Method and device for providing training data for training a data-based state model for determining a state of an electrical energy storage device

Abstract

Computer-implemented method for training a state model for determining a state of an electrical energy storage device using operating characteristic points of a plurality of energy storage devices, comprising the following steps: - Providing a data-based state model that assigns a modeled state variable, in particular an ageing state, to an operating characteristic point, - Providing a database with histories of real operating characteristic points from the plurality of energy storage devices for successive evaluation periods; - Training or updating the state model depending on at least one training data set, wherein the at least one training data set is generated with the following steps: ◯ Reducing the state uncertainties of one or more of the state variables resulting from the data-based state model at the real operating characteristic points using at least one rule based on domain knowledge; ◯ Providing or determining an operating characteristic point with an insufficient state uncertainty of the associated modeled state variable or by at least one known label, ◯ Selecting operating characteristic points from the real operating characteristic points using a clustering procedure depending on the determined operating characteristic point, ◯ Determination of the state variables for the selected operating characteristic points and associated state uncertainties, in particular using the data-based state model or by measurement; ◯ Generating the training data set from the mean value of the determined state variables for the selected operating characteristic points and an operating characteristic point that corresponds to a mean value or a centroid of the selected operating characteristic points.

Description

Technical area

[0001] The invention generally relates to the characterization of system states of electrical energy storage devices, such as in electrically driven motor vehicles, in particular electric vehicles or hybrid vehicles, and further to measures for determining a state of an electrical energy storage device, such as a vehicle battery. Technical background

[0002] Electrical energy storage devices such as batteries or fuel cells are generally used to supply technical devices with energy that is not tied to the grid. For example, electrically powered vehicles are powered by an electrical energy storage device, such as a vehicle battery. This supplies electrical energy to operate vehicle systems, particularly the drive system. The aging condition of the electrical energy storage device deteriorates noticeably over the course of its service life, resulting in a decreasing maximum storage capacity. The degree of aging of a vehicle battery depends on the individual load on the vehicle battery, i.e., the driver's usage behavior and the vehicle battery type.

[0003] Although a purely physical aging model can be used to determine the current aging state based on historical operating parameters, this model is often inaccurate. This inaccuracy of the conventional aging model makes it difficult to predict the aging state progression. However, predicting the aging state progression of a vehicle battery is an important technical parameter, as it enables an economic assessment of the residual value of the vehicle battery.

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

[0005] Both physical and data-based or hybrid state models require precisely recorded state labels for training data, as they form the basis for model parameterizations. Continuous label generation after the model has been deployed enables continuous improvement of the state model to determine and predict the state variable to be characterized.

[0006] Document EP 3 224 632 B1 discloses a wireless network-based battery management system (BMS) for hybrid or electric vehicles, comprising an off-board subsystem and an on-board subsystem, wherein the on-board subsystem is an on-board battery system, wherein the on-board subsystem comprises on-board data storage for storing historical data of the on-board subsystem and data from other battery systems, and an off-board data processing device configured to process the stored data and to create and validate accurate and complex off-board battery models. Disclosure of the invention

[0007] According to the invention, a method for providing training data for training a state model for modeling a state of an electrical energy storage device according to claim 1 and a corresponding device according to the independent claim are provided.

[0008] Further embodiments are specified in the dependent claims.

[0009] According to a first aspect, a computer-implemented method for training a state model for determining a state of an electrical energy storage device using operating feature points of a plurality of energy storage devices, comprising the following steps: - Providing a data-based state model that assigns a modeled state variable, in particular an ageing state, to an operating characteristic point, - Providing a database with histories of real operating characteristic points from the plurality of energy storage devices for successive evaluation periods; - Training or updating the state model depending on at least one training data set, where the at least one training data set is generated with the following steps: ◯ Reducing the state uncertainties of one or more of the state variables resulting either from the data-based state model at the real operating characteristic points or from at least one known label, using at least one rule based on domain knowledge; ◯ Providing or determining an operating characteristic point with an insufficient state uncertainty of the associated modeled state variable, ◯ Selecting operating characteristic points from the real operating characteristic points using a clustering method, ◯ Determination of the state variables for the selected operating characteristic points and associated state uncertainties, in particular using the data-based state model or by measurement; ◯ Generating the training data set from the mean value of the determined state variables for the selected operating characteristic points and an operating characteristic point that corresponds to a mean value or a centroid of a cluster of the selected operating characteristic points determined by the clustering method.

