Method for the computer-aided evaluation of measurements of electrical currents in a high-voltage electrical system of a given electrically powered motor vehicle

Abstract

Method for computer-aided evaluation of measurements of electrical currents (i1, i2, ..., i5) in a high-voltage electrical system (BN) of a given electrically powered motor vehicle (VE), wherein the high-voltage electrical system (BN) comprises a plurality of electrical components (HVB, EM, HE, CO, IN) including an electric machine (EM) for driving the given motor vehicle (VE) and an electrical energy storage device (HVB) for supplying electrical energy to the electric machine (EM), and each electric current (i1, i2, ..., i5) flows through the same node (N) in the high-voltage electrical system (BN), wherein: a) in the specified motor vehicle (VE), a first data-driven model (MO) is learned using a machine learning method (LE), and the learned first data-driven model (MO) is transmitted to a backend server (SE) by the specified motor vehicle (VE), wherein the first data-driven model (MO) is learned based on a plurality of measurement data sets (MD), each of which was measured during the operation of the specified motor vehicle (VE) at different times of operation, wherein each measurement data set (MD) for a respective time of operation comprises measured values ​​of input variables (u1, u2, u3) and measured values ​​of the electric currents (i1, i2, ..., i5), wherein the learned first data-driven model (MO) can predict measured values ​​of the electric currents (i1, i2, ..., i5) based on measured values ​​of the input variables (u1, u2, u3); b) in the backend server (SE) the learned first data-driven model (MO) and a plurality of further learned second data-driven models (MO'), each of which was learned in a different motor vehicle (VE') of the same type as the specified motor vehicle (VE) using the same machine learning method (LE) as the first data-driven model (MO) based on a plurality of measurement data sets measured in the respective other motor vehicle (VE') for the same input variables (u1, u2, u3) and the same electrical currents (i1, i2, ..., i5) as the first data-driven model (MO), are processed by calculating current averages (is) for the electrical currents (i1, i2, ...) for a model driving measurement series (MUD), which contains measurement values ​​for the input variables used in learning the first data-driven model (u1, u2, u3) for a plurality of model driving operating times within a specified model driving., i5) are determined for each model operation time, wherein a respective current mean value (is) for a respective electric current (i1, i2, ..., i5) and for a respective model operation time is the mean value of the measured values ​​of the respective electric current (i1, i2, ..., i5), which are predicted by the first learned data-driven model (MO) and the plurality of second learned data-driven models (MO') based on the measured values ​​of the input variables (u1, u2, u3) for the respective model operation time; . c) at least in the case that, depending on the measurement data sets (MD) measured in the specified motor vehicle (VE), abnormal operating behavior of the specified motor vehicle (VE) and the presence of measurement errors in the measurement of the electric currents (i1, i2, ..., i5) in the specified motor vehicle are detected, a parameter vector (PV) is determined from parameter values ​​(pw1) for parameters of a first function (f1) and from parameter values ​​(pw2) for parameters of a second function (f2) using the measured values ​​of the respective electric currents (i1, i2, ..., i5) predicted by the first learned data-driven model (MO) and the average current values ​​(is) for the respective electric currents (i1, i2, ..., i5), wherein the first function (f1) calculates current values ​​of the electric currents (i1, i2, ..., i5) under normal operating behavior of the specified motor vehicle (VE). ..., i5) in the case of abnormal operating behavior of the specified motor vehicle (VE) and wherein the second function (f2) maps current values ​​of the electric currents (i1, i2, ..., i5) in the absence of measurement errors to current values ​​of the electric currents (i1, i2, ..., i5) in the presence of measurement errors in the specified motor vehicle (VE), wherein the parameter values ​​(pw1, pw2) are determined by means of an optimization with the optimization goal of a minimum number of parameter values ​​(pw1, pw2) not equal to zero in the parameter vector (PV), wherein the parameter values ​​(pw2) of the parameters of the second function (f2) are transmitted to the specified motor vehicle (VE) for the correction of future measured values ​​of at least a part of the electric currents (i1, i2, ..., i5).

Description

[0001] The invention relates to a method and a system for computer-aided evaluation of measurements of electrical currents in a high-voltage electrical system of a given electrically powered motor vehicle.

[0002] In electrically powered vehicles, such as pure electric vehicles or hybrid vehicles, the precise determination of electrical currents in the high-voltage electrical system is of paramount importance. The high-voltage electrical system is the vehicle's electrical network used for electric driving. It comprises an electric drive motor and an energy storage device for this motor. If the corresponding electrical currents are not determined with sufficient accuracy, this must be taken into account when designing the electrical energy storage device within the high-voltage electrical system, resulting in inefficient use of the energy storage system.

[0003] When considering the measured currents at a node in a high-voltage electrical system, Kirchhoff's current law (Kirchhoff's current law) dictates that the current at this node should be 0 A. However, in conventional high-voltage electrical systems, a value significantly deviating from zero is typically measured, which is due to measurement errors. This value might be, for example, 1 A, resulting in a power dissipation of 400 W in a 400 V system. It is desirable to correct this power dissipation so that the storage protection limits of the electrical energy storage device in the high-voltage electrical system can be set closer to the actual physical load limits. These storage protection limits take into account not only the actual physical load limit but also the maximum possible power dissipation.In the motor operation of the vehicle, this means that an overly conservative storage protection limit results in drive power not being released during full acceleration requested by the driver, even though it would be available. In the case of recuperation, taking the maximum power loss into account means that the electrical energy storage device cannot be charged to its physical charge limit.

