side view of an electric vehicle 1 is shown, having front wheels 2a, rear wheels 2b, a passenger compartment 3, and an electric vehicle drivetrain, which includes a high-voltage battery 6 connected to an electric motor 5, for example, via a power converter such as an inverter. The output shaft of the electric motor 5 is drivingly connected to the front wheels 2a and / or the rear wheels 2b of the vehicle. The drivetrain also includes a charging inlet 4 configured to receive a charging connector from a charging station during charging of the high-voltage battery 6.

[0046] During charging of the high-voltage battery 6 , charge is supplied from the charging station to the high-voltage battery 6 via the charging interface 4 , and during vehicle travel, charge is supplied from the high-voltage battery 6 to the motor 5 for propulsion of the vehicle.

[0047] Driven by growing public awareness of environmental issues, stringent regulations on vehicle emissions, and advances in electric propulsion and energy storage technologies, large-scale electrification of the transportation sector has become an unstoppable global trend. Automakers are increasingly shifting from internal combustion engine-centric powertrain architectures to electrified solutions. Many have made bold commitments to electrify their entire product portfolio by 2025 and to cease sales of internal combustion engine vehicles as early as 2030. Lithium-ion batteries play a crucial role in this transition due to their high energy density, relatively low cost, and long service life. As one of the most critical and expensive vehicle components, batteries have garnered significant attention from both industry and academia. Among the various battery characteristics, aging is a key one, as it impairs performance and reliability, deteriorates the battery's state of health (SoH), and increases the risk of safety hazards. Furthermore, improper battery use can easily accelerate this aging process. Therefore, real-time understanding of aging processes and monitoring SoH are crucial for optimizing energy management, timely maintenance, and accurate residual value prediction.

[0048] In general, existing battery SoH estimation methods can be broadly categorized as empirical, model-based, and data-driven. Based on extensive laboratory cycling data, empirical methods typically use polynomial, exponential, or quadratic functions to fit battery aging trends. These methods often rely on the assumption of regular cycling conditions. However, batteries deployed in electric vehicles (EVs) encounter irregular and complex operating conditions, which inevitably undermine the accuracy and reliability of these estimation methods.

[0049] In contrast to the open-loop estimation of empirical methods, model-based approaches adopt a closed-loop approach. Two different types of models are commonly used, namely, equivalent circuit models (ECMs) developed from Kirchhoff’s laws and electrochemical models (EMs) derived from porous electrode theory

[10] . Due to their simple structure and ease of implementation, ECMs have been widely studied and applied to battery aging prediction. However, the fidelity of such models decreases when the conditions experienced in actual applications differ from laboratory characteristics and when the model parameters deviate from the true values due to the lack of regular reference performance tests. EMs have the ability to simulate the local dynamics inside the battery, including aging mechanisms such as solid electrolyte interface growth, lithium plating, and particle rupture. However, the high computational requirements and difficulty in model parameterization make such models difficult to use in online applications.

[0050] Many recent research efforts have sought to apply data-driven approaches to battery SoH estimation because of their flexibility, mechanism-independent nature, and ability to identify patterns and trends in complex dynamic situations. Feature construction is often a relevant step in such approaches, as the performance of ML algorithms often depends heavily on whether the selected battery features contain sufficient information to indicate the battery's state of aging. A typical battery management system (BMS) measures only the voltage of individual cells and the current of the battery pack, in addition to one or two temperature sensors per module in the battery pack. Therefore, feature selection is preferably based on the available information in real-world battery systems.

[0051] Depending on the implementation and the charge and discharge profiles, battery characteristics indicative of battery SoH can be derived from the charge and discharge processes. However, it is sometimes preferable to rely primarily or even solely on the charge process, such as in vehicle applications, as the uncontrolled operating conditions of the vehicle end user, especially intermittent regenerative braking during the discharge phase and long periods of parking that can lead to calendar aging, can pose significant challenges to the applicability of this approach. Compared to the often unpredictable and random discharge process, vehicle charging conditions are easily controlled, such as the constant current-constant voltage (CC-CV) scheme widely adopted in commercial BMSs. This has led to significant research interest in using CC-CV charging curves for SoH estimation.

[0052] The capacitance (IC) curve and differential voltage (DV) curve, which are commonly used to analyze battery aging mechanisms, can also be used to estimate battery SoH. To this end, characteristic values associated with the peaks or valleys of the IC / DV curve, such as height, are usually selected as features. However, in certain battery applications, such as battery electric vehicles, the SoC range may change frequently within different charging cycles in actual EV use due to random user behavior. Since the peaks and valleys of the IC curve only appear in specific voltage windows, the corresponding estimator becomes useless when such a window does not appear in the charging implementation. This requires researchers to work hard to develop reliable estimators from partial charging curves, so that the battery SoH estimator is particularly robust for battery electric vehicle implementations.

[0053] The partial charge profile is provided, for example, by first cycling the battery under a first partial charge stage, such as CC-CV charging starting at 20% SoC and ending at 80% SoC while recording relevant battery data, or CC charging starting before the IC peak and ending before reaching the CV stage while recording relevant battery data, and thereafter by cycling the battery under another partial charge stage while recording relevant battery data, and so on.

[0054] Alternatively, partial charge curves may be provided, for example, by using battery data generated by cycling a cell under full CC-CV charge, i.e., substantially 0-100% SoC, but where the partial charge curves are subsequently removed during data processing, either manually or by a computer, to form synthetic partial charge curves. These synthetic curves will differ from the natural partial charge curves of actual cells due to voltage polarization effects and internal resistance caused by the initial voltage rise, but may still provide valuable and useful results and insights.

[0055] Existing battery SoH estimators are typically based only on the full CC-CV charging curve. When applied based on a partial charging curve, this can lead to a model-plant mismatch, exposing existing battery SoH estimators to serious drawbacks and resulting in poor performance in vehicle applications. To the best of our knowledge, no existing SoH estimators based on IC / DV signatures completely, systematically, and statistically avoid this drawback.

[0056] This is at least partially addressed by applying a suite of machine learning (ML) models in conjunction with SoH estimation fusion to provide improved battery aging diagnostics.

[0057] Specifically, by combining and fusing the SoH estimation results of multiple different ML algorithms, the potential poor SoH estimation provided by a single ML algorithm can be significantly reduced. The advantage of the combined model over the individual models becomes even greater when the assumptions of the individual models do not strictly hold for certain situations.

[0058] The present disclosure describes an exemplary embodiment of a battery SoH estimation method that is efficient, practical, and easy to implement, which can improve accuracy, especially by more optimally combining various estimation values during actual deployment.

[0059] refer to Figure 2 According to some exemplary embodiments, a method for estimating the state of health of a rechargeable battery includes a first step S100: obtaining input data of a predetermined set of battery features that collectively indicate the state of health of a battery 5. Thereafter, the method includes a second step S200: applying multiple machine learning algorithms to perform battery 5 state of health estimation, wherein each machine learning algorithm calculates a battery 5 state of health estimate and a quantitative estimate of a confidence interval of the battery 5 state of health estimate based on the obtained input data of the battery features. Finally, the method includes a third step S300: applying a Kalman filter (KF)-based fusion algorithm to combine the state of health estimates from all of the multiple machine learning algorithms to provide a fused battery 5 state of health estimate.

[0060] All three steps S100-S300 are preferably performed online, i.e., in real time, during actual daily use of the vehicle, such as driving or parking. The sequence of steps constituting the method may be repeated periodically and / or in conjunction with a certain event, such as, in particular, a charging event, but may also or alternatively be repeated during other events, such as during driving, i.e., during a discharging event.

[0061] Where a battery SoH estimation method is applied at each charging event, the method will typically automatically generate an updated battery SoH estimate value after each charging event.

[0062] The second step S200 of applying multiple machine learning algorithms to perform battery health state estimation may, for example, include: applying two different types of ML algorithms based on the input data of the obtained battery characteristics, or applying three different types of ML algorithms based on the input data of the obtained battery characteristics, or applying four different types of ML algorithms based on the input data of the obtained battery characteristics, or applying five different types of ML algorithms based on the input data of the obtained battery characteristics, or applying more different types of ML algorithms based on the input data of the obtained battery characteristics.

