Method and apparatus for providing training data for training a data-based state model

Abstract

This invention relates to a computer-implemented method for training a state model to determine the state of an electric energy storage device, the method comprising the steps of: - providing a data-based state model that assigns modeled state parameters, particularly aging states, to operating feature points; - training or updating the state model based on at least one training dataset, wherein the training dataset is generated by: o providing or identifying operating feature points with insufficient state uncertainty for their modeled state parameters; o selecting operating feature points from a set of operating feature points, wherein those operating feature points similar to the identified operating feature points with insufficient state uncertainty are selected; o determining the state for the selected operating feature points; o reducing the state uncertainty by means of at least one domain-knowledge-based rule; o generating the training dataset based on the reduced state uncertainty.

Description

Technical Field

[0001] In general, the present invention relates to the characterization of the system state of electric energy storage devices, such as those in electrically driven motor vehicles, especially electric or hybrid vehicles, and also to measures for determining the state of electric energy storage devices, such as vehicle battery packs. Background Technology

[0002] For supplying energy to technical equipment that is not connected to the power grid, electrical storage devices, such as battery packs or fuel cells, are typically used. For example, electrical storage devices, such as vehicle battery packs, are used to supply energy to electrically powered motor vehicles. These storage devices provide electrical energy for the operation of the vehicle systems, particularly the drive systems. The aging condition of an electrical storage device deteriorates over its lifespan, leading to a reduction in its maximum energy storage capacity. The degree of aging of a vehicle battery pack depends on the individual load on the battery pack, that is, on the driver's usage behavior, and also on the type of battery pack.

[0003] While purely physical aging models can determine the current aging state based on historical operational parameter changes, these 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 allows for the economic evaluation of the remaining value of the vehicle battery pack.

[0004] Using conventional physical models, it is usually not possible to determine other states of a vehicle battery pack, such as the state of charge, with high accuracy.

[0005] Not only physical state models, but also data-based state models or hybrid state models require accurate state labels for training data, as these labels form the basis for model parameterization. Continuous label generation after its application enables ongoing improvement of the state model in order 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 to model the state of an electric energy storage device, as described in claim 1, and a corresponding device, as described in the parallel claims, are provided.

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

[0008] According to the first aspect, a computer-implemented method is specified for training a state model to determine the state of an electric energy storage device using operating characteristic points of multiple energy storage devices, the method comprising the following steps: Provides a data-based state model that assigns modeled state parameters, especially aging states, to running feature points; Provide a database containing the changes in real operational characteristic points of the plurality of energy storage units over a continuous evaluation period; The state model is trained or updated based on at least one training dataset. The at least one training dataset is generated using the following steps: By using at least one domain-knowledge-based rule, the state uncertainty of one or more state parameters in the state parameters is reduced, the state uncertainty being derived either from the data-based state model at these real operational feature points or from at least one known label; Provide or identify operational feature points with insufficient state uncertainty for the modeled state parameters; Clustering methods are used to select running feature points from these real running feature points; In particular, by using this data-based state model or through measurement, the state parameters and associated state uncertainties of the selected operational feature points can be determined; The training dataset is generated based on the average value of the state parameters determined for the selected running feature points and the running feature points corresponding to the average value or centroid of the clusters of the selected running feature points determined by the clustering method.

[0009] In particular, by using clustering methods to select running feature points from these real running feature points, it is possible to execute those running feature points that are similar to the identified running feature points with insufficient state uncertainty.

[0010] Furthermore, this state model can correspond to an aging state model used to provide an aging state based on the operating characteristics of the energy storage device, or a charging state model used to provide a charging state based on these operating characteristics.

[0011] 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 with quantified uncertainty calculation. Alternatively, the supervised learning model can be implemented as an integrator method or as a Bayesian neural network.

[0012] Operating characteristics can be derived from the time series of continuously monitored operating parameters, where these operating characteristics are determined for consecutive evaluation periods, and multiple operating characteristics for a specific evaluation period of the determined accumulator define operating characteristic points.