[0004] Publication [1] describes a fleet-based approach for detecting errors in the measurement of electrical currents in an electrically powered vehicle. The electrical current of the electric drive motor is predicted using a data-driven model and compared with corresponding predicted currents predicted by data-driven models in other vehicles. While this method can classify error sources, it cannot correct measurement errors in the respective vehicle.

[0005] The publication [3] discloses a method and system for monitoring the condition of a vehicle using machine learning and data mining technologies on data collected from multiple vehicles.

[0006] In publication [4] a method for estimating the range of an electric vehicle is described using a model that is trained by machine learning.

[0007] The object of the invention is to create a method for evaluating measurements of electrical currents in a high-voltage electrical system of a given electrically powered motor vehicle, with which measurement errors can be reliably detected and corrected.

[0008] The method according to the invention serves for the computer-aided evaluation of measurements of electrical currents in a high-voltage electrical system of a given electrically powered motor vehicle. The term "electrically powered motor vehicle" is to be interpreted broadly. An electrically powered motor vehicle can be a purely electric vehicle, but also a hybrid vehicle, and in particular a plug-in hybrid vehicle. The decisive factor is that the motor vehicle has an electric drive, although an internal combustion engine may also be provided. The high-voltage electrical system, which, for example, operates at a voltage of 400 V, is used for the electric driving of the given motor vehicle.It comprises a plurality of electrical components, including an electric motor for propelling the specified vehicle and an electrical energy storage device for supplying this electric motor with electrical energy. Each electric current flows through the same node in the high-voltage electrical system. Preferably, each electric current occurs at a different electrical component among the plurality of electrical components.

[0009] In the method according to the invention, in step a) a first data-driven model is learned in the specified motor vehicle using a machine learning method and this learned first data-driven model is transmitted to a backend server by the specified motor vehicle, wherein a backend server is understood to be a computer system not belonging to the motor vehicle consisting of one or possibly several interconnected computers.

[0010] The first data-driven model is learned based on a large number of measurement data sets, each measured during the operation of the specified vehicle at different times. Each measurement data set for a given operating time includes measured values ​​of input variables and measured values ​​of electrical currents. The learned first data-driven model can then predict measured values ​​of the electrical currents based on the measured values ​​of the input variables.

[0011] In one embodiment of the invention, the first data-driven model is updated at regular intervals in the specified vehicle based on newly acquired measurement data sets. The processing steps described below in the backend server are then repeated when an updated, learned first data-driven model is available.

[0012] In step b), the learned first data-driven model and a plurality of further learned second data-driven models are processed in the backend server. Each second data-driven model was learned in a different vehicle of the same type as the specified vehicle, using the same machine learning method as the first data-driven model. This learning process is based on a large number of measurement data sets measured in the respective other vehicle for the same input variables and the same electrical currents as the first data-driven model. During processing in the backend server, average current values ​​for the electrical currents are determined for each model driving measurement series. This series contains measurement values ​​for the input variables used in learning the first data-driven model and for a large number of operating points within a specified model driving scenario.The average current value for a given electric current and for a given point in time during model operation is the average of the measured values ​​of that electric current, predicted by the first learned data-driven model and the multitude of second data-driven models based on the measured values ​​of the input variables for that point in time during model operation. In other words, the average current value is an average of the predicted measured values ​​across the first and second data-driven models. A model operation is understood to be a predefined operation with specific operating parameters, which are defined as the values ​​of input variables at corresponding points in time during model operation.

[0013] After completing step b), the subsequent step c) is performed under certain conditions or, if necessary, always. Step c) is always performed if, based on the measurement data sets taken in the specified vehicle, abnormal operating behavior of the specified vehicle and the presence of measurement errors in the measurement of the electrical currents in the specified vehicle are detected. According to step c), a parameter vector is determined using the measured values ​​of the respective electrical currents and the average current values ​​for the respective electrical currents, as predicted by the first learned data-driven model. This parameter vector consists of parameter values ​​for parameters of a first function and parameter values ​​for parameters of a second function.These functions can be modeled in different ways; preferably, a linear function is used as the basis for the first and / or second function. The first function maps current values ​​of the electric currents under normal operating conditions of the given vehicle to current values ​​of the electric currents under abnormal operating conditions of the given vehicle, whereas the second function maps current values ​​of the electric currents in the absence of measurement errors to current values ​​of the electric currents when measurement errors are present in the given vehicle. Preferably, the first function is designed such that if the sum of the current values ​​of all electric currents to be mapped by the first function is zero, then the sum of the current values ​​of all electric currents resulting from the mapping with the first function is also zero.This approach takes advantage of the fact that Kirchhoff's node rule is not affected by abnormal operating behavior (i.e., by hardware errors in the given vehicle).

[0014] The parameter values ​​in step c) are determined by means of an optimization with the optimization goal of a minimum number of non-zero parameter values ​​in the parameter vector, whereby the parameter values ​​of the parameters of the second function are transmitted to the specified motor vehicle for the correction of future measured values ​​of at least part of the electric currents.

[0015] According to the invention, the knowledge is utilized that, in a fleet-based approach, suitable corrections for measurement errors can be determined using appropriately predicted measured values ​​and average current values ​​determined for a model drive. These corrections are reflected in the parameters of a corresponding second function. For the correction, after transmitting the parameters of the second function to the specified vehicle, the inverse of the second function is determined in the specified vehicle, whereby the value of this function for a corresponding measured value in the specified vehicle yields the corrected measured value.

[0016] Furthermore, the invention is based on the finding that corresponding parameter vectors cannot be uniquely determined without further boundary conditions. Therefore, a suitable optimization using the criterion of sparse population of the parameter vector is employed. The optimization can be carried out, for example, using the method described in document [2].