[0010] In particular, the selection of operating characteristic points from the real operating characteristic points can be carried out using a clustering procedure to select those operating characteristic points that are similar to the determined operating characteristic point with the insufficient state uncertainty.

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

[0012] It can be provided that the data-based state model comprises a data-based machine learning model configured to specify a state uncertainty for a modeled state variable, wherein the state model, in particular, comprises a Gaussian process model as a supervised learning model with quantified uncertainty calculation. Alternatively, the supervised learning model can be implemented, in particular, as an ensemble method or as a Bayesian neural network.

[0013] The operating characteristics can be derived from time series of continuously recorded operating variables, whereby the operating characteristics are determined for successive evaluation periods, whereby several operating characteristics for a specific evaluation period of a specific energy storage device define an operating characteristic point.

[0014] For electrical energy storage devices in higher-level technical devices such as motor vehicles, the determination of states is necessary in order to be able to operate the higher-level system in an improved or optimal manner. These states are specified as state variables. For example, the determination of the aging state of the electrical energy storage device and the charge state of a rechargeable battery are important state variables that are essential for the operation of the technical devices powered by them, such as an electrically powered motor vehicle. Furthermore, the technical devices can include a machine tool, a household appliance, a building energy supply, an aircraft, in particular a drone, and / or an entertainment electronics device, in particular a mobile phone.

[0015] Such state variables generally cannot be modeled with high accuracy using physically based models. Furthermore, the inaccuracy of the state variables' calculations is difficult or impossible to quantify online. Therefore, the use of data-based or hybrid state models is increasingly being considered for modeling these state variables. Especially when a large amount of training data is available, such state models can enable very accurate state calculations or predictions.

[0016] Training such data-based state models for supervised learning techniques requires highly accurate training data, which typically must be measured or evaluated in advance. The process of generating and compiling training data is commonly referred to as label generation.

[0017] After training, data-based state models can be implemented in devices operated with them in advance by transmitting the corresponding model parameters. It is also possible for devices that have a regular connection to a central unit (cloud) to regularly receive the model results, such as the state variables or updated model parameters of the data-based state model, from the central unit. In particular, if operating variables measured in a large number of devices equipped with the same electrical energy storage devices are evaluated 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 the similar devices in order to improve their state prediction. The state calculation orState prediction in the central unit (cloud) and the results are provided as states to the IOT devices.

[0018] In order to provide training data for training the data-based state model, a state variable can be derived from the obtained operating variables in a complex manner, usually by means of time-consuming diagnostic measurements, for one of the devices, which is determined as a label, in particular in the central unit, in order to provide a training data set there based on operating variables or operating characteristics derived therefrom and the associated state variable (label).For example, in order to determine an ageing state model in a central unit, operating variables for the operation of a vehicle battery of an electrically powered motor vehicle can be continuously transmitted to the central unit and an ageing state indication can be derived there by evaluating certain operating processes, such as a complete charging cycle under defined load and ambient conditions, so that a training data set is made available there.

[0019] In the initial phase of operation of a technical system with a central unit and a large number of similar devices that are in communication with the central unit and are based on a data-based state model for estimating a state variable, the data-based state model is often not yet trained with sufficient accuracy in all areas of an input data space at the start of operation. In order to continuously improve the performance of the data-based state model during its use, the data-based state model in the central unit can be regularly updated based on training data sets that can be determined using the operating variables recorded by the devices. Not all areas of the input variable space, which is defined by the operating variable points orIf operating characteristic points are defined, corresponding access variables can be easily determined by real measurements, so that the data-based condition model remains very inaccurate in these areas and may not be usable.