[0013] For electric energy storage devices in higher-level technical equipment such as motor vehicles, it is necessary to determine their state in order to enable the higher-level system to operate in an improved or optimal manner. These states are designated as state parameters. Thus, determining, for example, the aging state of the electric energy storage device and the state of charge of a rechargeable battery pack are important state parameters, essential for the operation of the technical equipment utilized therein, such as electrically powered motor vehicles. Furthermore, these technical devices may include machine tools, household appliances, building energy supply systems, aircraft, especially drones, and / or entertainment electronic devices, especially mobile phones.

[0014] Such state parameters often cannot be modeled with high accuracy using physics-based models. Furthermore, inaccuracies in the computation of state parameters in this context are almost impossible to quantify or can only be painstakingly quantified online. Therefore, data-based state models or hybrid state models are increasingly being considered for modeling these state parameters. Especially when a large amount of training data is available, such state models can achieve very accurate state computation or state prediction.

[0015] For supervised learning techniques, training such data-based state models requires highly accurate training data, which must typically be measured or evaluated beforehand. The process of generating and compiling training data is usually called label generation.

[0016] After training, the data-based state model can be implemented in advance in the devices that utilize it by transmitting the corresponding model parameters. Alternatively, devices with regular connections to the central unit (cloud) may periodically receive model results, such as state parameters or updated model parameters of the data-based state model, from the central unit. Especially when evaluating operating parameters measured in multiple 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 similar devices to improve state predictions there. Preferably, state calculation or state prediction is performed in the central unit (cloud), and the results are provided as state to the IoT devices.

[0017] To provide training data for training a data-driven state model, state parameters can be derived for one of these devices in a complex manner, often with the aid of time-consuming diagnostic measurements, based on the obtained operating parameters. These state parameters are specifically identified as labels in the central unit so that a training dataset can be provided there based on the operating parameters or the operating characteristics derived from them and the assigned state parameters (labels). Thus, to determine an aging state model in the central unit, for example, the operating parameters of a vehicle battery pack used to operate an electrically driven motor vehicle can be continuously transmitted to the central unit, and an aging state description can be derived there by evaluating the determined operating changes, such as a complete charging cycle under defined load and environmental conditions, thereby providing a training dataset there.

[0018] In the initial stages of operating a device with a central unit and multiple similar devices communicating with that central unit, based on a data-based state model used to estimate state parameters, the data-based state model is typically not yet trained with sufficient accuracy across all regions of the input data space. 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 from these devices. Not all regions of the input parameter space defined by operating parameter points or operating feature points can have their corresponding state parameters easily determined by real measurements, making the data-based state model highly inaccurate and potentially unusable in these regions.

[0019] The aforementioned method specifies that an artificially generated training dataset is used to improve the data-based state model. The artificially generated training dataset is determined based on the current training state of the state model, the already determined operational feature points of the real accumulator, or through at least one known label and at least one rule derived from domain knowledge to limit state uncertainty. Therefore, the artificially generated training dataset can be obtained without additional measurement and / or determination of the state parameters. This determined training dataset can then be used to update the data-based state model.

[0020] In the above method, a data-based state model is provided in the central unit for multiple similar technical devices. This data-based state model can also be implemented as a hybrid state model. Here, the data-based state model is updated and improved by evaluating the available training dataset and the data-based state model during actual operation of multiple technical devices, creating additional training datasets suitable for retraining or updating the data-based state model. Thus, the initially provided data-based state model can be retrained in regions of high state uncertainty in the input parameter space.

[0021] In particular, for state parameters, it depends on the systematic existence of correlations regarding the time-varying processes of these state parameters, which are known and can be governed or used as domain knowledge. This enables the reduction of large confidence intervals or high state uncertainties in state parameter predictions (modeled state parameters) derived from modeling using state models, by setting, for example, upper and lower bounds on the confidence intervals to exclude physically impossible ranges of values ​​for the relevant state parameters.