[0063] In some exemplary embodiments, at least one Bayesian ML algorithm and at least one frequentist ML algorithm are applied to calculate two separate battery SoH estimates, which are then fused in a Kalman filter based on confidence intervals for the at least two separately calculated estimates. By using at least one ML algorithm of each type, the risk of systematic errors associated with a particular type of ML algorithm is reduced.

[0064] In some exemplary embodiments, at least two Bayesian ML algorithms and two frequentist ML algorithms are applied to develop SoH estimation models, each SoH estimation model quantitatively calculating an estimate confidence interval.

[0065] The present rechargeable battery health status estimation method and system can be based on battery characterization data from various types of charging scenarios, such as full CC-CV charging from about 0-100% SoC, or partial CC-CV charging scenarios, or other types of charging scenarios.

[0066] Therefore, the present rechargeable battery state of health estimation method and system can define two or more different predetermined charging scenarios, for example, during a setup phase, and then train multiple machine learning algorithms to estimate the battery SoH for each predetermined charging scenario.

[0067] Thus, during subsequent use of the vehicle, i.e., after leaving the vehicle manufacturing plant and entering actual battery use, the vehicle's electronic microcontroller can be configured to: for each new charging event, analyze the new charging event and identify the predetermined charging scenario that is most similar to the new charging event from a set of predetermined charging scenarios, and thereafter apply a plurality of machine learning algorithms trained on the most similar predetermined charging scenario to calculate a set of individual battery SoC estimates based on the battery characteristic data obtained from the new charging event. As described above, the individual battery SoH estimates are then fused in a Kalman filter to calculate a single battery SoH estimate.

[0068] In the following, reference will be made to 3A to 16D Specific exemplary embodiments of methods and systems for estimating the state of health of a rechargeable battery are described in more detail.

[0069] First, all possible charging processes of an electric vehicle (EV) are divided into a set of different charging scenarios. In this specific exemplary embodiment, six feasible and mutually exclusive scenarios are selected, but the method according to the present disclosure is obviously not limited to this particular number of charging scenarios, or even to completely different charging scenarios. For each charging scenario, a set of highly correlated features is extracted. Secondly, two Bayesian ML algorithms and two frequentist ML algorithms are applied to develop battery SoH estimation models, each model quantitatively calculating the confidence interval of the estimate. Thirdly, a Kalman filter (KF) based fusion algorithm is introduced to systematically combine all these battery SoH estimation models. Model fusion can be achieved in real time with only a small amount of additional computational effort. Relative to the individual battery SoH estimation models, the online fusion algorithm achieves highly accurate and robust results while significantly tightening the confidence intervals.

[0070] Later herein, illustrative results from a large number of batteries with different chemistries demonstrate the efficacy of methods and systems according to the present disclosure.

[0071] The appropriate choice of dataset can help when solving the battery SoH estimation problem under arbitrary vehicle operating conditions. In this regard, the dataset used for model development and validation should be as close as possible to actual battery usage. The dataset used in this paper to validate the efficacy of the battery SoH estimation model was obtained from Sandia National Laboratories (SNL). Y. Preger, HM Brownholtz, A. Fresquez, DLCampbell, BW Juba, J. Roma`n-Kustas, SR Ferreira, B. Chalamala, J. prepared and analyzed the dataset in the research paper "Degradation of Commercial Lithium-Ion Cells as a Function of Chemistry and Cycling Conditions" and was published in Electrochem. Soc. 167 (12) (2020) 120532. doi: 10.1149 / 1945-7111 / abae37. In addition, the dataset can be obtained, for example, via the URL https: / / www.batteryarchive.org / study_summaries.html To get it.

[0072] The SNL dataset and basic experiments were initially conducted to investigate the effects of various stress factors, such as discharge rate, depth of discharge, and ambient temperature, on the degradation performance of commercial battery cells of different chemistries. Battery cells with nickel-manganese-cobalt (NMC) and nickel-cobalt-aluminum (NCA) cathodes were selected. Detailed cell specifications are listed in Table 1.

[0073]

[0074] Table 1: Battery cell specifications

[0075] in addition, Figure 3A and Figure 3B The capacity retention trends for NMC and NCA cells are shown separately. All cells were charged at a 0.5C rate during the test. The capacity during the charging phase was accumulated and used as the ground truth for the battery capacity. Figure 3C 、 Figure 3D and Figure 3E The current, voltage and temperature curves of a typical charge and discharge cycle are respectively illustrated. Specifically, Figure 3A shows the capacity decay curves of 31 NMC battery cells, Figure 3B Represents the capacity decay curve of 24 NCA battery cells. Figures 3C to 3E Typical cycling conditions for the investigated batteries are illustrated. The test campaign setup is shown in Table 2. As can be seen, the battery cells were exposed to different partial charge levels and cycled at three different ambient temperatures.

[0076]

[0077] Table 2: Battery unit testing activities

[0078] As mentioned above, today, car companies generally use CC-CV charging strategy. Figure 3C The CC-CV charging strategy is schematically illustrated in Figure 1, where the first charging phase is performed using CC charging, i.e., applying a constant charging current while the battery voltage increases. The second charging phase is performed using CV charging, i.e., applying a constant voltage while the battery charging current slowly decreases. The third phase is the CC discharge phase, i.e., a constant current discharge phase while the battery voltage again slowly decreases.

[0079] Generated at a sampling rate of 1 Hz Figures 3A to 12D All measured values are shown in Figure 4a. When the charging method is CC-CV, the raw current and voltage data are used to extract smooth and consistent IC curves. The charging current and voltage curves of a typical battery studied at different cycles are shown in Figures 4a and 4b, respectively. The corresponding IC curve is shown in Figure 4c.

[0080] Unlike battery cell cycling in a laboratory under controlled conditions with repeatable conditions, real-world battery usage is highly dynamic, with operating conditions changing frequently. When an EV is shared by different drivers, usage can become even more complex, and driving conditions are often unpredictable. Furthermore, seasonal and local temperature variations can significantly impact battery performance and internal aging. Compared to discharge conditions, charging conditions are easier to control, and the same control laws are often repeated. Therefore, in this exemplary embodiment, only the charging curve is used to estimate capacity values. This eliminates the assumption typically applied in laboratory testing: that all charging cycles begin with the same initial SoC level and are contained within a fixed ambient temperature. In other words, the methods and systems according to the present disclosure are not limited to strictly repeatable CC-CV charging cycles. Instead, the methods and systems according to the present disclosure are configured to provide reliable battery SoH estimates regardless of the application of any CC-CV charging condition or any other charging strategy.

[0081] Based on the above discussion, all existing charging situations are divided into five different and mutually exclusive scenarios S1-S5, as shown in Figure 4a.

[0082] like Figure 4C As shown, the IC peak region is located between a lower voltage level 8 of approximately 3.7V and a higher voltage level 9 of approximately 3.8V. These two voltage levels are also included in Figure 4A and the IC peak range is Figure 4A 8 is defined as a range extending from an intersection point between the lower voltage level 8 and the charging voltage curve of the 708th charging cycle to an intersection point between the higher voltage level 9 and the charging voltage curve of the second charging cycle.

[0083] In this exemplary embodiment, the five different charging scenarios S1-S5 are:

[0084] Charging scenario S1: Full CC-CV charging from 0% SoC to 100% SoC.

[0085] Charging scenario S2: At IC peak 10 (see Figure 4C ) and ends with a complete CV phase (see Figure 4a).

[0086] Charging scenario S3: Partial CC charging starting before IC peak 10 and ending without a CV phase.

[0087] Charging scenario S4: Partial CC-CV charging starting after IC peak 10 and ending with a full CV phase.

[0088] Charging scenario S5: CC charging starting after IC peak 10 and ending without CV phase.

[0089] Note that there is no mention of temperature in any of the above scenarios, which means that the usual constraints imposed on ambient temperature in SoH estimator design are removed.