[0020] The above method now proposes improving the data-based state model with artificially generated training data sets. The artificially generated training data sets are determined from the current training state of the state model, previously determined operating characteristic points of real energy storage systems, or by at least one known label and at least one rule resulting from domain knowledge for limiting state uncertainty. Thus, the artificially generated training data sets can be obtained without additional measurements and / or determination of state variables. The training data sets thus obtained can then be used to update the data-based state model.

[0021] In the above method, a data-based state model, which can also be implemented as a hybrid state model, is made available for a large number of similar technical devices in a central unit. The data-based state model is updated and improved by evaluating available training data sets and evaluating the data-based state model in real operation of the large number of technical devices, creating additional training data sets suitable for retraining or updating the data-based state model. This allows an initially provided data-based state model to be retrained in areas of the input variable space where there is a high level of state uncertainty.

[0022] In particular, system-related dependencies exist for state variables regarding their temporal progression, which are known and available or usable as domain knowledge. This makes it possible to reduce large confidence intervals or high state uncertainties of state variable predictions (modeled state variables) resulting from modeling with the state model by excluding physically impossible value ranges of the respective state variable by setting, for example, the upper and lower limits of the confidence interval.

[0023] Furthermore, the state uncertainties at the selected operating characteristic points can be reduced by at least one rule, wherein the rule depends on a temporal development of the operating characteristic points and state uncertainties of operating characteristic points at evaluation periods other than the evaluation periods of the selected operating characteristic points.

[0024] In particular, the state uncertainties at the selected operating characteristic points can be reduced by limiting the state uncertainty at each of the selected operating characteristic points to an upper or lower limit of a confidence interval determined by the state uncertainty at an evaluation period preceding the respective operating characteristic point, and / or by limiting the state uncertainty at each of the selected operating characteristic points to a lower or upper limit of a confidence interval determined by the state uncertainty at an evaluation period following the respective operating characteristic point. This can be attributed to domain knowledge, since a slowly time-varying state variable such as the aging state of an energy storage device changes only slowly.

[0025] Alternatively or additionally, the state uncertainties at the selected operating characteristic points can be reduced by limiting the state uncertainty at each of the selected operating characteristic points with an upper and a lower limit of a corresponding confidence interval, which is determined by interpolating upper limits of confidence intervals of evaluation periods preceding and following the respective operating characteristic point.

[0026] From the multitude of real operating characteristic points, operating characteristic points can be selected using a clustering method that lie within a range around an operating characteristic point that exhibits high state uncertainty or a large confidence interval. By combining the selection by the clustering method and the reduction of state uncertainty through the application of domain knowledge, a new training data set can be generated. This new training data set corresponds to a center point or centroid of the cluster of selected operating characteristic points determined by the clustering method and a mean value of the state variables determined for this cluster.

[0027] The validity of the training data set thus generated can be validated with statistical significance by the reduced state uncertainties, thus proving the accuracy requirement and thus the suitability as labels. According to one embodiment, the training or updating of the state model based on the at least one training data set can only be performed if a total state uncertainty does not exceed a predefined uncertainty threshold, wherein the total state uncertainty is determined from the calculated state uncertainties of the selected operating characteristic points according to the law of large numbers or using an error propagation method.

[0028] In other words, if a statistical accuracy requirement for the operating characteristic point of the generated training dataset is not met, the correspondingly artificially generated and statistically evaluated state variable with the selected operating characteristic points can be used as a new training dataset. In this way, additional training datasets can be generated for similar devices connected to a cloud to train a data-based state model, particularly for areas of operating characteristic points where the data-based state model has previously exhibited high state uncertainty. Such an approach for providing additional training data based on applied domain knowledge enables rapid improvement of a data-based state model for predicting state variables for a large number of similar devices.

[0029] It can be provided that the operating variables are transmitted from a plurality of technical systems to a central unit, wherein a database with operating characteristic points of real energy storage devices of the plurality of technical systems is provided, wherein the method is carried out in the central unit, wherein the operating characteristic points are selected from the database.

[0030] The operating characteristic point with the insufficient state uncertainty can be determined during an evaluation of a current or a historical operating characteristic point of a specific energy storage device.