[0022] Furthermore, state uncertainty at a selected operating feature point can be reduced by at least one rule, wherein the rule depends on the operating feature point and the evolution of the state uncertainty of the operating feature point over time in an evaluation period different from the evaluation period of the selected operating feature point.

[0023] In particular, the state uncertainty at selected operating characteristic points can be reduced by limiting the state uncertainty at each of these selected operating characteristic points to an upper or lower bound of a confidence interval, which is determined by the state uncertainty of the relevant operating characteristic point in the previous evaluation period; and / or by limiting the state uncertainty at each of these selected operating characteristic points to a lower or upper bound of a confidence interval, which is determined by the state uncertainty of the relevant operating characteristic point in the next evaluation period. This can be attributed to domain knowledge, since slowly time-varying state parameters, such as the aging state of an energy storage device, change only slowly.

[0024] Alternatively or additionally, the state uncertainty at the selected operating feature points can be reduced by limiting the state uncertainty at each of these selected operating feature points using the upper and lower bounds of the corresponding confidence interval, which are determined by interpolation of the upper bounds of the confidence intervals for the previous and next evaluation periods at the relevant operating feature point.

[0025] From multiple real-world running feature points, clustering methods can be used to select running feature points located in regions surrounding running feature points with high state uncertainty or large confidence intervals. Here, by combining selection via clustering with reducing state uncertainty through the application of domain knowledge, a new training dataset can be generated. This new training dataset corresponds to the centroids or centers of the clusters of the selected running feature points determined by the clustering method, along with the average values ​​of the state parameters determined for this purpose.

[0026] The effectiveness of the training dataset generated in this way can be verified by statistical significance using the reduced state uncertainty, thus confirming the accuracy requirements and the applicability of the labels. In one implementation, training or updating of the state model can only be performed based on the at least one training dataset if the total state uncertainty does not exceed a pre-defined uncertainty threshold, wherein the total state uncertainty is determined according to the law of large numbers or by means of an error propagation method based on the state uncertainty calculated from the selected running feature points.

[0027] In other words, if the required statistical accuracy for the running feature points is lower than that required for the generated training dataset, correspondingly artificially generated and statistically evaluated state parameters can be used together with the selected running feature points as a new training dataset. In this way, additional training datasets for training data-based state models can be generated for similar devices connected to the cloud, especially for regions where the running feature points currently exhibit high state uncertainty in the data-based state model. This approach of providing additional training data based on applied domain knowledge enables rapid improvement of the data-based state model to predict the state parameters of multiple similar devices.

[0028] It can be stipulated that: the operating parameters of multiple technical systems are transmitted to a central unit, wherein a database of operating characteristic points of real energy storage devices of the multiple technical systems is provided, wherein the method is implemented in the central unit, and operating characteristic points are selected from the database.

[0029] Operating characteristic points with insufficient state uncertainty can be identified during the assessment of the current or historical operating characteristic points of the energy storage device.

[0030] According to another aspect, a device is specified for training a state model to determine the state of an electric energy storage device by means of operating characteristic points of multiple energy storage devices, wherein the device is constructed for: Provides a data-based state model that assigns modeled state parameters, especially aging states, to running feature points; Provide a database containing the changes in real operational characteristic points of the plurality of energy storage units over a continuous evaluation period; The state model is trained or updated based on at least one training dataset. The at least one training dataset is generated using the following steps: By using at least one domain-knowledge-based rule, the state uncertainty of one or more state parameters in the state parameters is reduced, the state uncertainty being derived from the data-based state model at these real operational feature points or from at least one known label; Provide or identify operational feature points with insufficient state uncertainty for the modeled state parameters; Based on the determined operational feature points, a clustering method is used to select operational feature points from the actual operational feature points; In particular, by using this data-based state model or through measurement, the state parameters and associated state uncertainties of the selected operational feature points can be determined; The training dataset is generated based on the average value of the state parameters determined for the selected running feature points and the running feature points corresponding to the average value or centroid of the selected running feature points. Attached Figure Description