[0017] In a particularly preferred embodiment of the method according to the invention, a classification into the following classes is carried out before step c) is performed, depending on the measurement data sets measured in the specified motor vehicle: - a first class, according to which there is neither abnormal operating behavior of the specified motor vehicle nor measurement errors for the measured values ​​of the electrical currents in the specified motor vehicle; - a second class, according to which abnormal operating behavior of the specified motor vehicle exists, but no measurement errors are present for the measured values ​​of the electrical currents in the specified motor vehicle; - a third class, according to which there is no abnormal operating behavior of the specified motor vehicle, but measurement errors are present for the measured values ​​of the electrical currents in the specified motor vehicle; - a fourth class, according to which abnormal operating behavior of the specified motor vehicle exists and measurement errors exist for the measured values ​​of the electrical currents in the specified motor vehicle.

[0018] In this embodiment, the process is terminated if the first class or the second class is present, because the process aims to correct measurement errors that are not present when classified into the first and second classes.

[0019] In a particularly preferred embodiment of the above, the classification is based on first features and a second feature, wherein each first feature is associated with an electric current and is a measure of the absolute deviation of the measured values ​​of the associated electric current, predicted by the first learned data-driven model, from the average current values ​​for the model operation times. In contrast, the second feature is a measure of the magnitude of the sums of the measured values ​​of the electric currents for the respective measurement data sets at the operating time of the corresponding measurement data set. With such features, the above classification can be carried out in a simple manner.

[0020] In a preferred embodiment of the variant just described, classification is performed using a threshold classifier, whereby abnormal operating behavior of the specified motor vehicle is present when a first value, which increases with the size of the first features and decreases with the size of the second features, exceeds a first threshold. Conversely, measurement errors for the measured values ​​of the electrical currents in the specified motor vehicle are present when a second value, which increases with the size of the second feature, exceeds a second threshold.

[0021] Instead of the threshold classifier described above, a different classifier can also be used in another embodiment of the invention. In particular, the classification can optionally also be carried out using a machine-learned decision tree classifier, which is trained using training datasets comprising first features, a second feature, and a corresponding class from the first to fourth classes. Such decision tree classifiers are well known and are therefore not described in further detail.

[0022] In a further, particularly preferred embodiment of the invention, if the third class mentioned above is present, instead of the parameter vector of step c), a reduced parameter vector is determined from parameter values ​​for the parameters of the second function. In other words, in this case, not step c), but a modified step is performed in which a parameter vector with fewer parameter values ​​is determined. This parameter vector contains only parameter values ​​for the parameters of the second function and not for the parameters of the first function. In this case, the parameter values ​​can be determined by regressing the measured values ​​of the electrical currents predicted by the first data-driven model onto the average current values ​​for the model operation times.The parameter values ​​of the second function are then transmitted to the specified vehicle to correct future measured values ​​of at least some of the electrical currents. This correction can be performed by calculating the inverse of the function with the corresponding parameter values ​​and using this inverse to determine a correction value for a given measured current value.

[0023] As explained above, the high-voltage electrical system includes at least the electric motor and the electrical energy storage device. Preferably, however, the high-voltage electrical system also includes further electrical components. These components are, in particular, a heater and / or an air conditioner and / or an inverter for voltage conversion to a low-voltage electrical system, e.g., a 12 V system. Each of the aforementioned components may be present multiple times in the high-voltage electrical system.

[0024] Depending on the specific configuration of the method according to the invention, various types of first and second data-driven models can be learned. Preferably, the first data-driven model and the majority of second data-driven models each represent a neural network structure known per se, preferably a so-called LSTM network structure (LSTM = Low Short-Term Memory). An LSTM network structure enables regressive learning, which allows the data-driven models to be updated easily.

[0025] In an alternative, particularly preferred variant, the first data-driven model and the majority of second data-driven models each represent a state-space model. Such data-driven models are also known from the prior art. A data-driven model based on a state-space model is explained in more detail below.

[0026] In addition to the method described above, the invention relates to a system for the computer-aided evaluation of measurements of electrical currents in a high-voltage electrical system of a given electrically powered motor vehicle, wherein the high-voltage electrical system comprises a plurality of electrical components, including an electric motor for driving the given motor vehicle and an electrical energy storage device for supplying electrical energy to the electric motor, with each electrical current flowing through the same node in the high-voltage electrical system. This system is configured to carry out the method according to the invention or one or more preferred variants thereof.

[0027] The invention further relates to a backend server configured to function as a backend server in the method according to the invention or in one or more preferred embodiments thereof. In other words, the backend server includes means for carrying out those process steps that it performs in the method according to the invention or in one or more preferred embodiments thereof.

[0028] The invention further relates to an electrically powered motor vehicle, wherein the motor vehicle is configured to operate as a predetermined motor vehicle in the method according to the invention or in one or more preferred embodiments of the method according to the invention. In other words, the motor vehicle includes means for carrying out the steps performed by the predetermined motor vehicle in the method according to the invention or in one or more preferred variants of the method according to the invention.

[0029] An embodiment of the invention is described in detail below with reference to the accompanying figures.

[0030] They show: Fig. 1 a schematic representation of a high-voltage electrical system in an electrically powered motor vehicle, wherein a variant of the method according to the invention is used for this electrical system; Fig. 2 a schematic representation illustrating the process of a variant of the method according to the invention; and Fig. 3 a diagram illustrating the consideration of Kirchhoff's node rule for a classifier used in an embodiment of the method according to the invention.