[0090] For ML algorithms to work accurately and efficiently, it's beneficial to select relevant and informative battery features. The process of selecting appropriate battery features is also known as feature construction. Furthermore, for the algorithm to be effective, the selected features must be usable under actual charging cycles. For a defined charging scenario, different features are added to the feature pool based on their availability.

[0091] In some exemplary embodiments, a specific voltage window related feature can be used as a battery feature. The voltage and current curves during the charging phase will gradually change as the battery decays, such as Figures 4A-4B As shown, Figure 4A The diagram shows the change of voltage curve and different charging scenarios. Figure 4B Indicates the change of the current curve, Figure 4C The changes in the IC curve are shown.

[0092] Therefore, a predefined sub-range of the raw charging voltage data can be selected as a basis for constructing the feature. According to a first example, the time interval between the predetermined voltage windows can be selected as the battery feature, because an aged battery cell will gradually require less charging time than a newer battery cell, e.g. Figure 2 As shown in A.

[0093] According to a second example, the signal strength over time, which is usually regarded as the energy of the signal, can also be used as a battery feature, where the energy of the signal can be calculated as

[0094]

[0095] where s(t) is the signal.

[0096] According to a third example, the area under the current and voltage curves may be used as a battery characteristic, since the area also represents the state of aging of the battery and can be used to estimate the battery capacity, ie, the battery SoH.

[0097] According to the fourth example, by observing the change of the voltage curve with the battery life, it is noticed that the voltage curve in the CC phase becomes steeper as the battery ages. Therefore, the slope of the voltage curve can be selected as a candidate battery feature.

[0098] Similarly, if there is a CV phase in the charging situation, ie, the charging scenario, the slope of the current curve in the CV phase can be used as a battery characteristic.

[0099] In some exemplary embodiments, IC curve-related features can be used as battery characteristics. IC (capacitance increase) is defined as the ratio between the change in charge capacity and the corresponding change in voltage over a predetermined time interval. As a promising approach for non-destructive battery aging mechanism identification and battery characterization methods, IC analysis is also a source of information-rich health indicators because the underlying pattern of the IC curve changes during battery degradation.

[0100] However, the specified time interval, division calculation, and measurement noise from the current and voltage sensors will significantly affect the final IC curve. To make the obtained IC curves comparable over the entire battery life, a time interval of a specific length, such as 10 seconds, can be used, and a Kalman filter (KF) can be applied to smooth the calculated IC curve.

[0101] It is worth noting that when applying IC-related features for battery capacity estimation in real applications, the initial SoC level and battery cell temperature can significantly change the characteristics of the constructed IC curve due to battery cell polarization and internal resistance variations. To address the impact of this variation, initial charge SoC and / or charge temperature-related features can be included in the battery feature pool, allowing the ML algorithm to consider the initial SoC level and / or battery cell temperature when estimating the battery SoH.

[0102] According to some exemplary embodiments, for battery features extracted from the IC curve, the peak value [dQ / dV] and its corresponding voltage level or voltage interval may be selected because they change significantly during battery aging, such as Figure 4C shown.

[0103] Battery feature engineering can be performed in various ways. When feature engineering involves a manual feature selection process to incorporate domain knowledge, there is a risk of selecting irrelevant or redundant features into the battery feature pool. Therefore, a correlation analysis can be performed to rank battery features based on their correlation with the target variable, namely, capacity. Feature checking and control schemes can then be applied to ensure that redundant features are not selected for training the ML algorithm. By doing so, a good compromise between model accuracy and computational complexity can be achieved.

[0104] The correlation analysis may, for example, include performing a Spearman correlation analysis or another type of correlation analysis.

[0105] The estimator design can be implemented in various ways. In one example, the proposed estimator contains a capacity estimation model developed from time series data, where KF is employed to optimally fuse all model results from multiple machine learning models and provide a final capacity estimate. Figure 5A SoH estimation flow chart according to this exemplary embodiment is shown in FIG.

[0106] In this exemplary embodiment, three evaluation matrices are applied to quantify the estimation performance, namely the mean absolute percentage error (MAPE), the root mean square percentage error (RMSPE), and the 95% percentile confidence interval of the estimation results. They are mathematically defined as:

[0107]

[0108] where Q i is the measured capacity value, is the estimated capacity value, N is the total number of samples in the test set, and CI is the 95.4% probability of the Gaussian distribution, and the covariance is given by express.

[0109] Since the estimation results can imply decisions, corrective actions, or advancement to prognostic steps, an assessment of the uncertainty of the estimation becomes essential.

[0110] In an exemplary embodiment, four ML algorithms were selected to develop a model for capacity estimation, two of which were probabilistic and two were frequentist. They were all able to quantitatively propagate their estimation uncertainties to provide confidence intervals for their results. Hereafter, the ML algorithm output, i.e., battery capacity Q, will be referred to as y, and the corresponding features will be represented by x. Random search hyperparameter tuning was applied together with 5-fold cross-validation to find the optimal hyperparameters for each ML algorithm.

[0111] The first ML algorithm can be, for example, Gaussian process regression (GPR). As a nonparametric and probabilistic model, GPR attempts to learn the posterior distribution rather than a single value of the model parameters. In addition, rather than treating the model parameters as random variables as in Bayesian ridge regression (BRR), GPR treats the model function f(x) as a Gaussian process and uses it to calculate the posterior probability P(f(x)|y). Here, f(x) is defined as:

[0112] f(x)~GP(μ(x),κ(x,x′))

[0113] where x and x′ are two arbitrary feature samples, μ(x) is the mean function, and κ(x,x′) is the covariance function, or equivalently, the kernel function. Intuitively, the kernel function determines how strongly the correlation is between two samples x and x′. Given a series of training pairs {X,Y}, where X = {x1,x2,···,xn} and Y = {y1,y2,···,yn}, GPR predicts a new test sample x* by conditioning f(x) on the training data to find the posterior predictive distribution. GPR has high prediction accuracy even with only a relatively small dataset, and GPR is able to quantify the uncertainty of the prediction. These advantages make it a widely adopted method in battery diagnostics. However, for large datasets or high feature dimensions, the computational cost of GPR becomes very expensive due to matrix inversion.

[0114] The second ML algorithm can be, for example, Bayesian Ridge Regression (BRR). Applying Bayesian methods to linear regression models produces the so-called BRR. Instead of treating the coefficient θ as a single variable, BRR assumes that θ is a spherical Gaussian distribution, defined as P(θ) = N(θ; 0, Σ0), with zero mean and covariance Σ0. Here, Σ0 = Iα is chosen to simplify the calculation, where I is the identity matrix and α is a positive hyperparameter. β is defined as the noise variance of the output y and can be regarded as a regularization parameter. Here, α and β can be assumed to be uninformative priors obeying a gamma distribution. These two parameters can be optimized during training by maximizing the marginal likelihood. Using α, β and conjugate priors, the posterior can be explicitly calculated. The simplicity of training and testing and the good interpretability of the estimation results make it a suitable ML algorithm. However, the linear assumption may limit its applicability, especially for complex and nonlinear systems.

[0115] The third ML algorithm can be, for example, random forest regression (RFR). RFR consists of an ensemble of different decision trees and can achieve a good bias-variance tradeoff by aggregating the outputs of each tree. Through randomly selected bootstrap replication, a subset of the training dataset and features is copied to each individual tree. The infinitesimal jack-knife variance estimation method can be adapted to quantify the confidence interval of the estimate of the RFR result.

[0116] The fourth ML algorithm can be, for example, a deep ensemble neural network (DeNN). Deep neural networks (DNNs) have recently attracted great interest from academia and industry due to their outstanding performance in pattern recognition and trend identification for various dynamic and complex tasks. Ensemble-based methods can be used to combine the results of each DNN model and quantify the estimation uncertainty. The final layer can be constructed with two outputs, one is the predicted mean μ(x) and the other is the predicted variance σ(x) 2 . In addition, a negative log-likelihood function can be selected as the loss function. The capacity estimation model is of course not limited to the above examples, but can instead include more or fewer ML algorithms and / or other types of ML algorithms. For example, the capacity estimation model can include a support vector regression type machine learning algorithm, which is a frequentist ML algorithm.