[0031] The embodiments are then described in more detail with reference to the accompanying drawings. Wherein: Figure 1 A schematic diagram of a system for a fleet of vehicles with multiple vehicles and a central unit for providing a data-based state model is shown. Figure 2 A flowchart is shown to illustrate a method for training or updating a data-based state model, particularly a data-based state model in the form of an aging state model of a vehicle battery pack. Figure 3 The modeled aging state time series with large confidence intervals is shown. Figure 4 This illustrates how to reduce the confidence interval by moving the confidence limit; Figure 5 This illustrates how to reduce the confidence interval by moving the confidence limit using linear interpolation; Figure 6 This illustrates the clustering of identified vehicle battery packs with similar operational characteristics; and Figure 7 A diagram illustrating how the described method can improve a data-based aging state model. Detailed Implementation

[0032] The method according to the invention is described below with reference to vehicle battery packs in multiple motor vehicles of the same type. In the motor vehicle, a data-based aging state model for the corresponding vehicle battery pack can be implemented in a control unit. The aging state model exemplarily represents a state model. In the central unit, the aging state model can be continuously updated or retrained based on the operating parameters of the vehicle battery packs in the fleet.

[0033] The examples above represent a large number of stationary or mobile devices with power supplies independent of the power grid, such as vehicles, facilities, IoT devices, and the like, which are connected to a central unit (cloud) via corresponding communication connections (e.g., LAN, Internet). Here, state parameters are those that cannot be determined with high accuracy and model-based precision in similar devices in a simple manner, and can only be determined through costly calculations, internal diagnostic measurements, especially destructive measurements, or after a predefined operating cycle of the device.

[0034] 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 the central unit 2, calculations and predictions should be made based on the fleet data to assess the aging process of the battery packs of the corresponding motor vehicles 4 in the fleet 3.

[0035] One of these four motor vehicles is Figure 1 The details are shown in more detail below. These motor vehicles 4 each have: a vehicle battery pack 41 as a rechargeable electrical energy storage device; an electric drive motor 42; and a control unit 43. The control unit 43 is connected to a communication module 44 adapted to transmit data between the respective motor vehicle 4 and the central unit (cloud). The control unit 43 is also connected to a sensor unit 45 having one or more sensors for continuous monitoring of operating parameters.

[0036] The central unit 2 has: a data processing unit 21 in which the methods described below can be implemented; and a database 22 for storing the operating parameters and aging status of the vehicle battery packs that have been determined in the multiple vehicles 4 of the fleet 3.

[0037] Vehicle 4 sends operating parameters F to 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, operating parameters F can indicate the current battery pack current, current battery pack voltage, current battery pack temperature, and current state of charge (SOC). Operating parameters F are detected in fast time frames between 2 Hz and 100 Hz, during which changes in these operating parameters are periodically transmitted to central unit 2 in uncompressed and / or compressed form.

[0038] In central unit 2 or in other embodiments, operating characteristics relating to the assessment period can also be generated in the corresponding motor vehicle 4 based on operating parameters F. For determining aging conditions, this assessment period can range from several hours (e.g., 6 hours) to several weeks (e.g., one month). A commonly used value for this assessment period is one week.

[0039] These operating characteristics may include, for example, characteristics and / or cumulative characteristics related to the evaluation period and / or statistical parameters determined over the entire service life to date. In particular, these operating characteristics may include, for example, histogram data on the process of changing state of charge, temperature, battery pack voltage, battery pack current, especially histogram data on the distribution of battery pack temperature with respect to state of charge, the distribution of charging current with respect to temperature, and / or the distribution of discharging current with respect to temperature, the cumulative total charge (Ah), the average capacity increase during the charging process (especially for charging processes in which the charge increase exceeds a threshold share (e.g., 20%) of the total battery pack capacity), the maximum value of the differential capacity (dQ / dU: change in charge divided by change in battery pack voltage), and others.