[0031] According to a further aspect, a device for training a state model for determining a state of an electrical energy storage device using operating feature points of a plurality of energy storage devices is provided, wherein the device is designed to: - Providing a data-based state model that assigns a modeled state variable, in particular an ageing state, to an operating characteristic point, - Providing a database with histories of real operating characteristic points from the plurality of energy storage devices for successive evaluation periods; - Training or updating the state model depending on at least one training data set, where the at least one training data set is generated with the following steps: ◯ Reducing the state uncertainties of one or more of the state variables resulting from the data-based state model at the real operating characteristic points or from at least one known label, using at least one rule based on domain knowledge; ◯ Providing or determining an operating characteristic point with an insufficient state uncertainty of the associated modeled state variable, ◯ Selecting operating characteristic points from the real operating characteristic points using a clustering procedure depending on the determined operating characteristic point, ◯ Determination of the state variables for the selected operating characteristic points and associated state uncertainties, in particular using the data-based state model or by measurement; ◯ Generating the training data set from the mean value of the determined state variables for the selected operating characteristic points and an operating characteristic point that corresponds to a mean value or a centroid of the selected operating characteristic points. Brief description of the drawings

[0032] Embodiments are explained in more detail below with reference to the attached drawings. They show: Fig. 1 a schematic representation of a system for a vehicle fleet with a plurality of motor vehicles and a central unit for providing a data-based condition model; Fig. 2 a flowchart illustrating a method for training or updating a data-based state model, in particular in the form of an aging state model for a vehicle battery of the motor vehicle; Fig. 3 a time series of modeled aging states with a modeled aging state with a large confidence interval; Fig. 4 an illustration of the reduction of the confidence interval by shifting the confidence limits; Fig. 5 an illustration of the reduction of the confidence interval by shifting the confidence limits using linear interpolation; Fig. 6 an illustration of the clustering of the identified vehicle batteries with similar operating characteristic points; and Fig. 7 shows the improvement of the data-based aging state model by the described method. Description of embodiments

[0033] The method according to the invention is described below using vehicle batteries in a large number of motor vehicles as similar devices. A data-based aging state model for the respective vehicle battery can be implemented in a control unit in the motor vehicles. The aging state model represents an example of a state model. The aging state model can be continuously updated or retrained in a central unit based on operating parameters of the vehicle batteries from the vehicle fleet.

[0034] The above example represents a multitude of stationary or mobile devices with off-grid power supplies, such as vehicles, systems, IoT devices, and the like, which are connected to a central unit (cloud) via a corresponding communication link (e.g., LAN, internet). The state variables represent variables that cannot be easily determined with high accuracy and simultaneously based on models in similar devices, but can only be determined through complex calculations, internal diagnostic, particularly destructive, measurements, or according to predefined operating cycles of the device.

[0035] Fig. 1 shows a system 1 for providing fleet data of motor vehicles 4 of a vehicle fleet 3 in a central unit 2. In the central unit 2, a calculation and a prediction of a course of an aging state of a vehicle battery of a respective motor vehicle 4 of the vehicle fleet 3 is to be carried out based on the fleet data.

[0036] One of the motor vehicles 4 is in Fig. 1. The motor vehicles 4 each have a vehicle battery 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 communications module 44, which is suitable for transmitting data between the respective motor vehicle 4 and a central unit (cloud). The control unit 43 is connected to a sensor unit 45, which has one or more sensors for continuously recording operating variables.

[0037] The central unit 2 has a data processing unit 21 in which the method described below can be carried out, and a database 22 for storing operating variables and aging states of vehicle batteries that have been determined in a plurality of vehicles 4 of the vehicle fleet 3.

[0038] The motor vehicles 4 send the operating variables F to the central unit 2, which at least indicate variables on which the aging state of the vehicle battery depends. In the case of a vehicle battery, the operating variables F can indicate a current battery current, a current battery voltage, a current battery temperature, and a current state of charge (SOC). The operating variables F are recorded in a fast time frame between 2 and 100 Hz, with their curves being regularly transmitted to the central unit 2 in uncompressed and / or compressed form.