[0031] An embodiment of the inventive method for a high-voltage electrical system is described below, schematically in Fig. Figure 1 shows the high-voltage electrical system BN, which is represented there by the reference symbol BN. The high-voltage electrical system BN comprises a multitude of electrical components, but for the sake of simplicity, only five electrical components will be considered in the following description. The electrical system BN is installed in an electrically powered vehicle and includes an electric motor EM for propelling the vehicle. This motor is powered by the electrical energy from a high-voltage battery HVB. An electric current i1 occurs at the high-voltage battery, and an electric current i2 occurs at the electric motor.

[0032] Furthermore, several electrical consumers are part of the high-voltage electrical system BN. In particular, an electric heater HE, an air conditioner CO, and an inverter IN for voltage conversion to a low-voltage electrical system are provided. The electric current i3 occurs at the electric heater HE, the electric current i4 at the air conditioner CO, and the electric current i5 at the voltage inverter. Corresponding measured values ​​of these electric currents are taken at node N. According to Kirchhoff's current law, the sum of the currents i1 to i5 at node N should be zero, provided no measurement errors occur in the acquisition of these currents. The aim of the variant of the method according to the invention described here is to detect and appropriately correct such measurement errors in the currents i1 to i5.

[0033] Fig. Figure 2 shows the general procedure. The following considers a given motor vehicle VE in which the high-voltage electrical system BN of the Fig. 1 is installed. In the procedure, in addition to this vehicle, other structurally identical vehicles VE' from a fleet are considered, for which corresponding corrections for the current values ​​of the high-voltage electrical systems installed in them are determined in the same way as for the specified vehicle VE. However, the correction is described below only for the specified vehicle VE.

[0034] During a journey of the specified vehicle VE, corresponding measured values ​​of the currents i1 to i5 are recorded at numerous operating times using current sensors at node N. Furthermore, measured values ​​of other quantities, referred to below as input quantities, are also recorded at the respective operating times. These quantities are related to the currents occurring. In the variant described here, the acceleration, speed, and inclination of the specified vehicle relative to the horizontal are considered as input quantities. These quantities are in Fig. 1 is denoted by u1, u2, and u3. If necessary, other input variables that are related to the measured currents can also be considered. For example, the gear selected in the vehicle, the power demand and / or allocation to the electric motor, or the rotational speed of the electric motor can also be used as input variables.

[0035] The recorded measured values ​​of the input variables and the corresponding currents result in a multitude of measurement data sets for each operating time. These measurement data sets are in Fig. 1, denoted by MD. In the embodiment described here, they are used as training data for a data-driven model MO, which is learned in the motor vehicle VE using a known learning method LE. Depending on the configuration, different data-driven models can be used. In the embodiment described here, a data-driven model in the form of a state-space model is considered. Such models with the associated learning methods are known to those skilled in the art. The state-space model used in the variant described here is explained in more detail below.

[0036] The result of the learning process LE is the learned data-driven model MO, which corresponds to a first data-driven model as defined in the patent claims. The model comprises parameters in the form of corresponding matrices. After learning, the parameters, and thus the data-driven model, are transferred to a backend server SE. This server contains further data-driven models MO' of the other vehicles VE' in the fleet. These data-driven models, which correspond to second data-driven models as defined in the patent claims, were learned in the same way as for vehicle VE, using corresponding measurement data from the respective vehicle, and then transferred to the backend server SE.

[0037] In step S1, average current values ​​are determined in the backend server SE using the data-driven models MO and MO' for a model drive measurement series MUD of a model drive UD (UD = Unity Drive). The model drive was predefined and accurately reflects the typical operating conditions of the vehicles VE and VE' under consideration. The model drive measurement series contains measured values ​​of corresponding input variables u1, u2, and u3 for a multitude of model drive operating times. In step S1, the corresponding measured values ​​of currents i1 to i5 at the model drive operating times are predicted using the measured values ​​of input variables u1 to u3 with all data-driven models MO and MO'. This yields predicted measured values ​​for all data-driven models at the respective model drive operating time.These predicted measurements are averaged across all data-driven models, resulting in a multitude of average current values ​​(is) for each model operating time. In other words, average current values ​​(is) exist for currents i1 to i5 at every model operating time.

[0038] A corresponding average current value can be considered an approximation of the so-called "ground truth," which represents the actual current values ​​occurring during the model operation of a perfect vehicle—that is, a vehicle in which neither measurement errors in the current measurements nor hardware faults occur. It is assumed that these errors average out when calculated over a large number of vehicles. Hardware faults result from a malfunction of one or more hardware components in the respective vehicle. Such faults cause the vehicle's operating behavior to deviate from normal operating behavior (i.e., it exhibits abnormal operating behavior).

[0039] The average current values ​​obtained for the model operation times, as well as the current measurements used to learn the first data-driven model MO, which are transferred to the backend server SE along with the learned data-driven model MO, and also the current measurements of the vehicle VE predicted in step S1, are subsequently processed in step S2 by a classifier, according to which a classification into the four classes explained below is performed. The classification step S2 is explained in more detail below. According to the classification step, the following classes are considered: - the CL1 class, according to which there are neither hardware errors nor measurement errors for the measured values ​​of the electrical currents in the specified motor vehicle; - a class CL2, according to which hardware faults exist in the specified motor vehicle, but no measurement errors have occurred for the measured values ​​of the electrical currents in the specified motor vehicle; - a class CL3, according to which there are no hardware faults in the specified motor vehicle, but measurement errors have occurred for the measured values ​​of the electrical currents in the specified motor vehicle; - a class CL4, according to which hardware faults have occurred in the specified motor vehicle and furthermore, measurement errors are also present for the measured values ​​of the electrical currents in the specified motor vehicle.