[0040] Other insights can be gleaned from these operating characteristics: load patterns over time, such as charging and driving cycles, which are determined by the use of modes (such as rapid charging under high current intensity or strong acceleration or braking processes with regenerative braking); the usage time of the vehicle battery pack; the cumulative charge and discharge during the operating time; the maximum charging current; the maximum discharging current; the charging frequency; the average charging current; the average discharging current; the power throughput during charging and discharging; (especially the average) charging temperature; the average state of charge; the (especially the average) distribution of the state of charge; and so on.

[0041] State of Health (SOH) is a key parameter used to describe the remaining capacity or charge of a battery pack. It measures the aging of a vehicle's battery pack, battery module, or battery cells and can be specified as Capacity Retention Rate (SOH-C) or the increase in internal resistance (SOH-R) in the case of a battery pack. Capacity Retention Rate (SOH-C) is specified as the ratio of the measured current capacity to the initial capacity of a fully charged battery pack. The relative change in internal resistance (SOH-R) increases with the aging of the battery pack.

[0042] exist Figure 2 The flowchart describes in more detail the method for training or updating a data-based aging state model. This method is implemented in the data processing unit 21 of the central unit and can be implemented there as software and / or hardware.

[0043] In step S1, a data-based aging state model is provided, which is pre-trained to output modeled aging states and state inaccuracies based on operational feature points. A database is also provided that provides time series of operational feature points for each evaluation period for the electric energy storage devices of the fleet's vehicles.

[0044] In step S2, operating parameters F are continuously received from vehicles 4 of fleet 3. Operating parameters F may include, not only at the pack level and module level but also at the battery level, the current battery pack cells, the current battery pack voltage, the current battery pack temperature, the current state of charge change process, and so on.

[0045] In step S3, the operating characteristics for a continuous evaluation period are determined based on the change process of the operating parameter F. When determining the aging state of the vehicle battery pack, these evaluation periods can range from one day to one month, preferably one week. Typically, to determine the determined aging state of the vehicle battery pack based on a model, an aging state is assigned as a state parameter to the operating feature points using a data-based aging state model. These operating feature points are determined by various operating characteristics of the determined vehicle battery pack within the evaluation 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, preferably with a supervised learning component.

[0046] In step S4, the following running feature points are determined where the state uncertainty is insufficient, specifically exceeding a pre-defined uncertainty threshold. For this purpose, a data-based aging state model can be provided that, in addition to model predictions, also describes the prediction uncertainty, for example, in the form of standard deviation or confidence intervals. For example, hybrid models or purely data-based models can be considered, including supervised learning methods, preferably Gaussian process models, or alternatively, integrator methods or Bayesian neural networks.

[0047] The operational feature points with insufficient calculated state uncertainty can be determined, for example, by using a data-based aging state model to determine and evaluate the modeled aging state of all vehicle battery packs at different evaluation periods. In other words, for each actual, i.e., for an actual vehicle battery pack, the determined operational feature point is used to determine the aging state and its modeled state uncertainty. If the state uncertainty of one of the modeled aging states is higher than a pre-given uncertainty threshold, the relevant actual operational feature point is determined as an operational feature point with insufficient state uncertainty. Therefore, for the current embodiment, the description of the aging state of the determined (as identified above) operational feature points of the determined vehicle battery pack is considered too inaccurate.

[0048] For example, in Figure 3 The diagram shows a time series of aging states Z with their respective confidence intervals KI for consecutive evaluation periods. It can be seen that there is high uncertainty at time point t, while there is low uncertainty at the previous evaluation period tn and the next evaluation period t+n.