[0117] According to another exemplary embodiment, the capacity estimation model may include a linear regression type machine learning algorithm, which is a frequentist ML algorithm.

[0118] According to another exemplary embodiment, the capacity estimation model may include an Adaptive Boosting (Adaboost) type of machine learning algorithm, which is a frequentist ML algorithm.

[0119] According to another exemplary embodiment, the capacity estimation model may include a machine learning algorithm of the correlation vector regression type, which is a Bayesian ML algorithm.

[0120] For a given battery charging scenario, predefined scenarios S1-S5 may be inappropriate for various reasons. For example, predefined scenarios S1-S5 are all associated with CC-CV charging, so when the charging scenario is of a different kind, such as multi-stage CC charging, pulse charging, dynamic charging, etc., charging scenarios S1-S5 are inappropriate. The same is true when a given battery charging scenario corresponds to abnormal usage behavior (e.g., extremely shallow discharge / charge), data corruption, communication delays / failures, etc. In this case, none of the charging scenarios S1-S5 can be used to extract the selected features, making all the ML models discussed above infeasible.

[0121] Therefore, the method and system may include another category, a sixth charging scenario S6, which covers all remaining scenarios not covered by S1 to S5. Thus, all existing charging situations are divided into six different and mutually exclusive scenarios S1 to S6.

[0122] Specifically, the method and system for estimating the health status of a rechargeable battery according to the present disclosure can be supplemented by a capacity prediction model based on histogram data, which is immediately triggered when a given battery charging condition does not fall into any charging scenario covered by S1 to S5.

[0123] The capacity prediction model based on histogram data enables the capacity monitoring task at the current time step k to be viewed as a prediction problem rather than an estimation problem, using historical capacity estimates up to k-1 and features extracted from usage data.

[0124] In other words, in order to be able to use raw data in any format collected under various operating conditions to predict battery aging trajectory and lifespan online, a machine learning framework based on histogram data can be implemented. This framework is well suited for the capacity prediction task and is therefore recursively implemented whenever charging scenario S6 is triggered.

[0125] Figure 6 The schematic layout and flow chart of the battery SoH estimator design, which includes two types of internal models, are shown in Figure 2. One is a capacity estimation model developed from time series data, and the other is a capacity prediction model developed from histogram-based data. A KF is then employed to optimally fuse all model results and provide the final capacity estimate.

[0126] End-user usage data, such as cumulative discharge / charge energy throughput, discharge range, battery cell temperature, charge current, discharge current, voltage, vehicle parking time, and state of charge, whether in the form of a time series or a histogram of any dimension, are converted into one-dimensional (1D) histogram data. A set of statistical characteristics of the constructed 1D histogram are then extracted and used as candidate initial features.

[0127] Specifically, Figure 7 An exemplary embodiment of a flow chart for performing feature construction based on histogram-based data is schematically shown, starting with a time series of end-user usage data in the form of charge and discharge currents. A histogram transformation is then performed to transform this data into one-dimensional (1D) histogram data. Figure 7 In , the discharge current and the charge current are displayed as separate one-dimensional (1D) histograms. Then, a set of statistical characteristics of the constructed 1D histograms are extracted and used as candidate initial features. Figure 7 In the embodiment of the present invention, a set of statistical characteristics are extracted for each of a one-dimensional (1D) discharge current histogram and a one-dimensional (1D) charge current histogram.

[0128] User data can be in any format and does not need to be in time series. For example, Figure 8 Schematically shows the reference Figure 7The same set of statistical properties are extracted as described, but here starting with end-user usage data in the form of 2D histograms, which undergo a histogram transformation to transform the data into one-dimensional (1D) histogram data, and then extracting a set of statistical properties of the constructed 1D histogram to use as candidate initial features.

[0129] Thereafter, feature engineering can be performed to identify a set of final selected battery features. Feature engineering can, for example, include correlation analysis to determine the correlation between each feature and the system output, i.e., the change in battery capacity at a defined time interval, as well as the correlation between different features. Specifically, Spearman correlation analysis can be applied to measure the strength and direction of the monotonic association between two variables. Spearman correlation analysis is applied to all features in the cell that are correlated with capacity change to identify and discard redundant battery features.

[0130] The remaining battery characteristics, which are determined by x hd The selected features represented by are used to learn the battery aging behavior. Specifically, the task of battery aging prediction is formulated as a regression problem within the supervised machine learning framework. Figure 9 An example of the entire pipeline to accomplish the task, including an offline path for global model development and an online path for model adaptation with streaming data, is summarized in . The global model is first developed offline from N labeled input-output pairs in the available dataset, generated by a statistically significant number of batteries.

[0131]

[0132] where ΔQ ^ n represents the capacity loss between two consecutive samples of any battery in the dataset.

[0133] Therefore, a global model is developed solely from an offline training dataset involving multiple battery cells of the same type. From these cells, each model essentially attempts to learn the average aging behavior in response to selected features. The model development process includes hyperparameter tuning, method selection, model evaluation, and online deployment.

[0134] For offline regression problems, a suitable machine learning algorithm is used to develop the function f global (). The machine learning algorithm can be, for example, support vector regression (SVR), random forest regression (RFR), Gaussian process regression (GPR) or artificial neural network (ANN).

[0135] After obtaining a reasonable global model in (5), the second step is to adapt it online to any considered single cell, indexed by m∈{1,···,M}. This is done by expressing the decay of cell m as λ m,k f global(Xhd,m,t) To achieve this, where λ m,k is the correction factor. Using the historical capacity estimates stored in the BMS memory (e.g., obtained when any scenario from S1 to S5 is valid), the model is then corrected by applying the correction factor to the battery cell, i.e., λ* m,k f global (x hd,m,t ), and the global model, i.e., f global (x hd,m,t ), a trade-off is made between determining a personalized model of the battery cell for the decay capacity of battery m, where λ* is optimized online m,k , to obtain the best possible fit to the historical data, i.e. The capacity decay from time step k-1 to any future time step t≥k is expressed as:

[0136] ΔQ^ m,t =(1-w* k,t )f global (x hd,m,t )+w* k,t λ* m,k f global (x hd,m,t ) (6)

[0137] where w* k-1 is the weight coefficient calculated offline to estimate the global value f global (x hd,m,t ) and the individually corrected estimate λ* m,k f global (x hd,m,t ) to make the best compromise between them. By assigning t to k+1 in the above formula (6), the following step capacity prediction model is obtained:

[0138]

[0139] During the online deployment process, the subscript m will be removed for any battery.

[0140] Figure 10The complete workflow of the proposed online adaptive algorithm for battery capacity prediction is schematically illustrated in Figure 1. Overall, the proposed adaptive algorithm operates in a closed loop that not only considers the aging characteristics of the battery cells in the database but also explicitly includes the properties and operating conditions of the considered battery cells. Consequently, issues such as cell variation, measurement noise, and interference are systematically addressed.

[0141] The ML models used to provide SoH estimation in combination with any of the charging scenarios S1-S5 discussed above have their advantages and disadvantages. The best performing algorithm may vary depending on different data sets and operating conditions. Therefore, fusing the results of all algorithms can give a more accurate and reliable estimate of the battery capacity. A Kalman filter (KF) is used to optimally combine the above estimation model and prediction model. The overall capacity estimation flow chart is shown in Figure 11 Shown in.

[0142] By defining the process noise as w and the measurement noise as v, the dynamic system of battery capacity is formulated as:

[0143] Q k =Q k-1 +ΔQ k-1 +w k-1 (9)

[0144] y k =CQ k +v k (10)

[0145] When the true capacity is not measured during operation, the system output y is a vector of estimated results from the above ML model and is defined by:

[0146] y=Q^ GPR Q^ BRR Q^ RFR Q^ DeNN T (11)

[0147] Means C = 1111 T .