[0040] If a classification into class CL1 or CL2 occurs, the procedure is terminated without further steps, as no measurement errors occur in these classes and therefore no correction is necessary. If class CL4 is present, a parameter vector PV is determined based on a combination of two functions f1 and f2. This parameter vector specifies parameter values ​​pw1 for parameters of function f1 and parameter values ​​pw2 for parameters of function f2. According to function f1, hardware errors are taken into account; that is, the function describes a mapping from the current values ​​of the electrical currents i1 to i5 without hardware errors to current values ​​when hardware errors occur. The first function f1 uses the restriction based on Kirchhoff's current law, according to which if the sum of the current values ​​to be mapped by function f1 is zero, then the sum of the current values ​​resulting from the mapping with function f1 is also zero.This takes into account that hardware failures do not change physical laws and therefore do not alter Kirchhoff's node rule.

[0041] Unlike function f1, function f2 takes measurement errors into account, so the restriction given above by Kirchhoff's current law does not apply to this function. Function f2 maps current values ​​without measurement errors to corresponding current values ​​with measurement errors in the given vehicle. The parameter values ​​pw1 and pw2 of the parameter vector PV are determined using a special mathematical method that assumes the parameter vector has a sparse population of non-zero parameter values. This will be explained in more detail below.

[0042] After determining the parameter vector PV, the parameter values ​​pw2 of the second function f2, i.e., the parameter values ​​relating to measurement errors, are transferred to the specified vehicle VE. Subsequently, the vehicle VE can use these parameter values ​​pw2 to correct the measured values ​​of the electric currents i1 to i5. This is achieved simply by inverting the function f2, which underlies the parameter values ​​pw2, within the vehicle VE. This allows for the determination of corrected current values ​​from the current values ​​containing measurement errors, thus achieving higher accuracy in the measurement of the currents within the vehicle VE.

[0043] If the classification in step S2 results in class CL3, determining the corresponding parameters is simpler than in the case of class CL4. Specifically, it is no longer necessary to consider the function f1, which models hardware errors, because no hardware errors are present in class CL3. Therefore, a parameter vector PV' with fewer parameter values ​​pw' is used only for the function f2. The function f2 can be determined by regressing the current values ​​predicted by the first data-driven model MO onto the average current values ​​is. In the case of classification into class CL4, the parameter values ​​pw' are transmitted to the vehicle VE and can be used there to correct measured values ​​of electrical currents by inverting the function f2.

[0044] The following section will describe in detail the process based on… Fig. The two explained methods are described. Generally, a high-voltage electrical system consists of K currents i1, i2, ..., i K considered, whereby in the example of the Fig. 2 K = 5 applies. For every current i k (k = 1, 2, ..., K) a measured value of i is obtained. k provided according to the MD measurement data sets. This measured value is subsequently referred to as i k,m denoted. All K currents flow through node N of the Fig. 1, so that according to Kirchhoff's knot rule the sum ∑k=1Kik which takes on the value zero. It should be noted that this criterion may not apply in the case of measurement errors. To simplify the notation, i and i are now used. m the K-dimensional vectors (i1, i2, ..., i K ) T or (i 1,m , i 2,m , ..., i K,m ) T designated.

[0045] During the journey of the specified motor vehicle, the measured currents result {i m(t)} 1≤t≤T from the measurement data sets MD, where t corresponds to a respective operating time. Based on this data, the goal is now to create a correction map i. m ↦ i to determine. This mapping can then be used to correct measurement signals in real time. As described above, this correction mapping is defined by corresponding parameter values ​​pw' and pw2, which are transmitted to the specified vehicle VE. The problem with determining the correction mapping is that only the measured values ​​i m (t) are available, but not the ground truth i(t) for the vehicle without measurement error. Since this ground truth is unknown, the corresponding mapping cannot be determined a priori via supervised learning.

[0046] To solve the above problem, the ideal ground truth i is therefore first established. sestimated, as described below. This ground truth corresponds to the size of is out Fig. 2. Next, a classifier determines whether the specified vehicle has hardware defects or not, as explained further below. Finally, depending on the classification result, a routine is executed to train the desired mapping for correcting measurement errors.

[0047] The following section will first describe the estimation of the ideal ground truth i. s explained. For the ideal ground truth, it is assumed that there is a perfect vehicle, i.e., a vehicle with normal operating behavior and no hardware faults, whose current values ​​are recorded without measurement errors. The k-th current in such a perfect vehicle is subsequently referred to as i k,s denoted. Thus, the current vector i exists. s = (i 1,s , i 2,s , ..., i K,s) for a perfect vehicle. The vector i s can be used as the target value of i m be considered, i.e. it should apply i s = i m However, for a less-than-perfect vehicle, the following results: s ≠ i m , whereby this deviation may be caused by hardware errors or measurement errors, or a combination of both types of errors.

[0048] Hardware failures (e.g., a flat tire) typically cause the vehicle to exhibit abnormal dynamic operating behavior that deviates from the operating behavior of a perfectly functioning vehicle. This results in the current i s The current i changes. Nevertheless, Kirchhoff's junction rule remains valid, i.e., the sum of the measured currents remains zero, so that 1 T i = t T i s If 0 holds (1 = (1,1, ...,1) T ).

