Method and apparatus for training an artificial intelligence-based state model

Abstract

A computationally implemented method for training a state model for determining a state of an electrical energy storage of a device operated with an electrical energy storage, comprising: providing a data-based state model which assigns a modeled state quantity, in particular an aging state and a state uncertainty measure, of the energy storage to an operating feature point; providing a database with a course of real operating feature points of a multitude of energy storages over successive evaluation time periods; training or updating the state model from a training data set, the training data set being generated as follows: determining an improvement measure for each current operating feature point of the energy storage determined at a current evaluation time period, which states how the state uncertainty at all real operating feature points improves when the state quantity for the respective current operating feature point is determined; selecting the operating feature point with the best improvement measure; determining the corresponding state quantity at the selected operating feature point as a label; generating the training data set from the selected operating feature point and the determined corresponding state quantity.

Description

Technical Field

[0001] This invention generally relates to characterizing the system state of an electric energy storage device, such as an electric energy storage device in an electrically driven device, such as a motor vehicle, especially an electric vehicle or a hybrid vehicle, and also to measures for determining the state of an electric energy storage device, such as a device battery pack or fuel cell system. Background Technology

[0002] For off-grid-locked energy supply of technical equipment, electric energy storage devices, such as battery packs, are typically used. Energy storage devices, in the sense described below, also include energy converters, such as fuel cell systems. For example, electric motor vehicles are powered by means of, for instance, vehicle battery packs or fuel cell systems. These vehicle battery packs or fuel cell systems provide electrical energy for operating the vehicle system, and especially the drive system. The aging condition of an electric energy storage device deteriorates significantly over its lifespan, impacting its maximum storage capacity. The aging of a device's battery pack depends on the individual load on the device's battery pack, i.e., on the user's usage behavior and the type of device's battery pack.

[0003] While it's possible to determine the instantaneous aging state using purely physical aging models based on historical operational parameter changes, such models are typically inaccurate. This inaccuracy of conventional aging models makes predicting the aging state change process difficult. However, predicting the aging state change process of a vehicle battery pack is an important technical parameter because it enables the economic evaluation of the remaining value (Restwert) of the vehicle battery pack.

[0004] Other states of the vehicle battery pack, such as the state of charge, cannot usually be determined with high precision using conventional physical models.

[0005] Not only physical state models, but also data-based or hybrid state models require accurate state labels for training data, as these labels form the basis for model parameterization. The adoption of continuous label generation following the use of (Nutzungsaufnahme) labels in applications enables the continuous improvement of state model parameterization to determine and predict the state parameters to be represented. Summary of the Invention

[0006] According to the present invention, a method for providing training data for training a state model as described in claim 1 is provided, wherein the state model is used to model the state of an electric energy storage device, and a corresponding apparatus according to the parallel independent claims is provided according to the present invention.

[0007] Other configuration options are described in the dependent claims.

[0008] According to the first aspect, a computer-implemented method is set up for training a state model, wherein the state model is used to determine the state of an electric energy storage device operating as an electric energy storage device, the method comprising the following steps: - Provides a data-based state model that assigns modeled state parameters of the accumulator to operating feature points, especially the aging state of the accumulator and state uncertainty measures. - Provide a database containing the variation of real operating characteristic points from a large number of energy storage devices for successive evaluation time periods; - Train or update the state model based on at least one training dataset; The at least one training dataset is generated using the following steps: O determines an improvement metric for each of the current operating characteristic points of the energy storage device for the current evaluation time period, wherein the improvement metric describes how the state uncertainty at all real operating characteristic points is improved when the state parameters for the corresponding current operating characteristic point are determined. O Select at least one running feature point with the best improvement metric; O determines the corresponding state parameter at the at least one selected operating feature point as a label; O generates the at least one training dataset from the at least one selected running feature point and the determined corresponding state parameters.

[0009] Furthermore, the improvement metric for each of these current operating feature points can be approximately determined through simulation by updating the provided state model using a simulation training dataset created for the corresponding current operating feature point to obtain a simulation state model. Simulation state uncertainty is then determined for each of the operating feature points in the changing process of these operating feature points of the real energy storage unit using the simulation state model, and the improvement metric is determined based on the simulation state uncertainty.

[0010] In particular, the improvement metric is determined as the sum of the simulation state uncertainties, wherein at least one of these simulation state uncertainties is reduced before determining the improvement metric, in particular by means of at least one pre-given domain-based rule.