[0148] Σ w and Σ v are the covariance of process noise and the covariance of measurement noise, respectively. The covariance of measurement noise consists of a quantification of the uncertainty of the estimates using the four ML models.

[0149] Assumption 1: The noise w and v are uncorrelated.

[0150] Justification 1: Using probabilistic or frequentist ML algorithms to develop estimation models from time series features. In contrast, using RFR derives prediction models from correlated histogram-based features. Therefore, the corresponding noise terms w and v are naturally uncorrelated when the inputs, ML models, and training procedures are different.

[0151] Assumption 2: w and v are zero-mean, white Gaussian noise.

[0152] Argument 2: All estimates of the system output y are independent and random, so the system input ΔQ is obtained k Therefore, we can assume that vk and wk are at least close to white noise.

[0153] Based on the nominal model of the dynamic system (9)-(10), the standard KF is designed, that is,

[0154]

[0155] where P is the state covariance matrix, the superscripts - and + denote the prior and posterior, respectively, and K k is the Kalman gain.

[0156] Although each of the estimation and prediction models used can provide an estimate of capacity based on assumptions 1 and 2, the KF optimally fuses the estimates from each model, thereby improving the estimation accuracy and robustness to noise.

[0157] During online deployment, the ground truth of battery capacity is rarely available from current BMS. Then, KF will work as an open-loop filter in most cases to fuse the estimated result, i.e., yk defined in (11), and the predicted result Q according to the uncertainty of the estimated result. ^ -k.

[0158] All adopted ML algorithms adopt random search hyperparameter tuning, and the corresponding hyperparameter ranges are shown in Table 3.

[0159]

[0160] Table 3: Ranges of hyperparameter values used in random search hyperparameter tuning

[0161] According to the above feature selection and engineering process, the features selected for each charging scenario of NMC battery cells are Figures 12A-12E It is worth noting that when expressed as When an IC peak exists in any given charging situation, it will always be selected and ranked highest, as shown by S1, S2, and S3. This means that This finding is consistent with previous ICA studies, which showed that is a good indicator for battery health diagnosis. Another finding is that when using partial charge cases, the initial SoC level is selected as one of the key features (see subfigures S3 and S5). This is expected because even if the battery experiences the same current and operating temperature, changing the initial SoC will significantly change the corresponding voltage behavior due to polarization effects. It also verifies the above argument about maintaining the practicality of the SoH estimator developed based on partial charge cases. Temperature-dependent features are also selected, and they overlap heavily between different charging scenarios. This result reflects the fact that temperature is an important stress factor for battery degradation.

[0162] Sorting from S1 to S6, the available charging situations become narrower, and correspondingly, the available features in the pool that meet our specified selection criteria become fewer, which means that the situation for the ML model to learn battery characteristics is more severe.

[0163] To verify the efficacy of the developed models, their performance was first verified for each charging scenario in laboratory tests, i.e., they were applied to data from a battery cell that was subjected to the same charging scenario throughout its lifecycle. The partitioning of the training and test datasets was different for the six considered charging scenarios. For example, for S1, part of the charging data could not be used to train the ML model. Specifically, Figure 13 Table 4 shows the capacity estimation results of NMC batteries under different charging scenarios. Table 4 shows the results of different SoH estimation algorithms for NMC batteries under charging scenarios S1-S6.

[0164]

[0165] Table 4

[0166] First, the estimation errors of different SoH estimation algorithms are quantitatively studied for NMC batteries, and the results are listed in Table 4. It can be seen that from S1 to S5, all the estimated ML models and KFs achieve reasonable estimation values, with the best case MAPE of 0.629% and the worst case MAPE of 2.27%. When the battery operates at S6 throughout its life cycle, although this is a very rare case for EV applications, the proposed model fusion method can still estimate the capacity trajectory with a MAPE of 3.899%. It is worth mentioning that at S6, the estimation model (10) derived from GPR, BRR, RFR and DeNN is not feasible, so the capacity estimation can only be performed by the prediction model (9). The relatively poor estimation results are simply attributed to the fact that a very small and sparse training dataset is used. It can also be observed from the table that for the six scenarios, the estimation tends to become more accurate when there are more qualified features in the pool. This means that capacity estimation should always be performed under the scenario with the richest features.

[0167] For example, Figures 14A-14C The estimation results of randomly selected NMC battery cells in scenario S1 are shown.

[0168] The results in Table 4 also verify the superiority of the proposed KF-based fusion method, which generally performs better or as good as the best performing individual models. Looking more closely, Figure 13 It shows that under the first five scenarios, KF rarely exceeds ±5% error, while under S1-S2, all estimation results are within ±2.5% error; from the instantiation of the random selected battery cell under S1 Figures 14A-14C Judging from the estimated capacity trajectory of , KF follows the measured capacity better than the best individual ML model (i.e., RFR in this case). The numerical results obtained are largely consistent with Remark 1. However, with the adopted battery dataset, KF does not achieve global optimality. This is because the zero-mean part of the measurement noise v in Assumption 2 does not hold. Figure 14C As shown, for the four estimated models in (10), the mean values deviate slightly from zero and lie in the range [0.4%, 0.6%]. The remainder of Assumption 2 is valid since v and w have (approximately) Gaussian distributions, respectively.

[0169] All the models developed here can provide 95th percentile confidence intervals for their estimates, thus providing valuable information for predictive battery maintenance and usage optimization. Moreover, by fusing the estimates from all the individual models using the KF, the confidence intervals are substantially tightened; in other words, the uncertainty is significantly reduced. Figure 15 The standard deviation of the results from the best model, the worst model, and the model ensemble of randomly selected cells is shown. Figure 15shows the standard deviation σ of the estimates from KF KF is always smaller than the standard deviation of any individual model. After metric 7, the standard deviations of the best and worst individual models are σ KF 2 times and 5 times.

[0170] The present disclosure is specifically designed to provide reliable SoH estimation results during daily vehicle use. In practice, it's quite rare for a battery to experience only a single charge case throughout its lifetime. Therefore, it's worth exploring the estimation performance of a specific battery across a variety of charge cases. In the current testing of the method, none of the battery cells in the database were cycled using multiple charge cases. Therefore, test results from cells undergoing full CC-CV charging were used, and then their charge cases were sliced to simulate partial charges. By doing so, the effectiveness of different models under actual charge conditions can be verified and compared. While the corresponding results may not accurately reflect actual use, they serve to test the SoH estimation method and, if implemented in actual vehicle use, would certainly yield realistic partial charge cases.

[0171]

[0172] Table 5: SoH estimation results for various real-world charging scenarios

[0173] In this SoH method test, the charging condition of each battery cell should be periodically rotated between the six scenarios throughout its life cycle. The detailed rotation protocol and results are given in Table 5. The corresponding features are extracted according to the availability of each scenario. Specifically, for charging scenario S6, since no time series features are available, it is not feasible to estimate the capacity value using a separate estimation model (10). In this case, whenever charging scenario S6 is triggered, the capacity remains unchanged relative to its previous time step estimate.

[0174] Table 5 shows that within each protocol, the KF consistently provides better estimation results than any individual ML model. When S6 is activated more frequently, the estimation results of the individual models generally deteriorate. Most of their estimates deviate from the measured values with an RMSPE greater than 2%, which is unacceptable for vehicle applications. In contrast, the KF remains highly reliable and consistently follows the ground truth with an RMSPE of approximately 1%. Comparing different protocols and comparing the results under laboratory testing with vehicle usage reveals that the KF's advantage becomes more pronounced, particularly when charging scenario S6 occurs more frequently. This is because the KF combines a capacity prediction model with four capacity estimation models. Even when the estimation model loses its effectiveness under charging scenario S6, the prediction model developed from the histogram data under arbitrary operating conditions consistently estimates capacity, as does the KF, which acts as an open-loop model-based predictor based on battery usage information. These results confirm the necessity of using the KF in real-world vehicle battery applications.