[0049] On the other hand, measurement errors (e.g., sensor offsets or sensor drifts) cause a second deviation between i m and i caused. Unlike hardware errors, this deviation usually leads to a change in the sum of the measured currents, so that generally 1 T i m ≠ 0 applies. Overall, the deviation of measured current values ​​from the current values ​​of a perfect motor vehicle can be described as follows: im−isTotal deviation=im−iMeasurement error+i−isHardware error

[0050] In the procedure described here, a first data-driven model is learned for the given vehicle VE. Similarly, corresponding second data-driven models were also learned for further vehicles VE' in the fleet under consideration. Thus, the fleet of N (imperfect) vehicles, including vehicle VE and the vehicles VE', is considered. The measured current of an (arbitrary) j-th vehicle (j = 1, ..., N) from the fleet is then used. imj This is referred to as the state-space model. For each vehicle in the fleet, a discrete state-space model is learned, corresponding to the first data-driven model for the given vehicle VE and the second data-driven models for the other vehicles VE'. State-space models and their learning are familiar to those skilled in the art. The state-space model used here is described in more detail in document [1].

[0051] As part of learning the state-space model, system matrices A are used. j , B j , C j and D j with state vector x, input vector u from the above input variables u1, u2 and u3 and output vector {imj(t)}1≤t≤Tj based on measured currents {imj(t)}1≤t≤Tj learned how this is also described in document [1]. In the next step, the flows {imj}1≤j≤N estimated, which would be measured over the fleet of motor vehicles if all motor vehicles operated under the same operating conditions (see also document [1]). In other words, the above model run is considered, which is in Fig. 2 is denoted by UD and contains corresponding values ​​of input vectors u at corresponding model operation times. For this model operation, corresponding current values ​​are determined using the trained state-space models. {isimj(t)}1≤j≤N Predicted. For a given model railway operating time, this prediction is described by the following equation: x(t+1)=Ajx(t)+Bjuud(t)isimj(t)=Cjx(t)+Djuud(t)

[0052] The following applies: x0 = x(0) (arbitrary initialization) and u(t) = u ud (t), t ∈ {1,2, ..., T ud} (i.e., t denotes a model-riding operating time).

[0053] Using the designation i s (t) For the current value of a perfect vehicle at time t during the model drive UD, the following assumption is used in the procedure described here: Assumption 1: For a sufficiently large N and at each time step t, i s (t) exactly by the mean of the simulated or predicted currents isimj(t) approximate the N vehicles in the fleet, i.e., the following applies: is(t)≈i˜s(t):=1N∑j=1Nisimj(t)

[0054] This assumption is based on observation 1N∑j=1Nisimj=1N∑j=1N(is+imj−ij+ij−is+isimj−imj) =is+1N∑j=1N(imj−ij)+1N∑j=1N(ij−is)+1N∑j=1N(isimj−imj) as well as the assumption that for a large N, the measurement errors, hardware errors, and simulation errors average out to zero across the fleet at a given time. Accordingly, based on the above assumption, the current mean vector is determined and processed in the subsequent steps. Similarly, the vector of simulated current values ​​for the corresponding vehicle is also determined. isimj(t) further processed.

[0055] The following section refers to the specified motor vehicle VE. Fig. 2. Reference is made, i.e., the index j now refers to the specified vehicle VE. The steps described below can be carried out in the same way for the other motor vehicles VE' in the fleet. As a rule, the procedure described here is also carried out for all motor vehicles in the fleet.

[0056] According to the calculations above, an estimate for imj and i s for each time step t ∈ {1,2, ..., T ud} of the model railway. In other words, it applies imj(t)≈isimj(t) and i s (t) ≈ ĩ s (t). The problem considered here would thus be almost solved for the given motor vehicle if the occurrence of hardware failures could be ruled out. In this case, i j (t) = i s (t) ≈ i s (t) apply, which then together with isimj(t) can be used to create the desired image imj↦ij to learn (i.e., the mapping can be done with the tuples) {(isimj(t),i˜s(t))}1≤t≤Tud (to be learned).

[0057] To reduce the complexity of the procedure, a classification corresponding to step S2 of the process is used in the embodiment described here. Fig. 2 corresponds. This classification can be used to predict whether hardware errors will occur in the given vehicle. If this is not the case, the above figure can be easily replaced with the values. {(isimj(t),i˜s(t))}1≤t≤Tud The learning process is as described in the paragraph above. However, if hardware errors are present, a more complex algorithm is used to determine measurement errors, as described in more detail below.

[0058] The classification performed will first be explained below. The classification is based on K + 1 characteristics f1, f2, ..., f1. K+1It is assumed that the absolute deviation between imj and i s The average error increases when not only measurement errors but also significant hardware errors occur in the specified vehicle. Therefore, by means of isimj(t) and i s (t) for the given motor vehicle and for each current k the average squared deviation between ik,mj and i k,s determined. In other words, the k-th feature corresponding to a respective first feature in the sense of the patent claims is determined for the j-th motor vehicle (i.e., the specified motor vehicle) as follows: fkj=1Tud∑t=1Tud(ik,simj(t)−i˜k,s(t))2≈E((ik,mj−ik,s)2)

[0059] The following applies: 1 ≤ k ≤ K. The quantity E corresponds to the expected value.

[0060] Furthermore, the (k + 1)th feature corresponding to a second feature as defined in the claims is defined as a measure of the significance of measurement errors in the specified motor vehicle. Since measurement errors correspond to a current vector imj i m For a system whose entries are highly unlikely to sum to zero, this characteristic can be expressed as the root mean square of 1Timj to be estimated, i.e., the following applies: fK+1j=1Tj∑t=1Tj(1Timj(t))2)≈E((1Timj)2)

[0061] For this purpose, the original measured currents are used. {imj(t)}1≤t≤Tj required, i.e., the measured currents from the MD measurement data sets according to Fig. 2.