[0011] Furthermore, the state model can correspond to an aging state model for providing an aging state based on operating characteristics or a charging state model for providing a charging state based on the operating characteristics of the energy storage device.

[0012] It can be specified that the data-based state model includes a data-based machine learning model constructed to describe state uncertainties for modeled state parameters, wherein the state model particularly includes a Gaussian process model as a supervised learning model utilizing quantified uncertainties. Alternatively, the supervised learning model can be implemented, in particular, as an ensemble method or a Bayesian neural network.

[0013] These operational characteristics can be derived from the time series of continuously detected operational parameters, wherein these operational characteristics are determined for successive evaluation time periods, and multiple operational characteristics for a specific evaluation time period for a particular accumulator define operational characteristic points.

[0014] For higher-level technical equipment, such as an electric energy storage device in a motor vehicle, multiple states need to be determined in order to enable the higher-level system to operate in an improved or optimized manner. These states are described as state parameters. Therefore, determining, for example, the aging state of the electric energy storage device and the charging state of the rechargeable battery pack are important state parameters, which are crucial for operating technical equipment that operates with said electric energy storage devices and battery packs, such as electric motor vehicles.

[0015] Physics-based methods often cannot model such state parameters with high accuracy. Furthermore, the inaccuracy of state parameter calculations in this case is almost impossible to quantify online, or only difficult to quantify online. Therefore, data-based or hybrid state models are clearly considered for modeling these state parameters. Especially when a large amount of training data is available, such state models can achieve very accurate state calculations or predictions.

[0016] Training such a data-driven state model requires high-precision training data for supervised learning techniques, which must typically be measured or evaluated beforehand. The process of assembling (Zusammenstellen) training data is generally called label generation.

[0017] After training, the data-based state model can be implemented in advance in devices operating with said state model by transmitting the corresponding model parameters. It is also possible that devices with regular connections to the central unit (cloud) regularly receive model results, such as state parameters of the data-based state model or updated model parameters, from the central unit. In particular, when evaluating measured operating parameters in a large number of devices equipped with the same energy storage in the central unit, the data-based state model can be continuously retrained or updated in the central unit, and the corresponding model parameters of the updated or retrained state model can be transmitted back to devices of the same type to improve state predictions there. Preferably, state calculations or state predictions are performed in the central unit (cloud), and these results are provided as states to these devices.

[0018] To provide training data for training a data-based state model, state parameters can be derived from the acquired operating parameters for one of these devices in a costly manner, typically by means of time-consuming diagnostic measurements. These state parameters are determined as labels, particularly in the central unit, so that a training dataset is provided there based on the operating parameters or the operating characteristics derived therefrom and the assigned state parameters (labels). Thus, operating parameters for the operation of a vehicle battery pack in an electrically powered motor vehicle can be continuously transmitted to the central unit, for example, to determine an aging state model in the central unit, and an aging state description can be derived there, for example, by evaluating the determined operating changes, such as a complete charging cycle under defined load and environmental conditions, triggered by diagnostic measurements, thereby providing a training dataset.

[0019] In the initial phase of operating a technical system based on a data-based state model used for estimating state parameters, the data-based state model is typically not yet trained with sufficient accuracy across the entire range of the input data space. This technical system has a central unit and multiple similar devices communicatively connected to the central unit. To continuously improve the performance of the data-based state model during its use, it can be periodically updated in the central unit based on a training dataset determined using operating parameters detected by these devices. It is not possible to easily determine the corresponding access parameters (Zugangsgröße) for the entire range of the input parameter space through measurements from real devices, thus leaving the data-based state model highly inaccurate and potentially unusable in these ranges, where the input parameter space is defined by these operating parameter points or operating feature points.

[0020] The above method now specifies that the data-based state model is evaluated predictively, particularly by means of simulation, to identify at least one operational feature point. The identified operational feature point can serve as the basis for determining the training dataset, which allows the state model to be updated in an optimized manner to improve state uncertainty at real operational feature points. Fleet data can be used for this purpose, which illustrates the modeled changes in state parameters derived from operational parameters with respect to real operational feature points. This is preferably carried out by means of simulation, using a simulated training dataset that presents possible states at the current operational feature point and determines an overall improvement in the reliability / confidence of state predictions. Those operational feature points that minimize overall uncertainty correspond to the identified operational feature points. The identified operational feature points are the starting point for generating new training datasets.