[0049] In step S5, for each of the real operating feature points, or at least those real operating feature points whose state uncertainty is higher than a pre-given uncertainty threshold, a vehicle-specific reduction of the corresponding confidence interval KI is performed using domain knowledge. To this end, the time-varying process of the aging state (always for an individual vehicle battery pack) around the evaluation period of the considered real operating feature point can be analyzed. Each operating feature point is assigned an evaluation period, where other operating feature points of the relevant vehicle battery pack typically exist before and after that evaluation period.

[0050] In particular, confidence intervals can be limited to aging states that are not higher than the upper limit of the confidence interval for previously determined or known aging states (in the previous evaluation period) and not lower than the lower limit of the aging state determined in the subsequent evaluation period. This is in Figure 4As exemplarily shown in the figure, the restricted confidence interval is characterized by KI'. This restriction can be based on domain knowledge: the change of the aging state over time in a sufficiently long time window can only decrease monotonically, especially when considering the effective stress factors that promote aging. Therefore, the confidence limit G, which will be restricted to an uncertain aging state, is shifted toward the statistical expectation of the modeled value of the aging state, resulting in a narrowed confidence interval.

[0051] Similarly, other rules for limiting confidence intervals can be derived from domain knowledge. Thus, as a possible further rule, it can be assumed that the aging state undergoes a continuous and slow, or piecewise almost linear, change over a finite period of a few assessment periods, for example, 3 to 10 assessment periods. This limitation can be made if, for example, the load curve and the resulting stress factors related to the aging of the vehicle under consideration in terms of energy throughput, temperature conditions, charging behavior, driving behavior, etc., are known from the fleet load curve, to be constant over the period between the previous and next assessment periods. Thus, the upper and lower bounds of the limited confidence interval KI' for the relevant aging state can be obtained by piecewise interpolation of the upper and lower confidence bounds for the previous and next assessment periods. This is in... Figure 5 As exemplarily shown in the figure.

[0052] In this way, the confidence interval for the aging state at the actual operating characteristic points of the vehicle battery pack can be limited. The high uncertainty at the actual operating characteristic points can be reduced by domain knowledge of the time inertia or the time-varying process of the state parameters. Therefore, the knowledge of the low state uncertainty at the evaluation time point can be attributed either to accurate modeling of the aging state using an aging state model, to the results of measurements, or a combination of both.

[0053] In step S6, based on the actual operating feature points, clustering methods are used to identify those operating feature points similar to the determined operating feature points. Here, all actual operating feature points identified from the database for past evaluation periods of all vehicle battery packs 41 are considered. The clustering method identifies actual operating feature points similar to the determined operating feature points from these actual operating feature points, particularly those with a pre-given Euclidean distance from the centroid, which is no greater than a pre-given threshold. The centroid corresponds to artificial operating feature points not assigned to the determined energy storage units. Similarity is defined by operating feature points. If, for example, the Euclidean distance between operating feature points in multidimensional space is sufficiently small, i.e., below a threshold associated with accuracy requirements, then the two operating feature points are sufficiently similar. In other words, the Euclidean distance determines the degree of similarity between two operating feature points. If this degree of similarity is higher than the pre-given threshold, then the two operating feature points are similar.

[0054] K-means++ and competitive learning can be used as possible unsupervised clustering methods.

[0055] In step S7, the corresponding aging state can be estimated for each of the selected operating feature points according to a data-based state model. Furthermore, a confidence range is derived from the model estimate as the state uncertainty.

[0056] Now, within a range of similar operating characteristic points, there exists a selection of real operating characteristic points for the identified vehicle battery pack, each assigned an aging state with a confidence interval. Reduced confidence intervals determined by applying domain knowledge can be assigned to the selected real operating characteristic points. For this, at least one point in the cluster must be selected. This is, for example, in… Figure 6 As shown in the figure, each region in region B corresponds to a running feature point MP with a corresponding restricted confidence interval.