[0062] In the following, the first features and the second feature are represented by the vector fj=(f1j,f2j,...,fK+1j)T described. Based on this vector f jA corresponding classification is performed. In one variant, a simple threshold classifier is used. With this classifier, the decision as to whether hardware errors are present is made using the following formula: ηj:=∑i=1Kfij−1KfK+1j>δh

[0063] In other words, according to equation (7) above, a hardware fault exists if the size η j the appropriately chosen threshold δ h exceeds. For an intuitive explanation, the above criterion can be rewritten as follows: ηj≈∑k=1KE((ik,mj−ik,s)2)−1KE((1Timj)2)=E(∑k=1K(ik,mj−ik,s)2−(1TimjK)2) =E(‖imj−is‖22−‖imj−∏H(imj)‖22)>δh

[0064] The above formula (5) was substituted into the criterion according to equation (7). The quantity Π ℌ is the projection operator onto the hypersurface ℌ, which is given by ℌ = {x ∈ ℝ K |1 T x = 0}. Since the current i sIf Kirchhoff's knot rule is satisfied, then, due to the Pythagorean theorem, the following relationship holds: ‖imj−is‖22−‖imj−∏H(imj)‖22=‖∏H(imj)−is‖22

[0065] The above relationship is shown again in the diagram of the Fig. 3 for K = 2, i.e., two electric currents i1 and i2, are represented. The hyperplane ℌ is in Fig. 3 is represented as a dashed line H. According to Kirchhoff's node rule, this line describes the relationship i1 + i2 = 0. The point i s is the expected current (i.e., the ground truth) which, according to Kirchhoff's node rule, lies in the hyperplane ℌ. Hardware errors move the point i. s to point i j the hyperplane ℌ, as indicated by the arrow HE in Fig. 3 is indicated. In contrast, measurement errors move point i. j out of the hyperplane ℌ to the point imj, we through the arrow ME into Fig. 3 is indicated. The projection of the point imj returning to the hyperplane ℌ leads to the point ∏H(imj) In equation (8) above, the Pythagorean theorem is applied to a triangle formed by the points i s , imj and ∏H(imj) This triangle has a right angle at point [point missing]. ∏H(imj).

[0066] Knowing that i j Kirchhoff's knot rule is fulfilled (i.e., i) j ∈ ℌ), the vector ∏H(imj) as the best possible approximation for i j It can be considered in the absence of further information. Thus, the above criterion is an approximation of the mean squared distance. ‖imj−is‖22, i.e. an approximation of the mean squared deviation caused by hardware errors (see also equation (1)).

[0067] The preceding section described the classification for the presence of hardware errors. Similarly, a corresponding threshold can be used to classify whether a measurement error is present. For this purpose, a threshold can be defined for the second characteristic. ƒK+1j This characteristic can be used. If this characteristic exceeds the threshold, measurement errors have occurred. If this is not the case, no measurement errors have occurred.

[0068] Instead of the threshold classifier described above, a modified embodiment can also use a decision tree classifier, which is known per se, for the given motor vehicle. This decision tree classifier is pre-defined based on suitable training data in the form of feature vectors f. jThe system learns from the associated information about which class the respective feature vector belongs to. Decision tree classifiers and corresponding learning methods for these classifiers are well-known and therefore will not be described in further detail.

[0069] The above classification results in the following classes or cases: - Class 1: Neither hardware faults nor measurement errors are detected. In this case, the vehicle is virtually perfect, so the procedure is aborted. - Class 2: Only hardware errors are detected. In this case, too, the procedure is terminated, since according to the invention only measurement errors are to be corrected. - Class 3: Only measurement errors are detected. In this case, a correction image can be easily generated. imj↦ij using the tuples [(isimj(t), is(t))]t to be learned. - Class 4: Both hardware and measurement errors are present, i.e., the deviation between imj and i s It consists of two components, as shown in equation (1) above. However, without further assumptions, this problem cannot be uniquely solved. To arrive at a suitable solution, assumptions are made that make it possible to find a solution. The relevant assumptions and the solution for grade 4 will be explained in more detail later.

[0070] The above classes 1 to 4 are in Fig. 2 are designated as CL1 to CL4. The following section first explains the error models considered when classes 3 and 4 are present. For clarity, the index j is omitted hereafter. Since hardware errors do not alter the validity of physical laws, i.e., in this case, Kirchhoff's current law, these errors can generally be described by a continuous mapping or function f. h : i m ↦ i are modeled for which the following holds: is∈H⇒fh(is)∈H

[0071] This function corresponds to the first function within the meaning of the patent claims or to function f1 from Fig. 2. In the embodiment described here, a linear model is used for f. h used, namely fh(is)=(1+dh)is+oh.

[0072] This involves d h ∈ ℝ a drift scalar and o h ∈ ℌ is an offset vector. The following can be derived: is,oh∈H⇒1Tis=1Toh=0⇒(1+dh)1Tis+1Toh=0⇒1Tfh(is)=0⇒fh(is)∈H.

[0073] Similarly, measurement errors can be represented as a linear mapping or function f. m : i ↦ i m can be modeled. However, this function is not subject to the restrictions according to equation (10) above, since the noisy vector i m It does not have to satisfy Kirchhoff's knot rule. The function f m corresponds to a second function within the meaning of the patent claims or to function f2 from Fig. 2. The function f m In the embodiment described here, it is modeled as the following linear function: fm(i)=(I+Dm)i+om

[0074] Here, I is the identity matrix, D m ∈ ℝ K×K denotes a diagonal drift matrix and o m ∈ ℝ K is an offset vector. The aim of the embodiment described here is to determine the parameters D m ∈ ℝ K×K and om ∈ ℝ K to learn and the inverse function fm−1(im) to determine. As explained above, this inverse function can then be used to appropriately correct the corresponding measured current values. The parameters D m and o m are done using the tuples (i m , i s ) learned.