[0175] Figure 16A The results of capacity estimation of randomly selected NMC type batteries under actual charging cycles are shown. Figure 16B and Figure 16C As shown in the detailed estimation results, the best performing individual model, GPR in this case, still has several estimation points that violate the ±2.5% error range, while KF manages to keep all estimation points within this error range. Underestimating capacity will lead to conservative use, while overestimating capacity may lead to misuse and may even cause thermal safety issues under extreme conditions. Figure 16D As shown in Figure 2, KF has a narrower probability distribution of estimation error, which is closest to zero among all estimators. This means that KF is less likely to under- or over-estimate the capacity than all the individual methods.

[0176] Since the voltage window plays an important role in the developed method, especially when dealing with partial charging scenarios where it is impossible to select all voltage windows for feature extraction. In view of this, the impact of different voltage windows on the estimation performance is studied. In order to make a fair comparison, only voltage window related features and temperature related features are included. The detailed numerical results are shown in Table 6, where only battery cells that have experienced charging scenario S1 are used in order to access any voltage window. In summary, with a voltage interval of 100mV, the results of KF are acceptable and meet the average industry standard of 2% MAPE, demonstrating the practicality of the proposed model fusion method. It is worth mentioning that the results obtained for different scenarios in Table 4 have adopted a voltage window of 3.65-3.75V for the voltage window related features. It can be seen from Table 6 that when the voltage window of 4.05-4.15V is selected, the designed estimators, including KF and separate models, will perform better. One assumption for this superiority is because this voltage window happens to cover the IC peak, as can be seen from Figure 4C As can be seen in the figure, the transition from CC to CV charging is also approaching. However, many vehicle uses may not fall into such a voltage window, which means that fewer data samples are available.

[0177] Table 6: SoH estimation results using different voltage windows

[0178]

[0179] To verify the applicability of the developed method to different types of battery cell chemistries, it was also applied to NCA type batteries and the results are shown in Table 7. In general, the conclusions drawn for NMC type batteries are also valid here. This shows that the developed method is general enough to cover batteries with different chemistries and cycling conditions. In contrast, the estimation accuracy is slightly worse than the results achieved for NMC type batteries, especially for the partial charge case. One hypothesis for this is that the number of data samples in the training set used for the NCA type battery dataset is smaller than the number used for NMC type batteries. Another possible reason may be that the differences between battery cells undergoing similar cycling conditions are generally larger than those in the NMC type battery dataset, which naturally increases the estimation challenge.

[0180]

[0181] Table 7: Results of different SoH estimation algorithms for NCA batteries

[0182] In summary, dynamic operating conditions, random user behavior, and cell-to-cell variations pose significant difficulties and challenges to high-fidelity SoH estimation for actual EV batteries. To address this issue, this disclosure describes a practical method for accurately estimating SoH under arbitrary usage conditions. The technical contributions come from three aspects. First, all possible charging situations are divided into six feasible and mutually exclusive scenarios. For each scenario, relevant features are extracted from time series or histogram data. Second, for each scenario, a Bayesian ML algorithm and a frequentist ML algorithm are used separately to derive a SoH estimation model using time series data. Finally, based on the SoH prediction model previously derived from histogram data, a KF is applied to systematically fuse the results of all models.

[0183] Experimental data from different types of batteries demonstrate that the proposed model fusion approach is able to significantly improve estimation accuracy and robustness, while significantly tightening the confidence intervals of the estimation results. For example, under real-world operating protocols, the estimated values obtained for the entire life cycle of an NMC battery cell have a MAPE of 0.631%, and for all studied protocols, the error is less than 0.8%.

[0184] Thus, typically, the input data obtained by the method according to the present disclosure is obtained in conjunction with, or more particularly primarily during, a battery charging phase.

[0185] In addition, reference Figure 17A ,refer to Figure 2 The described method typically also includes a setup phase 11 performed before the first step S100 of obtaining input data, wherein the setup phase 11 includes a step S40, which includes selecting a suitable machine learning algorithm and training the machine learning algorithm based on offline battery data. Therefore, the setup phase 11 is typically performed offline before starting real-time use, that is, during the vehicle development phase.

[0186] The offline battery data should preferably include relevant battery characteristics, such as time series of charge current, battery voltage, individual battery cell voltage levels, discharge current, battery temperature, and the like.

[0187] The time series of offline data may include at least 100 charging events, specifically at least 500 charging events, and more specifically at least 1000 charging events, wherein the charging event may be a full charge, eg, 0-100% SoC, or a partial charge, eg, 30-80% SoC.

[0188] The time series of offline data may extend continuously or discontinuously over a battery usage period of at least 1 week, specifically at least 1 month, and more specifically at least 6 months.

[0189] The offline battery data should preferably be derived from battery cells of the same or at least similar type as the battery cells intended to be monitored by the present method and system to achieve more reliable estimation results by the machine learning algorithm.

[0190] The input data obtained in the first step S100 can also be referred to as online data because it originates from measurements of the battery to be monitored by the present battery SoH method. Initially, no online data from the battery to be monitored is available because the battery is new. However, over time, if the battery is used regularly, online data from the battery to be monitored will slowly build up.

[0191] The input data may be obtained, for example, from battery sensors, such as one or more current sensors, voltage sensors, and / or temperature sensors associated with the battery or an electrical device connected to the battery.

[0192] The online battery SoH estimation performed by the method / system according to the present disclosure can be implemented in an electronic control unit adjacent to and / or associated with the battery. In other words, the online battery SoH estimation can be provided more or less in real time, for example, in conjunction with a recent charging event. However, according to an alternative, the online battery SoH estimation performed by the method / system according to the present disclosure can be implemented in a remote cloud computing server or a similar type of remote server solution. Such a solution would then require data transmission from the battery controller to the remote server. Furthermore, the online battery SoH estimation can alternatively be performed after a certain delay, such as within 24 hours of the recent charging event.

[0193] refer to Figure 17B According to some exemplary embodiments, the setting phase 11 of the method may further include a step S20A of selecting a predetermined set of battery characteristics. The selection step includes selecting one or more of the following battery characteristics based on recent battery charging events: a voltage profile; a current profile; a time interval between predetermined voltage windows; a signal strength over time, which is calculated as where s(t) is the signal; the area under the current curve; the area under the voltage curve; the slope of the voltage curve; the slope of the current curve [which can be used as a feature during the CV phase if there is one in the charging case]; the initial SoC; the final SoC, a charge temperature-dependent feature; the peak value of the ramp-up curve; the voltage level of the ramp-up curve at the peak; the final total battery output voltage; the final individual cell voltages; and the differential voltage curve.

[0194] refer to Figure 17CAccording to some exemplary embodiments, the setting stage 11 of the method may further include step S10: selecting a set of unique full and / or partial charging scenarios (S1-S5), each charging scenario having a unique charging start position and / or charging end position; and for each of the selected charging scenarios S1-S5, training a machine learning algorithm based on a data set corresponding to the selected charging scenario S1-S5, respectively, wherein the step S100 of obtaining input data includes: determining to which of the unique full and / or partial charging scenarios (S1-S5) the obtained input data corresponds, and wherein the step S200 of applying multiple machine learning algorithms to perform battery health state estimation includes: for each of the machine learning algorithms, applying a machine learning algorithm trained based on data associated with the determined charging scenario (S1-S5) to calculate the battery health state estimation value and the quantitative estimate of the confidence interval / value of the battery health state estimation value.

[0195] Reference again Figure 17C , then, the step S20B of selecting predetermined battery characteristics may include: selecting a predetermined set of battery characteristics indicative of a battery state of health for each charging scenario (S1-S5) in the set of unique charging scenarios (S1-S5), and the predetermined battery characteristics of at least one charging scenario (S1-S5) in the set of unique charging scenarios (S1-S5) are different from the predetermined battery characteristics of another charging scenario (S1-S5) in the set of unique charging scenarios (S1-S5). Thus, the battery characteristics of each charging scenario are typically selected to be more or less optimally adapted to the particular charging scenario. In some exemplary embodiments, each charging scenario has a unique set of selected battery characteristics, and in other exemplary embodiments, some charging scenarios may have the same set of selected battery characteristics, but generally not all of them are the same, depending on each specific implementation.