[0057] Now, the aging state assigned to the selected operating feature points can be averaged for all vehicles in step S8, and the confidence intervals can be correspondingly merged into the total confidence interval or merged into a common standard deviation. In, for example, using the square root law when the confidence interval is the same size. To be combined in the total confidence interval or to be combined in a common standard deviation, where The same standard deviation is given, and n corresponds to the number of aging states considered.

[0058] In cases where there are varying and constrained confidence intervals, error propagation can be used to calculate the total confidence interval in order to determine the variance or standard deviation of the averaged aging states. For example, model-based error propagation can be used to determine the variance or standard deviation of the averaged aging states by utilizing model sensitivity to calculate the combined label uncertainty. This can be achieved, for example, by evaluating and combining the partial derivatives of the data-based state model, such as using Taylor series expansion.

[0059] In the subsequent step S9, it is checked whether the total confidence interval thus determined is lower than a pre-given maximum confidence interval, which indicates the maximum allowable size of the confidence interval. If so (either option: yes), the method continues to step S10. Otherwise (either option: no), it jumps back to step S1.

[0060] In step S10, a new training dataset is generated based on the average value or centroid (the running feature point representing the centroid of the relevant cluster) of the identified vehicle battery pack 41, or another running feature point assigned to the cluster and the average value of the aging state to which it belongs.

[0061] In the subsequent step S11, the data-based aging state model is retrained or updated based on the new training dataset. Before updating, multiple such training datasets may be identified first, and / or training datasets collected over a specific time period may be collected.

[0062] Figure 7 The predictions of the aging state of a specific running feature point MP are shown before (1) updating the data-based aging state model and after (2) updating the data-based aging state model.

[0063] In subsequent step S12, the updated aging state model parameters, or the state determined by the model, can be transmitted back to the vehicles in fleet 3, so that the data-based aging state model can be used in vehicle 4 to determine the aging state (SOH) of vehicle battery pack 41. This step is optional. The aging state model can also be run in central unit 2 and implemented there.

Claims

1. A computer-implemented method for training a state model to determine the state of an electric energy storage device using operating characteristic points of multiple energy storage devices, the method comprising the following steps: Provide a data-based state model that assigns modeled state parameters to running feature points; A database is provided, which contains the changes in real operating characteristic points of the plurality of energy storage devices over a continuous evaluation period; The state model is trained or updated based on at least one training dataset. The at least one training dataset mentioned above is generated using the following steps: By using at least one domain-knowledge-based rule, the state uncertainty of one or more state parameters in the state parameters is reduced, and the state uncertainty is derived at the actual operational feature point by the data-based state model. Provide or determine the running feature points with insufficient state uncertainty for the modeled state parameters, or provide or determine the running feature points by at least one known label; Based on the determined operational feature points, a clustering method is used to select operational feature points from the actual operational feature points; Determine the state parameters and associated state uncertainties of the selected operational feature points; The training dataset is generated based on the average value of the state parameters determined for the selected running feature points and the running feature points corresponding to the average value or centroid of the selected running feature points.

2. The method according to claim 1, wherein the modeled state parameter is an aging state.

3. The method of claim 1, wherein the state parameters and associated state uncertainties of the selected operational feature points are determined by means of the data-based state model or by measurement.

4. The method of claim 1, wherein, in order to select running feature points from the real running feature points by means of a clustering method, those running feature points similar to the determined running feature points with insufficient state uncertainty are executed.

5. The method according to any one of claims 1 to 4, wherein the state model corresponds to an aging state model for providing an aging state based on the operating characteristics of the energy storage device or a charging state model for providing a charging state based on the operating characteristics.

6. The method according to any one of claims 1 to 4, wherein the data-based state model comprises a data-based machine learning model, the data-based machine learning model being constructed to describe state uncertainty for modeled state parameters, wherein the state model comprises supervised learning methods, such as Gaussian process models or Bayesian neural networks.