[0039] From the operating variables F, operating characteristics can be generated in the central unit 2 or, in other embodiments, already in the respective motor vehicles 4, which relate to an evaluation period. The evaluation period for determining the aging state can range from a few hours (e.g., 6 hours) to several weeks (e.g., one month). A typical value for the evaluation period is one week.

[0040] The operating characteristics can, for example, include characteristics related to the evaluation period and / or accumulated characteristics and / or statistical variables determined over the entire service life to date. In particular, the operating characteristics can include, for example: histogram data on the state of charge curve, the temperature, the battery voltage, the battery current, in particular histogram data regarding the battery temperature distribution over the state of charge, the charging current distribution over the temperature and / or the discharging current distribution over the temperature, accumulated total charge (Ah), an average capacity increase during a charging process (in particular for charging processes in which the charge increase exceeds a threshold proportion (e.g., 20%) of the total battery capacity), a maximum of the differential capacity (dQ / dU: charge change divided by change in battery voltage), and others.

[0041] Further information can be obtained from the operating characteristics: a temporal load pattern such as charging and driving cycles, determined by usage patterns (such as rapid charging at high currents or strong acceleration or braking with recuperation), a usage time of the vehicle battery, a cumulative charge quantity over the running time and a cumulative discharge quantity over the running time, a maximum charging current, a maximum discharging current, a charging frequency, an average charging current, an average discharging current, a power throughput during charging and discharging, a (particularly average) charging temperature, a (particularly average) spread of the state of charge and the like.

[0042] The state of health (SOH) is the key parameter for indicating the remaining battery capacity or remaining battery charge. The state of health represents the aging of the vehicle battery, a battery module, or a battery cell and can be expressed as the capacity retention rate (SOH-C) or as the increase in internal resistance (SOH-R). The capacity retention rate (SOH-C) is expressed as the ratio of the measured instantaneous capacity to the initial capacity of the fully charged battery. The relative change in internal resistance (SOH-R) increases with increasing battery age.

[0043] In Fig. 2 uses a flowchart to describe the procedure for training or updating the data-based aging state model.

[0044] The method is executed in the data processing unit 21 of the central unit and can be implemented there as software and / or hardware.

[0045] In step S1, a data-based aging state model is provided, which is pre-trained to output a modeled aging state and a state inaccuracy depending on operating characteristic points. Furthermore, a database is provided that provides time series of operating characteristic points for each evaluation period for the electrical energy storage devices of the vehicles in the fleet.

[0046] In step S2, operating variables F are continuously received from the vehicles 4 of the vehicle fleet 3. The operating variables F can include curves of the current battery current, the current battery voltage, the current battery temperature, the current state of charge at the pack, module, and cell levels, and the like.

[0047] In step S3, operating characteristics for successive evaluation periods are determined from the courses of the operating variables F. In the case of determining the aging state for the vehicle battery, these evaluation periods can range from one day to one month, preferably one week. Typically, for the model-based determination of the aging state of a specific vehicle battery, an aging state is assigned as a state variable to an operating characteristic point, which is determined by the individual operating characteristics for a specific vehicle battery within an evaluation period, using the data-based aging state model. This data-based model can also be implemented as a hybrid model, in particular a combination of physical and data-driven models, preferably with a supervised learning component.

[0048] In step S4, an operating characteristic point is determined for which the condition uncertainty is insufficient, i.e., in particular, it exceeds a predefined uncertainty threshold. For this purpose, the data-based aging condition model can be provided as a model that, in addition to a model prediction, can also specify a prediction uncertainty, e.g., in the form of a standard deviation or a confidence interval.

[0049] For example, hybrid models or purely data-based models can be considered, which include supervised learning methods, preferably Gaussian process models or alternatively ensembler methods or Bayesian neural networks.

[0050] The operating characteristic point for which a calculated state uncertainty is insufficient can be determined, for example, by determining and evaluating modeled aging states of all vehicle batteries at different evaluation periods using the data-based aging state model. In other words, aging states and their modeled state uncertainties are determined for each real operating characteristic point, i.e., for a real vehicle battery. If the state uncertainty for one of the modeled aging states exceeds a specified uncertainty threshold, the respective real operating characteristic point is identified as an operating characteristic point for which the state uncertainty is insufficient.For the present exemplary embodiment, the indication of the aging state for the determined (as described above, identified) operating characteristic point of the specific vehicle battery is thus recognized as too inaccurate.