[0196] According to some example embodiments, step S20B of selecting predetermined battery features may include: for each predetermined battery feature set indicative of a battery health state, first identifying a preliminary battery feature set that collectively indicates a battery health state, and then performing a correlation analysis of the preliminary battery features.

[0197] The correlation analysis may be, for example, a Spearman analysis performed by an algorithm in a computer, etc. The purpose of the correlation analysis is to eliminate redundant features, thereby improving estimation quality and / or computational efficiency.

[0198] As mentioned above, especially Figures 6 to 12E As described, according to some exemplary embodiments, Figure 17DAs shown in , the method may include: step S250, based on the battery input data obtained in step S100, calculating a battery SoH prediction value and a confidence interval / quantitative estimate of the battery SoH prediction value using a machine learning prediction model based on histogram data, and then in step 300, applying the Kalman filter (KF)-based fusion algorithm to combine the battery state of health estimation values from all of the multiple machine learning algorithms and the battery SoH prediction value from the machine learning prediction model based on histogram data to provide a fused battery state of health estimation value. Therefore, even when the most recent charging event does not match any of the predetermined charging scenarios S1-S5, that is, when the ML algorithm in step S200 cannot provide an accurate SoH estimation value, a reasonable, reliable and accurate battery SoH estimation value can still be provided.

[0199] According to some exemplary embodiments, the step of calculating a battery SoH prediction value by a machine learning prediction model based on histogram data may include an offline setting phase, which includes: a step S50 of obtaining historical battery usage data; a subsequent step S60 of converting the battery usage data into a 1D histogram and extracting statistical characteristics from the 1D histogram; then, a step S70 of determining battery characteristics based on the extracted statistical characteristics; and finally, a step S80 of providing a global model by selecting and offline training a machine learning algorithm based on the obtained historical battery usage data, and wherein during the online use of the battery, the step S250 of calculating a battery SoH prediction value by a machine learning prediction model based on histogram data includes: calculating a global battery SoH prediction value based on the global model; and online adjusting the global battery SoH prediction value based on the measured historical battery capacity estimation value of the current battery (5) to provide a final battery SoH prediction value.

[0200] Step 70 of determining battery characteristics based on the extracted statistical characteristics may further include first identifying a preliminary set of battery characteristics that collectively indicate the battery state of health based on the extracted statistical characteristics, and then performing a correlation analysis on the preliminary battery characteristics. As described above, the correlation analysis may be, for example, a Spearman analysis performed by an algorithm in a computer. The purpose of the correlation analysis is to eliminate redundant characteristics, thereby improving estimation quality and / or computational efficiency.

[0201] According to some exemplary embodiments, the set of unique charging scenarios (S1-S5) may, for example, include one or more of the following battery charging scenarios: full constant current (CC)-constant voltage (CV) charging; partial CC-CV charging including starting after the peak of the capacitance (IC) curve and ending with a complete constant voltage (CV) phase; partial constant current (CC)-constant voltage (CV) charging when starting after the peak of the capacitance (IC) curve and ending without a constant voltage (CV) phase; partial constant current (CC)-constant voltage (CV) charging when starting before the peak of the capacitance (IC) curve and ending with a complete constant voltage (CV) phase; partial constant current (CC)-constant voltage (CV) charging when starting before the peak of the IC curve and ending with a complete constant voltage (CV) phase. These battery charging scenarios are believed to represent charging scenarios commonly used in actual use of battery applications, such as, in particular, in actual use of battery electric vehicles.

[0202] In other words, a set of individual charging scenarios (S1-S5) may include a full charging scenario and one or more partial charging scenarios (S1-S5).

[0203] According to some exemplary embodiments, the method includes setting the final battery state of health estimate equal to the battery SoH prediction value derived by the machine learning prediction model based on the histogram data when the obtained input data does not correspond to any of the set of unique charging scenarios (S1-S5). In other words, when the recent charging event includes, for example, pulse charging or dynamic charging, or some other type of charging condition that does not fall into any of the predetermined charging scenarios S1-S5, the resulting battery state of health estimate provided by the method may be set equal to the battery SoH prediction value derived by the machine learning prediction model based on the histogram data, because the estimation model will not be able to provide any reliable estimation results in such charging events.

[0204] refer to Figure 1The present disclosure also relates to a system for real-time estimation of the state of health of a rechargeable battery. The system includes: a rechargeable battery 5; and a set of sensors 12 configured to sense a set of battery features about the rechargeable battery 5. The set of battery features is selected for its capacity to collectively provide an indication of the state of health of the battery 5. The system also includes an electronic control unit 13, i.e., an ECU, which is connected to the set of sensors 12 and is configured to obtain input data related to a predetermined set of battery features and apply multiple, in particular four, machine learning algorithms to perform real-time battery state of health estimation, wherein each machine learning algorithm is configured to calculate a battery state of health estimate and a confidence interval / value of the battery 5 state of health estimate based on the input data of the obtained battery features. Finally, the ECU is configured to apply a Kalman filter (KF)-based fusion algorithm to combine the health state estimates from all of the multiple machine learning algorithms to provide a fused health state estimate.

[0205] refer to Figure 1 , the present disclosure also relates to a vehicle 1 comprising the above-described system.

[0206] The present disclosure also relates to an electronic control unit 13 comprising a processor configured to execute the above-described method.

[0207] Furthermore, the present disclosure relates to a computer program comprising instructions which, when executed by a computer, cause the computer to perform the steps of the above-mentioned method.

[0208] The present disclosure has been presented above with reference to specific embodiments. However, other embodiments than those described above are also possible and within the scope of the present disclosure. Within the scope of the present disclosure, method steps different from those described above may be provided, and the method may be performed by hardware or software.

[0209] The methods disclosed herein can be implemented in a general-purpose computer, a processor, or a processor core. Suitable processors include, by way of example, a general-purpose processor, a special-purpose processor, a conventional processor, a digital signal processor (DSP), a plurality of microprocessors, one or more microprocessors associated with a DSP core, a controller, a microcontroller, an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA) circuit, any other type of integrated circuit (IC), and / or a state machine.

[0210] The methods or flow charts provided herein can be implemented in a computer program, software, or firmware that is incorporated into a computer-readable storage medium for execution by a general-purpose computer or processor. Examples of computer-readable storage media include read-only memory (ROM), random access memory (RAM), registers, cache memory, semiconductor storage devices, magnetic media such as internal hard disks and removable disks, magneto-optical media, and optical media such as CD-ROM disks and digital versatile disks (DVDs).

[0211] Thus, according to an exemplary embodiment, a non-transitory computer-readable storage medium is provided having one or more programs stored thereon, the one or more programs being configured to be executed by one or more processors of a system for estimating the state of health of a rechargeable battery, the one or more programs comprising instructions for executing a method according to any of the above-described embodiments. Alternatively, according to another exemplary embodiment, a cloud computing system may be configured to perform any of the method aspects presented herein. The cloud computing system may include distributed cloud computing resources that collectively execute the method aspects presented herein under the control of one or more computer program products. In addition, the processor may be connected to one or more communication interfaces and / or sensor interfaces for receiving and / or sending data with an external entity, such as a sensor disposed on a surface of the vehicle, an off-site server, or a cloud-based server.

[0212] The processor associated with the system for estimating the state of health of a rechargeable battery may be or include any number of hardware components for performing data or signal processing or for executing computer code stored in memory. The system may have associated memory, and the memory may be one or more devices for storing data and / or computer code for performing or facilitating the various methods described in this specification. The memory may include volatile memory or non-volatile memory. The memory may include a database component, an object code component, a script component, or any other type of information structure for supporting the various activities of this specification. According to exemplary embodiments, any distributed or local memory device may be used with the systems and methods of this specification. According to exemplary embodiments, the memory is communicatively connected to the processor (e.g., via a circuit or any other wired, wireless, or network connection) and includes computer code for performing one or more processes described herein.