[0021] The determination of the operational feature points for the training dataset to be generated is therefore carried out using an improvement metric, which is obtained by simulating each operational feature point among the current operational feature points determined for each accumulator or for the accumulator in the current evaluation time period. The improvement metric describes how the state uncertainty at all real operational feature points in the database is improved when a simulated training dataset with state parameters for the corresponding current operational feature points is determined and the state model is updated accordingly. The improvement metric is determined approximally through the simulation, in which the state model is evaluated, respectively (only for the simulation), based on the added value related to fleet relevance and the accompanying reduction in global model uncertainty, in order to retrain the aging state model.

[0022] The simulated state models can now be used individually to determine the improvement metric as the sum of the simulated state uncertainties. Here, the individual simulated state uncertainties can be reduced according to rules based on domain knowledge, and the corresponding reduced simulated state uncertainties are considered in order to determine the improvement metric.

[0023] Since the current operating characteristic points of a large number of energy storage units are known, this evaluation can be based on reducing the overall model uncertainty of the change process of the actual operating characteristic points of the energy storage units. This reduction can be quantified according to the method described above based on a corresponding improvement metric for each of these current operating characteristic points, wherein the improvement metric describes how the state uncertainty at all actual operating characteristic points, i.e., for the operating characteristic points appearing in actual operation as a whole, is improved when the state parameters for the corresponding current operating characteristic point are determined through label generation. Here, it is preferable to take the uncertain, modeled state parameters of the state model as the labeled state parameters.

[0024] This enables the targeted generation of labels for the accumulator that, based on the specific improvement metric among these accumulators, indicates the reduction as much as possible in the overall model uncertainty of the fleet (with respect to its current operational feature point), thereby quantifying label relevance.

[0025] Therefore, a label generation method for statistics can be used, in which the determination of the corresponding state parameters is performed, especially in combination with domain knowledge-based rules for reducing state uncertainty.

[0026] In particular, the state parameters exhibit dependencies on their time-varying processes in a system-bedded manner, and these dependencies are known and available as domain knowledge. This enables the reduction of high state uncertainty or large confidence intervals in state parameter predictions (modeled state parameters) derived from modeling using state models, by excluding physically impossible ranges of values ​​for the state parameters through, for example, setting upper and lower bounds on the confidence intervals.

[0027] Furthermore, the reduction of state uncertainty at the selected operational feature point can be achieved through at least one rule, wherein the rule depends on the state uncertainty of the operational feature point and the temporal development (Entwicklung) of the operational feature point in an evaluation time period different from the evaluation time period of the selected operational feature point.

[0028] In particular, the reduction of state uncertainty at the selected operating characteristic points can be achieved by limiting the state uncertainty at each of the selected operating characteristic points to an upper or lower bound of a confidence interval, wherein the upper or lower bound is determined by the state uncertainty over the evaluation time period prior to the operating characteristic point, and / or by limiting the state uncertainty at each of these selected operating characteristic points to an upper or lower bound of a confidence interval, wherein the lower or upper bound is determined by the state uncertainty over the evaluation time period following the operating characteristic point. This can be attributed to domain knowledge: slowly time-varying state parameters, such as the aging state of an accumulator, change only inertially.

[0029] Alternatively or additionally, the reduction of state uncertainty at the selected operating feature points can be achieved by limiting the state uncertainty at each of the selected operating feature points using the upper and lower limits of the corresponding confidence intervals, wherein the upper and lower limits are determined by interpolating the upper limits of the confidence intervals for the evaluation time periods preceding and following the relevant operating feature point.

[0030] Alternatively, the determination of the corresponding state parameters can be performed by requesting the technical system assigned to the at least one selected running feature point to perform the measurement of the state parameters, wherein the measured state parameters are used as the label target parameters to be measured after the measurement in order to create a training dataset for updating the state model.

[0031] In the above method, a data-based state model is provided in the central unit for a large number of similar technical devices. This state model can also be implemented as a hybrid state model. The data-based state model is updated and improved by creating additional training datasets through evaluating available training datasets and evaluating the data-based state model in real-world operation of a large number of technical devices. These additional training datasets are suitable for retraining or updating the data-based state model. This allows for retraining of the initially provided data-based state model within the input parameter space, where a high degree of state uncertainty exists.