[0051] For example, in Fig. Figure 3 shows a diagram showing a time series of aging states Z, each with an associated confidence interval CI for successive evaluation periods. It can be seen that there is a high degree of uncertainty at time t, while there is a low degree of uncertainty for the previous evaluation period tn and the subsequent evaluation period t+n.

[0052] In step S5, a vehicle-specific reduction of the corresponding confidence interval KI is performed using domain knowledge for each of the real operating characteristic points, or at least for those real operating characteristic points whose state uncertainty lies above a specified uncertainty threshold. For this purpose, the temporal progression of the aging state (always for the individual vehicle battery) can be analyzed around the evaluation period of the real operating characteristic point under consideration. Each operating characteristic point is assigned to an evaluation period, with additional operating characteristic points for the vehicle battery in question generally available before and after the evaluation period.

[0053] In particular, the confidence interval can be restricted to aging states that are not greater than an upper limit of the confidence interval of a previously determined or known aging state and not lower than a lower limit of an aging state determined in a later (subsequent) evaluation period. This is exemplified in Fig. 4, where the limited confidence interval is denoted by CI'. Such a restriction can be made based on the domain knowledge that temporal courses of aging states can only develop monotonically decreasingly given a sufficiently long time window, especially considering the effective stress factors that contribute to aging. Thus, the confidence limits G of the aging state defined as uncertain are shifted toward the statistical expectation of the modeling value of the aging state, resulting in a reduced confidence interval.

[0054] Other rules derivable from domain knowledge for restricting the confidence intervals can also be defined. For example, a continuous and sluggish, or piecewise nearly linear, progression of an aging condition within a limited period of a few, for example, between 3 and 10, evaluation periods can be assumed as a further possible rule. Such a restriction can be applied if, for example, the fleet load profile shows that the load profile and the associated stress factors related to aging for the vehicle under consideration, with regard to energy throughput, temperature conditions, charging behavior, driving behavior, and the like, were constant during the period under consideration between the previous evaluation period and the subsequent evaluation period.Thus, an upper and lower limit of a limited confidence interval CI' for the respective aging state can be determined by piecewise interpolation of the upper confidence limits of a previous and a subsequent evaluation period and the lower confidence limits of the previous and subsequent evaluation periods. This is illustrated by way of example in . Fig. 5 shown.

[0055] In this way, the confidence intervals for the aging states at the real operating characteristic points of the vehicle batteries can be limited. The high uncertainty at real operating characteristic points can be reduced through domain knowledge of the temporal inertia or the temporal course of the state variable. Thus, knowledge about low state uncertainties at evaluation times can be traced back either to precise modeling of the aging state using the aging state model or to the results of a measurement, or a combination of both.

[0056] In step S6, a clustering method is used to identify from the real operating characteristic points those operating characteristic points that are similar to the determined operating characteristic point. All real operating characteristic points from the database that were determined for previous evaluation periods of all vehicle batteries 41 are taken into account. The clustering method determines from the real operating characteristic points the real operating characteristic points that are similar to the determined operating characteristic point, in particular operating characteristic points with a predetermined Euclidean distance from a centroid point that is not greater than a predetermined threshold value. The centroid point corresponds to an artificial operating characteristic point that is not assigned to a specific energy storage device. Similarity is defined via operating characteristic points. If, for example,If the Euclidean distance between operating characteristic points in a multidimensional space is sufficiently small, for example, if it lies below a threshold, which may be linked to an accuracy requirement, then the two are sufficiently similar. In other words, the Euclidean distance determines the degree of similarity between two operating characteristic points. Two operating characteristic points are similar if the degree of similarity lies above a specified threshold.

[0057] K-means++ and competitive learning can be used as possible unsupervised clustering methods.