[0213] It should be understood that the foregoing description is merely exemplary in nature and is not intended to limit the present disclosure, its application, or uses. Although specific examples have been described in the specification and shown in the drawings, those skilled in the art will understand that various changes may be made and equivalents may be substituted for elements thereof without departing from the scope of the present disclosure as defined in the claims. In addition, modifications may be made to adapt a particular situation or material to the teachings of the present disclosure without departing from the basic scope thereof. Although the above discussion is based on Figures 17A-17E Although the method of the present invention is described in the flowchart of the present invention, it should be understood that one or more operations can be omitted from the method discussed. In addition, the operations can be performed in any order and the order provided is not necessarily implied. Instead, the method discussed is merely one embodiment of the present disclosure as intended.

[0214] Therefore, it is intended that the present disclosure not be limited to the particular examples shown in the drawings and described in the specification as the best mode presently contemplated for carrying out the teachings of the present disclosure, but that the scope of the present disclosure will include any embodiment falling within the above description and the appended claims. Reference signs mentioned in the claims shall not be construed as limiting the scope of the subject matter protected by the claims, and their sole function is to make the claims easier to understand.

[0215] Reference numerals

[0216] 1. Vehicle

[0217] 2. Front wheel

[0218] 3. Rear wheel

[0219] 4. Passenger compartment

[0220] 5. High voltage battery

[0221] 6. Motor

[0222] 7. Charging port

[0223] 8. Low voltage level

[0224] 9. High voltage level

[0225] 10.IC peak

[0226] 11. Setting the Stage

[0227] 12. Sensors

[0228] 13.ECU

[0229] The peak on the IC curve.

[0230] ΔI CV The slope of the current during the CV charging phase.

[0231] Δt is the time spent in the defined voltage window.

[0232] ΔV is the slope of the voltage within a defined voltage window.

[0233] E U The energy of the voltage signal.

[0234] SoC0 Initial SoC level when charging starts.

[0235] T max Maximum temperature during charging.

[0236] T min Minimum temperature during charging.

[0237] t cc The time spent in CC charging phase.

[0238] The IC has a voltage corresponding to the peak value.

Claims

1. A method for estimating the state of health of a rechargeable battery (5), the method comprising: obtaining input data of a predetermined set of battery characteristics that collectively indicate a state of health of a battery (5); Applying a plurality of machine learning algorithms to perform battery (5) state of health estimation, wherein each machine learning algorithm calculates a battery (5) state of health estimate and a confidence interval / quantitative estimate of the battery (5) state of health estimate based on input data obtained of battery characteristics; as well as A Kalman filter-based fusion algorithm is applied to combine the health state estimates from all of the plurality of machine learning algorithms to provide a fused health state estimate.

2. The method according to claim 1, wherein The input data obtained is acquired in conjunction with the battery charging phase.

3. A method according to any one of the preceding claims, wherein The method further comprises a setup phase performed before the step of obtaining input data, wherein the setup phase comprises training a machine learning algorithm.

4. A method according to any one of the preceding claims, wherein The predetermined battery characteristics include one or more of the following battery characteristics based on recent battery charging events: a voltage profile; a current profile; a time interval between predetermined voltage windows; a signal strength over time, which is calculated as Where s(t) is the signal; the area under the current curve; the area under the voltage curve; the slope of the voltage curve; the slope of the current curve; the initial SoC; the final SoC; and the charging temperature-related characteristics. The peak value of the accelerometer curve; the voltage level of the accelerometer curve at the peak value; the final total battery output voltage; the final individual battery cell voltage; and the differential voltage curve.

5. The method according to any one of the preceding claims, in, The method further includes a setup phase, the setup phase comprising: selecting a set of unique charging scenarios (S1-S5), each unique charging scenario having a unique charging start location and / or charging end location; and training a machine learning algorithm for each of the selected charging scenarios S1-S5 based on a data set corresponding to the selected charging scenario S1-S5, The step of obtaining input data includes: determining which of the unique charging scenarios (S1-S5) the obtained input data corresponds to, and The step of applying multiple machine learning algorithms to estimate the health state of the battery (5) includes: for each of the machine learning algorithms, applying a machine learning algorithm trained based on data associated with the determined charging scenario (S1-S5) to calculate the health state estimate value of the battery (5) and the confidence interval / value of the health state estimate value of the battery (5).

6. The method according to claim 5, wherein: A predetermined set of battery characteristics indicative of a state of health of a battery (5) is selected for each charging scenario (S1-S5) in the set of unique charging scenarios (S1-S5), and wherein the predetermined battery characteristics of at least one charging scenario (S1-S5) in the set of unique charging scenarios (S1-S5) are different from the predetermined battery characteristics of another charging scenario (S1-S5) in the set of unique charging scenarios (S1-S5).

7. A method according to any one of the preceding claims, wherein Each of the predetermined battery characteristic sets indicative of the state of health of the battery (5) is determined by: First, a preliminary set of battery characteristics that collectively indicate the state of health of the battery (5) is identified, Perform correlation analysis on preliminary battery characteristics.

8. The method according to any one of the preceding claims, further comprising: Calculating a battery SoH prediction value and a confidence interval / quantitative estimate of the battery SoH prediction value by a machine learning prediction model based on the histogram data, and The Kalman filter-based fusion algorithm is applied to combine the battery state of health estimates from all of the plurality of machine learning algorithms and the battery SoH prediction value from the histogram data-based machine learning prediction model to provide a fused battery state of health estimate.

9. The method according to claim 8, wherein The step of calculating the battery SoH prediction value by a machine learning prediction model based on histogram data includes a setup phase, wherein the setup phase includes: Get historical battery usage data, Converting battery usage data into a 1D histogram and extracting statistical features from the 1D histogram, Determine battery characteristics based on extracted statistical properties, Providing a global model by selecting and offline training a machine learning algorithm based on the acquired historical battery usage data, And wherein, during the online use of the battery, the step of calculating the battery SoH prediction value by using a machine learning prediction model based on histogram data includes: Calculating a global battery SoH prediction value based on the global model; and The global battery SoH prediction value is adjusted online based on the measured historical battery capacity estimate value of the current battery (5) to provide a final battery SoH prediction value.

10. The method according to any one of claims 5 to 9, wherein The group of unique charging scenarios (S1-S5) includes one or more of the following battery charging scenarios: full constant current (CC)-constant voltage (CV) charging; partial CC-CV charging that starts after the peak of the capacitance (IC) curve and ends with a complete constant voltage (CV) stage; partial constant current (CC)-constant voltage (CV) charging that starts after the peak of the capacitance (IC) curve and ends without a constant voltage (CV) stage; partial constant current (CC)-constant voltage (CV) charging that starts before the peak of the capacitance (IC) curve and ends with a complete constant voltage (CV) stage; partial constant current (CC)-constant voltage (CV) charging that starts before the peak of the IC curve and ends with a complete constant voltage (CV) stage.

11. The method according to any one of the preceding claims 8 to 10, comprising: When the obtained input data does not correspond to any charging scenario in the set of unique charging scenarios (S1-S5), the battery state of health estimate is set equal to the battery SoH prediction value derived by the machine learning prediction model based on the histogram data.

12. A system for estimating the state of health of a rechargeable battery (5), the system comprising: Rechargeable battery (5); a set of sensors configured to sense a set of battery characteristics about the rechargeable battery (5), wherein the set of battery characteristics collectively indicate a state of health of the battery (5); An electronic control unit is connected to the set of sensors and is configured to: obtaining input data associated with a predetermined set of battery characteristics; Applying a plurality of machine learning algorithms to perform battery (5) state of health estimation, wherein each machine learning algorithm calculates a battery (5) state of health estimate and a confidence interval / quantitative estimate of the battery (5) state of health estimate based on the obtained input data of the battery characteristics; and A Kalman filter-based fusion algorithm is applied to combine the health state estimates from all of the plurality of machine learning algorithms to provide a fused health state estimate.

13. A vehicle comprising the system according to claim 12. 14 . A data processing control unit, comprising a processor, wherein the processor is configured to execute the steps of the method according to claim 1 .

15. A computer program comprising instructions which, when executed by a computer, cause the computer to perform the steps of the method according to any one of claims 1 to 11.

Images