[0032] It can be stipulated that a large number of technical system operating parameters are transmitted to a central unit, wherein a database is provided, the database having operating characteristic points of a large number of real energy storage devices of technical systems, wherein the method is executed in the central unit, wherein the operating characteristic points are selected from the database.

[0033] According to another aspect, a device for training a state model is provided, particularly a control unit in a central unit that is in communicative connection with a large number of devices having energy storage, wherein the state model is used to determine the state of the energy storage, and the device is configured to: - Provides a data-based state model that assigns modeled state parameters of the accumulator to operating feature points, especially the aging state of the accumulator; - Provide a database containing the variation of real operating characteristic points from a large number of energy storage devices for successive evaluation time periods; - Train or update the state model based on at least one training dataset, wherein the at least one training dataset is generated using the following steps: O determines an improvement metric for each of the current operating characteristic points of the energy storage device for the current evaluation time period, wherein the improvement metric describes how the state uncertainty at all real operating characteristic points is improved when the state parameters for the corresponding current operating characteristic point are determined. O Select at least one running feature point with the best improvement metric; O determines the corresponding state parameter at the at least one selected operating feature point as a label; O generates the at least one training dataset from the at least one selected running feature point and the determined corresponding state parameters. Attached Figure Description

[0034] The following describes several implementation methods in further detail with reference to the accompanying drawings. Among them: Figure 1A schematic diagram of a system for a fleet of vehicles with a large number of vehicles and a central unit for providing a data-based state model is shown. Figure 2 A flowchart illustrating a method for training or updating a data-based state model is shown, in particular in the form of an aging state model for a vehicle battery pack. Figure 3 A diagram illustrating the overall measure of reducing the uncertainty of all determined states at those running feature points by generating labels at those running feature points. Detailed Implementation

[0035] The method according to the invention is described below based on vehicle battery packs, which serve as energy storage devices in a large number of motor vehicles that are similar devices. In these motor vehicles, a data-based aging state model for the corresponding vehicle battery pack can be implemented in a control unit. This aging state model, for example, represents a state model characterizing the internal, non-measurable state of the energy storage device. This aging state model can be continuously updated or retrained in a central unit based on operating parameters from the vehicle battery packs of the fleet.

[0036] The examples above representatively illustrate a wide range of fixed or mobile devices with power supply systems independent of the power grid, such as vehicles, facilities, IoT devices, power supply facilities, and autonomous robots connected to a central unit (cloud) via corresponding communication connections (e.g., LAN, Internet). These state parameters represent parameters that cannot be determined easily and accurately, and simultaneously based on models, in similar devices; rather, they can only be determined through costly calculations, internal, and especially destructive measurements, or after a predetermined operating cycle of the device.

[0037] Figure 1 A system 1 is shown for providing fleet data of motor vehicles 4 of a fleet 3 in a central unit 2. In this central unit 2, the aging process of the vehicle battery packs of the corresponding motor vehicles 4 in the fleet 3 should be predicted based on the fleet data.

[0038] exist Figure 1 The diagram details one of these motor vehicles 4. Each of these motor vehicles 4 has a vehicle battery pack 41 as a rechargeable electric energy storage device, an electric drive motor 42, and a control unit 43. The control unit 43 is connected to a communication module 44, which is adapted to transmit data between the respective motor vehicle 4 and the central unit 2 (cloud). The control unit 43 is also connected to a sensor unit 45, which has one or more sensors for continuously detecting operating parameters.

[0039] The central unit 2 has a data processing unit 21 and a database 22 for storing the aging status and operating parameters of the vehicle battery packs in the multiple vehicles 4 of the fleet 3, wherein the methods described below can be implemented in the data processing unit.

[0040] These vehicles 4 transmit operating parameters F to the central unit 2. These operating parameters at least indicate the parameters upon which the aging state of the vehicle battery pack depends. In the case of the vehicle battery pack, these operating parameters F can indicate instantaneous battery pack current, instantaneous battery pack voltage, instantaneous battery pack temperature, and instantaneous state of charge (SOC). The operating parameters F are detected using a fast time grid from 2Hz to 100Hz, wherein the changes in these operating parameters are periodically transmitted to the central unit 2 in uncompressed and / or compressed form.

[0041] The operational characteristics related to the evaluation period can be generated in the central unit 2 by the operational parameter F, or in other embodiments, the operational characteristics related to the evaluation period can already be generated in the corresponding vehicle 4 by the operational parameter F. To determine the aging state, the evaluation period can range from several hours (e.g., 6 hours) to several weeks (e.g., one month). A commonly used value for the evaluation period is one week.