[0058] In step S7, an estimate of the respective aging state can be made for each of the selected operating characteristic points according to the data-based state model. Furthermore, the model estimate results in a respective confidence interval as the state uncertainty.

[0059] A selection of real operating characteristic points for identified vehicle batteries in a range of comparable (similar) operating characteristic points is now available, each of which is assigned an aging state with a confidence interval. The reduced confidence interval 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 possible, for example, in Fig. 6. Each of the areas B shown corresponds to an operating characteristic point MP with a corresponding limited confidence interval.

[0060] The aging states assigned to the selected operating characteristic points can now be averaged across all vehicles in step S8 and the confidence intervals can be calculated accordingly in an overall confidence interval or in a common standard deviation σ gesbe consolidated, for example, for confidence intervals of the same size using the root-n law σtotal=σn., where σ corresponds to the same standard deviation and n to the number of aging states considered.

[0061] If different, limited confidence intervals are present, the overall confidence interval can also be calculated using error propagation to determine the variance or standard deviation of the averaged aging state, for example, using model-based error propagation that exploits model sensitivity to calculate the combined label uncertainty. This can be done, for example, by evaluating and combining the partial derivatives of the data-based state model, for example, using a Taylor series expansion.

[0062] In a subsequent step S9, a check is performed to determine whether the total confidence interval thus determined falls below a predefined maximum confidence interval, which specifies the maximum permissible size of the confidence interval. If this is the case (alternative: yes), the method continues with step S10. Otherwise (alternative: no), the method returns to step S1.

[0063] In step S10, a new training data set is generated based on the mean value or centroid (operating characteristic point representing the centroid of the respective cluster) of the cluster of identified vehicle batteries 41 or another operating characteristic point associated with the cluster and the mean value of the associated aging states.

[0064] In a subsequent step S11, the data-based aging state model is retrained or updated based on the new training data set. It is also possible to initially determine several training data sets obtained in this way and / or to collect training data sets over a specific period of time before an update takes place.

[0065] Fig. Figure 7 shows the prediction of the aging state for a specific operating characteristic point MP before (1) and after (2) the update of the data-based aging state model.

[0066] In a subsequent step S12, the model parameters of the updated aging state model or the states determined by the model can be transmitted back to the vehicles of vehicle fleet 3, so that the data-based aging state model can be used in motor vehicle 4 to determine the aging state SOH of vehicle battery 41. This step is optional. The aging state model can also be operated and executed in central unit 2.