7. The method according to any one of claims 1 to 4, wherein operating characteristics are derived from a time series of continuously detected operating parameters, wherein the operating characteristics are determined for consecutive evaluation periods, wherein multiple operating characteristics for a particular evaluation period within the consecutive evaluation periods of the determined energy storage device define operating characteristic points.

8. The method of claim 7, wherein operating parameters of a plurality of technical systems are transmitted to a central unit, wherein a database of operating characteristic points of real energy storage devices of the plurality of technical systems is provided, wherein the method is implemented in the central unit, wherein operating characteristic points are selected from the database.

9. The method according to any one of claims 1 to 4, wherein during the evaluation of the determined current operating characteristic point of the energy storage, an operating characteristic point with insufficient state uncertainty is provided or determined.

10. The method according to any one of claims 1 to 4, wherein the state uncertainty at the selected operating characteristic point is reduced by at least one rule, the rule depending on the operating characteristic point of the corresponding energy storage device and the development of the state uncertainty of the corresponding energy storage device over time in an evaluation period different from the evaluation period of the selected operating characteristic point.

11. The method according to any one of claims 1 to 4, wherein the state uncertainty at the actual operating feature point is reduced by: limiting the state uncertainty at each of the selected operating feature points to an upper or lower limit of a confidence interval, the upper or lower limit being determined by the state uncertainty of the relevant operating feature point in the previous evaluation period; and / or limiting the state uncertainty at each of the selected operating feature points to a lower or upper limit of a confidence interval, the lower or upper limit being determined by the state uncertainty of the relevant operating feature point in the next evaluation period.

12. The method according to any one of claims 1 to 4, wherein the state uncertainty at the selected operating feature points is reduced by limiting the state uncertainty at each of the selected operating feature points using upper and lower limits of a corresponding confidence interval, which are determined by interpolation of the upper limits of the confidence intervals for the previous and next evaluation periods of the relevant operating feature point.

13. The method according to any one of claims 1 to 4, wherein training or updating of the state model is performed only on the at least one training dataset if the total state uncertainty does not exceed a pre-given uncertainty threshold, wherein the total state uncertainty is determined according to the state uncertainty of the selected running feature points by the law of large numbers or by means of an error propagation method.

14. The method according to any one of claims 1 to 4, wherein the model parameters of the state model are transmitted to a plurality of devices having an energy storage device.

15. The method according to any one of claims 1 to 4, wherein the state model is implemented in a central unit that maintains communicative connections with a plurality of devices, wherein the state parameters of each of the plurality of devices can be queried by transmitting operating parameters to the central unit, and the central unit transmits the state parameters determined by means of the state model to the requesting device.

16. An apparatus for training a state model to determine the state of an electric energy storage device using operating characteristic points of multiple energy storage devices, wherein the apparatus is configured to: Provide a data-based state model that assigns modeled state parameters to running feature points; A database is provided, which contains the changes in real operating characteristic points of the plurality of energy storage devices over a continuous evaluation period; The state model is trained or updated based on at least one training dataset. The at least one training dataset mentioned above is generated using the following steps: By using at least one domain-knowledge-based rule, the state uncertainty of one or more state parameters in the state parameters is reduced, the state uncertainty being derived at the actual operational feature point by the data-based state model or by at least one known label; Provide or identify operational feature points with insufficient state uncertainty for the modeled state parameters; Based on the determined operational feature points, a clustering method is used to select operational feature points from the actual operational feature points; Determine the state parameters and associated state uncertainties of the selected operational feature points; The training dataset is generated based on the average value of the state parameters determined for the selected running feature points and the running feature points corresponding to the average value or centroid of the selected running feature points.

17. The device of claim 16, wherein the modeled state parameter is an aging state.

18. The device of claim 16, wherein the state parameters and associated state uncertainties of the selected operating feature points are determined by means of the data-based state model or by measurement.

19. 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 15.

20. 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 15.

Images