[0042] Operating characteristics may include, for example, characteristics relating to the evaluation period and / or cumulative characteristics and / or statistical parameters determined over the entire service life to date. Specifically, the operating characteristics may include, for example, histogram data on state of charge changes, temperature, battery pack voltage, battery pack current, cumulative total charge (Ah), average capacity increase during charging (particularly for charging processes where the charge increase exceeds a threshold share of the total battery pack capacity (e.g., 20%), differential capacity maximum (dQ / dU: charge change divided by battery pack voltage change), etc., wherein the histogram data is particularly histogram data relating to the battery pack temperature distribution at the state of charge, the charging current distribution at the temperature, and / or the discharging current distribution at the temperature.

[0043] Further details can be obtained from the operating characteristics: load patterns over time, such as charging cycles and driving cycles, determined by the use of modes (e.g., fast charging or strong acceleration under high current intensity or with regenerative braking processes), usage time of the vehicle battery pack, accumulated charge and accumulated discharge during operating time, maximum charging current, maximum discharging current, charging frequency, average charging current, average discharging current, power throughput during charging and discharging, (especially average) charging temperature, (especially average) dispersion of the state of charge, etc.

[0044] State of Health (SOH) is a key variable used to describe the remaining capacity or charge of a battery pack. It represents a measure of the aging of a vehicle's battery pack, battery module, or battery cells and can be expressed as Capacity Retention Rate (SOH-C) or as an increase in internal resistance (SOH-R). Capacity Retention Rate (SOH-C) is expressed as the ratio of the measured instantaneous capacity to the initial capacity of a fully charged battery pack. The relative change in internal resistance (SOH-R) increases with increasing battery pack aging.

[0045] exist Figure 2 The flowchart further describes the method for training or updating a data-based aging state model. This method is executed in the data processing unit 21 of the central unit and can be implemented there as software and / or hardware.

[0046] In step S1, a data-based aging state model is provided, which is pre-trained to output the modeled aging state and modeling inaccuracies based on operational feature points. Further, a database is provided that provides time series of operational feature points of the energy storage devices for each evaluation period of a specific vehicle battery pack in the fleet.

[0047] In step S2, operating parameters F are continuously received from vehicles 4 of the convoy 3. These operating parameters F may include changes in the current battery pack current, current battery pack voltage, current battery pack temperature, current state of charge, etc.

[0048] In step S3, the operating characteristics for successive evaluation time periods are determined by the change process of the operating parameter F. These evaluation time periods can range from one day to one month, preferably one week, when determining the aging state of a vehicle battery pack. Typically, to determine the aging state of a specific vehicle battery pack based on a model, a data-based aging state model is used to assign operating characteristic points to the aging state, wherein these operating characteristic points are determined by various operating characteristics for a specific vehicle battery pack within the evaluation time period. This data-based model can also be implemented as a hybrid model, particularly a combination of a physical model and a data-driven model, which preferably has a supervised learning component.

[0049] In step S4, an improvement metric is now determined based on all current operating characteristic points relating to the current operating state of these vehicle battery packs 41, i.e., all vehicle battery packs 41. The improvement metric describes how the state uncertainty at all actual operating characteristic points is improved when state parameters for the corresponding current operating characteristic points are assumed. Therefore, vehicle battery packs for which the improvement metric is derived for the assigned current operating characteristic points can be identified.

[0050] The improvement metrics are determined using a simulated state model, which is created by retraining using a simulated training dataset. The simulated training dataset contains training data points at the current run point that are assigned state labels, where the state labels correspond, for example, to model values ​​of the aging state at the current run point based on an aging state model.

[0051] Based on the simulated state model, state uncertainties are now determined for all operational characteristic points. That is, the aging state and its associated state uncertainty are determined for each evaluation period for all energy storage units considered in the central unit.

[0052] The state uncertainty is derived from the aging state model. However, individual state uncertainties can be reduced using domain knowledge at each operational feature point. The potential for reduction at relevant feature points can be assessed, for example, by using the average uncertainty from the training process to date. Therefore, the domain knowledge used in the training process is incorporated into the simulation model. Furthermore, for each observed operational feature point, the corresponding confidence interval (described by the state uncertainty) can be reduced on a vehicle-by-vehicle basis using domain knowledge. For this, the temporal changes in the aging state (always for the individual identified vehicle battery pack) before and after a given evaluation period can be analyzed, where the evaluation period is assigned to the corresponding operational feature point of the identified vehicle battery pack. Each operational feature point is assigned to an evaluation period, where other operational feature points typically exist before and after the evaluation period for the relevant vehicle battery pack.