Claims

[1] Computer-implemented method for training a state model for determining a state of an electrical energy storage device using operating feature points of a plurality of energy storage devices, comprising the following steps: - Providing a data-based state model that assigns a modeled state variable, in particular an ageing state, to an operating characteristic point, - Providing a database with histories of real operating characteristic points from the plurality of energy storage devices for successive evaluation periods; - Training or updating the state model depending on at least one training data set, wherein the at least one training data set is generated with the following steps: ◯ Reducing the state uncertainties of one or more of the state variables resulting from the data-based state model at the real operating characteristic points using at least one rule based on domain knowledge; ◯ Providing or determining an operating characteristic point with an insufficient state uncertainty of the associated modeled state variable or by at least one known label, ◯ Selecting operating characteristic points from the real operating characteristic points using a clustering procedure depending on the determined operating characteristic point, ◯ Determination of the state variables for the selected operating characteristic points and associated state uncertainties, in particular using the data-based state model or by measurement; ◯ Generating the training data set from the mean value of the determined state variables for the selected operating characteristic points and an operating characteristic point that corresponds to a mean value or a centroid of the selected operating characteristic points. [2] Method according to claim 1, wherein for selecting operating feature points from the real operating feature points using a clustering method, those operating feature points are selected which are similar to the determined operating feature point with the insufficient state uncertainty. [3] Method according to claim 1 or 2, wherein the state model corresponds to an aging state model for providing an aging state depending on the operating characteristics or a state of charge model for providing a state of charge depending on the operating characteristics of the energy storage device. [4] Method according to one of claims 1 to 3, wherein the data-based state model comprises a data-based machine learning model configured to specify a state uncertainty for a modeled state variable, wherein the state model comprises a supervised learning method, such as a Gaussian process model or a Bayesian neural network. [5] Method according to one of claims 1 to 4, wherein the operating characteristics are derived from time series of continuously recorded operating variables, wherein the operating characteristics are determined for successive evaluation periods, wherein a plurality of operating characteristics define an operating characteristic point for a specific evaluation period of a specific energy storage device. [6] Method according to claim 5, wherein the operating variables are transmitted from a plurality of technical systems to a central unit, wherein a database with operating characteristic points of real energy storage devices of the plurality of technical systems is provided, wherein the method is carried out in the central unit, wherein the selection of the operating characteristic points takes place from the database. [7] Method according to one of claims 1 to 5, wherein the providing or determining of the operating characteristic point with the insufficient state uncertainty occurs during an evaluation of a current operating characteristic point for a specific energy storage device. [8] Method according to one of claims 1 to 7, wherein the reduction of the state uncertainties at the selected operating characteristic points is carried out by at least one rule which depends on a temporal development of the operating characteristic points of the respective energy storage device and state uncertainties of operating characteristic points of the respective energy storage device at evaluation periods other than the evaluation periods of the selected operating characteristic points. [9] Method according to one of claims 1 to 8, wherein the state uncertainties at the real operating characteristic points are reduced by limiting the state uncertainty at each of the selected operating characteristic points to an upper or lower limit of a confidence interval which is determined by the state uncertainty at an evaluation period preceding the relevant operating characteristic point, and / or by limiting the state uncertainty at each of the selected operating characteristic points to a lower or upper limit of a confidence interval which is determined by the state uncertainty at an evaluation period following the relevant operating characteristic point. [10] Method according to one of claims 1 to 9, wherein the state uncertainties at the selected operating characteristic points are reduced by limiting the state uncertainty at each of the selected operating characteristic points with an upper and a lower limit of a corresponding confidence interval, which is determined by interpolating upper limits of confidence intervals of evaluation periods preceding and following the operating characteristic point in question. [11] Method according to one of claims 1 to 10, wherein the training or updating of the state model based on the at least one training data set is only carried out if a total state uncertainty does not exceed a predetermined uncertainty threshold, wherein the total state uncertainty is determined from the state uncertainties of the selected operating feature points according to the law of large numbers or by means of an error propagation method. [12] Method according to one of claims 1 to 11, wherein model parameters of the state model are transmitted to a plurality of devices with energy storage devices. [13] Method according to one of claims 1 to 12, wherein the state model is implemented in a central unit which is in communication with a plurality of devices, wherein the state variable of each of the plurality of devices can be queried by transmitting operating variables to the central unit and the central unit transmits the state variable determined with the aid of the state model to the requesting device. [14] Device for training a state model for determining a state of an electrical energy storage device using operating feature points of a plurality of energy storage devices, the device being designed to: - Providing a data-based state model that assigns a modeled state variable, in particular an ageing state, to an operating characteristic point, - Providing a database with histories of real operating characteristic points from the plurality of energy storage devices for successive evaluation periods; - Training or updating the state model depending on at least one training data set, wherein the at least one training data set is generated with the following steps: o Reducing the state uncertainties of one or more of the state variables resulting from the data-based state model at the real operating characteristic points or from at least one known label, using at least one rule based on domain knowledge; ◯ Providing or determining an operating characteristic point with an insufficient state uncertainty of the associated modeled state variable, ◯ Selecting operating characteristic points from the real operating characteristic points using a clustering procedure depending on the determined operating characteristic point, ◯ Determination of the state variables for the selected operating characteristic points and associated state uncertainties, in particular using the data-based state model or by measurement; ◯ Generating the training data set from the mean value of the determined state variables for the selected operating characteristic points and an operating characteristic point that corresponds to a mean value or a centroid of the selected operating characteristic points. [15] Computer program product comprising instructions which, when the program is executed by at least one data processing device, cause the device to carry out the steps of the method according to one of claims 1 to 12. [16] Machine-readable storage medium comprising instructions which, when executed by at least one data processing device, cause it to carry out the steps of the method according to one of claims 1 to 12.

Images