[0053] Figure 3 An example is shown of a measure to reduce the overall uncertainty of the state at other running feature points when labels are generated at specific running feature points. Figure 3 Based on the sequential detection of operational feature points over time within different temperature ranges, the modeled m-aging states with corresponding confidence intervals for two different vehicles, F1 and F2, are presented. The aging model used exhibits higher model uncertainty at high temperatures, resulting in a correspondingly larger confidence interval. This could be due, for example, to the fact that high temperatures of 70°C have not been adequately represented in the training database.

[0054] If, under specific conditions, no label exists for vehicle F1 at a particular time point t, where the label is neither generated by fleet vehicles nor determined from diagnostic measurements, then the label must be determined during operation. In the case of vehicle F1, the modeled aging state has low uncertainty at 23°C and high uncertainty at 40°C. Due to the high uncertainty at time point t-1, it is now of particular interest to have a label at time point t in order to minimize the confidence (Konfidenz) in the model at t-1.

[0055] To evaluate the potential effect of label generation for vehicle 1 at time t regarding the added value and relevance to the overall fleet 3, a possible chain reaction (Schwarmintelligenz) is simulated for the fleet, as shown for vehicle F2. Here, the slowly changing state parameters, such as aging states, do not change drastically over short time intervals. The effect of label generation at time t is simulated for vehicle F1. The effect of label generation is directly visible in the case of vehicle F1 (indicated by arrow P1, representing the first chain reaction), because the uncertainty at 40°C at time t-1 can also be minimized in this case, which can lead to an updated model value at t-1, especially if multiple vehicles can be analyzed in this state.

[0056] Given that label generation is now successful at 40°C, the reduced uncertainty of the model at 40°C and other model uncertainties are minimized, for example, at T=70°C. This reduction in state uncertainty is indicated by arrows P2 and P3.

[0057] In this example, given certain characteristics (e.g., temperature 50°C), the label generation for vehicle 1 at time point t not only allows for precise label determination for that running feature point, but also allows for the additional inference of parameter gaps at 40°C (vehicle F1) and 70°C (vehicle F2).

[0058] Through global fleet analysis, the effectiveness of label generation for each vehicle (currently operating feature point) can be evaluated quantitatively and proactively in a holistic manner, and thus, considering the added value for fleet 3, the effectiveness of possible label generation in each available parameter space at the current time point. The advantageous configuration of the invention allows for maximizing a predictable overall measure of uncertainty in the form of an optimization problem.

[0059] In particular, confidence intervals can be limited to aging states that are neither higher than the upper limit of the confidence interval for previously determined aging states nor lower than the lower limit of the aging states determined during subsequent evaluation periods. This restriction is imposed based on the fact that the domain knowledge can only be developed in a strictly monotonically decreasing manner according to the temporal change process of the aging states. Therefore, the confidence limit for aging states, defined as uncertain, is shifted in the direction of the modeled values ​​of the aging states, resulting in a narrowed confidence interval.

[0060] Similarly, other rules for limiting confidence intervals can be defined, derived from domain knowledge. Therefore, as possible additional rules, it can be assumed that the aging state undergoes a smooth and inert or piecewise almost linear change within a limited number of evaluation periods, such as 3 to 10 evaluation periods. This constraint can be applied if, for example, it is known from the fleet load distribution (Flottenlastprofil) that the load distribution for the observed vehicle is constant in terms of energy throughput, temperature conditions, charging behavior, driving behavior, etc., during the observed time period between the previous and next evaluation periods. Therefore, the upper and lower bounds of the limited confidence interval for the aging state can be derived by piecewise interpolation of the upper and lower confidence limits of the previous and next evaluation periods.

[0061] This method allows for defining confidence intervals based on the aging condition of the identified vehicle battery pack.

[0062] The improvement metric can be assumed to be a total metric for all determined state uncertainties. For example, the improvement metric is determined as the sum of all state uncertainties determined using one of these simulated state models.

[0063] In step S5, one or more energy storage devices (current operating feature points) are selected, wherein the maximum improvement metric is obtained for the energy storage device.

[0064] In this way, it is possible to identify which or those running feature points also contribute the best possible effect to reducing state uncertainty for other running feature points. Therefore, the above behavior also utilizes the determination of running feature points for which labels should be generated or for which training datasets should be generated to reduce state uncertainty for other running parameter points, thereby maximizing the label relevance for other fleet participants.

[0065] In step S6, a method for tag generation, i.e., for determining the aging state, is executed for the one or more selected accumulators. Therefore, one or more training datasets can be generated for retraining the aging state model.

[0066] In the next step S7, the data-based aging state model is retrained or updated based on the new training dataset. Alternatively, multiple such training datasets may be determined beforehand, and / or training datasets may be collected over a specific time period.

[0067] The determination of the corresponding state parameters can be performed by statistical methods, especially in combination with domain-based rules for reducing state uncertainty, and by means of clustering methods and overall observation of operating feature points similar to the identified operating feature points.

[0068] The clustering method identifies vehicle battery packs that share similar operational characteristics, particularly those with operational characteristics that have a pre-defined Euclidean distance to the centroid (Zentroid-Punkt) not exceeding a pre-defined threshold. The centroid corresponds to artificially assigned operational characteristics that are not specifically assigned to any particular accumulator. The goal of fleet clustering analysis is to identify vehicle battery packs that lie within a comparable operational characteristic space (parameter space). Possible unsupervised clustering methods include K-means++ or competitive learning.

[0069] In the next step S8, the updated aging state model parameters or the states determined by the model can be transmitted back to the vehicles of the fleet 3, thereby enabling the use of the data-based aging state model in the vehicle 4 to determine the aging state (SOH) of the vehicle battery pack 41. This step is optional. The aging state model can also be run and executed in the central unit 2.

[0070] Alternatively, the determination of the corresponding state parameters can be performed by requesting the device assigned to the at least one selected operating feature point to perform the measurement of the state parameters, wherein the measured state parameters are used after the measurement to create a training dataset.

[0071] Therefore, the training dataset can also be determined by measuring vehicle 4 via diagnostic measurements to determine the corresponding training dataset. For this, active learning can be used, whereby active learning selects the accumulator among all possible accumulators that yields the desired effect of reducing overall model uncertainty as best as possible in the running feature space to perform the diagnostic measurements, especially after step S5 when multiple highly correlated accumulators have been selected. The vehicles can be ultimately selected from the fleet based on their maximized label relevance using pool-based sampling, uncertainty sampling, variance reduction, and expected error reduction methods.

[0072] The above method enables the training and improvement of a state model that improves modeling uncertainty at a large number of running feature points during ongoing operation.

Claims

1. A computationally implemented method for training a state model, wherein the state model is used to determine the state of the electric energy storage device operating as an electric energy storage device, the method comprising the following steps: A data-based state model is provided, which assigns modeled state parameters and state uncertainty measures of the energy storage device to operating feature points; A database is provided, which contains the variation process of real operating characteristic points from a large number of energy storage devices for successive evaluation time periods; The state model is trained or updated based on at least one training dataset, wherein the at least one training dataset is generated using the following steps: Determine an improvement metric for each of the current operating characteristic points of the energy storage device for the current evaluation time period, wherein the improvement metric describes how the state uncertainty at all real operating characteristic points is improved when the state parameters for the corresponding current operating characteristic point are determined. Select at least one running feature point that has the best improvement metric; Determine the corresponding state parameters at the at least one selected operational feature point as labels; The at least one training dataset is generated from the at least one selected running feature point and the corresponding determined state parameters.

2. The method of claim 1, wherein the improvement metric is determined by simulation for each of the current operating feature points by updating the provided state model using a simulation training dataset created for the corresponding current operating feature point to obtain a simulated state model, wherein the simulated state uncertainty is determined by means of the simulated state model for each of the operating feature points of the actual energy storage device during the change process of these operating feature points, wherein the improvement metric is determined based on the simulated state uncertainty.

3. The method of claim 2, wherein the improvement metric is determined as the sum of the uncertainties of the simulated state.

4. The method of claim 3, wherein the reduction of the simulation state uncertainty at the selected operating feature point is performed by at least one rule, wherein the rule depends on the simulation state uncertainty of the corresponding energy storage device at the operating feature point in an evaluation time period different from the evaluation time period of the selected operating feature point and the temporal development of the operating feature point of the corresponding energy storage device.

5. The method of claim 3, wherein the reduction of the simulation state uncertainty at the selected operating feature points is performed by limiting the simulation state uncertainty at each of the selected operating feature points to an upper or lower limit of a confidence interval, wherein the upper or lower limit is determined by the state uncertainty of the evaluation time period prior to the operating feature point, and / or the simulation state uncertainty at each of the selected operating feature points is limited to an upper or lower limit of a confidence interval, wherein the lower or upper limit is determined by the state uncertainty of the evaluation time period following the operating feature point.

6. The method according to any one of claims 3 to 5, wherein the reduction of the simulation state uncertainty at the selected operating feature point is performed by limiting the simulation state uncertainty at each of the selected operating feature points using the following upper and lower limits of corresponding confidence intervals, wherein the upper and lower limits are determined by interpolating the upper limit of the confidence interval for the evaluation time period preceding and following the relevant operating feature point.

7. The method according to any one of claims 1 to 3, wherein the determination of the corresponding state parameters is performed by statistical methods.

8. The method according to any one of claims 1 to 3, wherein the determination of the corresponding state parameter is performed by requesting a device assigned to the at least one selected operating feature point to perform a measurement of the state parameter, wherein the measured state parameter is used after the measurement to create the training dataset.

9. The method according to any one of claims 1 to 3, wherein the state model corresponds to an aging state model for providing an aging state based on operating characteristics or a charging state model for providing a charging state based on the operating characteristics of the energy storage device.

10. The method of claim 9, wherein the operating characteristics are derived from a time series of continuously detected operating parameters, wherein the operating characteristics are determined for successive evaluation time periods, and wherein multiple operating characteristics for the determined evaluation time period of the determined energy storage device define operating characteristic points.

11. The method of claim 10, wherein operating parameters from a large number of technical devices are transmitted to a central unit in communication connection with the large number of technical devices, wherein the operating characteristics are determined by the operating parameters for each evaluation time period, wherein the method is executed in the central unit.

12. The method according to any one of claims 1 to 3, wherein the model parameters of the state model are transmitted to a plurality of devices having the energy storage device, wherein the devices are operated by means of the state model.

13. The method according to any one of claims 1 to 3, wherein the energy storage device (41) is used to operate the equipment.

14. The method according to claim 1, wherein the data-based state model assigns the aging state of the energy storage device and a state uncertainty measure to the operating feature points.

15. The method of claim 3, wherein at least one of the simulation state uncertainties is reduced by means of at least one pre-given domain knowledge-based rule before determining the improvement metric.

16. The method of claim 7, wherein the determination of the corresponding state parameters is performed by statistical methods in combination with domain-based rules for reducing state uncertainty.

17. The method according to claim 13, wherein the energy storage device (41) is used for operating motor vehicles, electric bicycles, aircraft, machine tools, entertainment electronic devices, autonomous robots and / or household appliances.

18. The method of claim 17, wherein the aircraft is a drone.

19. The method of claim 17, wherein the device of the entertainment electronic device is a mobile phone.

20. An apparatus for training a state model, wherein the state model is used to determine the state of an electric energy storage device, wherein the apparatus is configured to: A data-based state model is provided, which assigns modeled state parameters of the energy storage device to operating feature points; A database is provided, which contains the variation process of real operating characteristic points from a large number of energy storage devices for successive evaluation time periods; The state model is trained or updated based on at least one training dataset, wherein the at least one training dataset is generated using the following steps: Determine an improvement metric for each of the current operating characteristic points of the energy storage device for the current evaluation time period, wherein the improvement metric describes how the state uncertainty at all real operating characteristic points is improved when the state parameters for the corresponding current operating characteristic point are determined. Select at least one running feature point that has the best improvement metric; Determine the corresponding state parameters at the at least one selected operational feature point as labels; The at least one training dataset is generated from the at least one selected running feature point and the corresponding determined state parameters.

21. The apparatus of claim 20, wherein the data-based state model assigns the aging state of the energy storage device to operating feature points.

22. A computer program product comprising instructions that, when executed by at least one data processing device, cause the data processing device to perform the steps of the method according to any one of claims 1 to 19.

23. A machine-readable storage medium comprising instructions that, when executed by at least one data processing device, cause the data processing device to perform the steps of the method according to any one of claims 1 to 19.

Images