[0075] From the equations (11) and (12) above, the following relationship can be derived: im=fm∘fh(is)=(I+Dm)((1+dh)is+oh)+om=(I+Dm)(1+dh)is+(I+Dm)oh+om

[0076] In general, the parameters (especially D) can be m and o m ) not from the tuples in the form (i m , i s ) (and especially not from the tuples above (i sim , ĩ s )) be determined. To solve this problem, the above cases of Class 3 and Class 4 are distinguished in the embodiment described here.

[0077] If a classification or categorization into class 3 has been carried out, there are no hardware errors, so the above parameters d h and o h can be set to 0. From this, the relationship i can be derived from equation (13). m = (I + D m )i s + o m The parameters D m and o m In this case, the predicted current values ​​i can be easily determined by a linear regression. sim (t) on the average current values ​​i s (t) are determined. In other words, a total of K regressions are performed, where for each k a regression of the values ​​i sim,k (t) to the values ​​i m,k (t) is performed. The parameter D m and o m The formed vector corresponds to in Fig. 2 the parameter vector PV' with parameter values ​​pw'.

[0078] If the above class 4 applies, the problem cannot be uniquely solved without further knowledge. For example, if the vector (d h , o h , D m , o m If the above equation (13) is satisfied, then this also applies to the vector (d). h , o h + e, D m , o m - (I + D m )e), as long as e ∈ ℌ. The vector above (d h , o h , D m , o m ) corresponds to in Fig. 2 the parameter vector PV with corresponding parameter values ​​pw1 (i.e. values ​​of d h and o h ) and parameter values ​​pw2 (i.e., values ​​of D) m and o m To arrive at a clear solution in the event of classification in class 4, the following assumption is used.

[0079] Assumption 2: The vector x, which includes all parameters (d h , o h , D m , o m) contains a sparsely populated vector, i.e., a vector containing only a few parameter values ​​that are not equal to 0.

[0080] This assumption is based on the understanding that it is highly unlikely that all, or even many, sources of error would be present simultaneously in a motor vehicle. The presence of all sources of error at the same time would mean that all K sensors used for current measurement would exhibit offset or drift errors, or both. Furthermore, K currents would invariably be affected by hardware failures.

[0081] The assumption that there are relatively few sources of error in the same vehicle at any given time is therefore plausible.

[0082] Using the above assumption, the parameter vector x, which corresponds to the parameter vector PV of the Fig. 2 corresponds to the following. First, the expressions (I + D) are determined. m )(1 + d h ) and (I + D m )Oh + o m through the tuples (i sim (t), ĩ s (t)). This estimation uses a linear regression of i. sim (t) on i s (t) is performed. In the following, the k-th diagonal element of D is calculated. m as d m,k and the k-th elements of o h or o m as o h,k or o m,k denoted. For each k ∈ {1,2, ..., K}, the following relationship then results: (1+dm,k)+(1+dh)≈ak(1+dm,k)oh,k+om,k≈bk

[0083] Here, a denotes k or b k Estimates for the slope or the y-intercept. These estimates are obtained through the regression of (i sim,k (1), ..., i sim,k (T ud )) T on (ĩ s,k (1), ..., ĩ s,k (T ud )) T obtained. Based on the definitions y = (a1, ..., a K ,f, b1, ..., b K ) T and x̃ = (d h, o h,1 ..., o h,K , d m,1 , ..., d m,K , o m,1 , ..., o m,K , 1) T ∈ ℝ 3K+2 Each of the estimated terms can be expressed in a quadratic form x̃ T Qx̃ of x̃ can be brought into the following form, for example: (1+dm,k)+(1+dh)=(x˜Te3K+2)2+x˜Te3K+2eK+1+kTx˜+x˜Te3K+2e1Tx˜+x˜TeK+1+k e1Tx˜=x˜T(e3K+2e3K+2T+e3K+2eK+1+kT+e3K+2e1T+eK+1+kTe1T)x˜:=x˜TQkx˜,es

[0084] The notation e 3K+2 the canonical vector with respect to 3K+2. Similarly, one obtains (1 + d m,k )O h,k + o m,k = x̃ T Q K+k x̃ and can therefore write equation (14) as follows: ∀i∈{1,2,…,2K}x˜TQix˜≈yTei.

[0085] Finally, the parameter vector x (or x̃) is obtained by solving the following optimization problem: minx‖x˜‖1 subject to the condition ∑i=12K(x˜TQIx˜−yTei)2≤ε,x˜Te3K+2=1, x˜T(∑i=1Ke1+i)=0

[0086] Here, ∈ denotes an error threshold. The restriction x̃ T e 3K+2 The condition = 1 ensures that the last element of x̃ takes the value 1. x˜T(∑i=1Ke1+i)=0 ensures that the hardware offsets o h,1 , ..., o h,K Summing to zero satisfies equation (10). The optimization problem from equation (17) represents a minimization of the sparsity of the vector x. In other words, this optimization aims to ensure that the vector x has the fewest possible non-zero parameter values. This optimization problem can be solved, for example, using the "Quadratic Basic Pursuit" algorithm from publication [2]. This algorithm transforms the above problem into a convex problem, allowing it to be solved using a computer.

[0087] The overall result is thus the corresponding parameter vectors for the classifications in class 3 and in class 4, which are then transmitted from the backend server SE to the vehicle VE under consideration and are used there to correct current measurements by inverting the corresponding function.

[0088] In a modified embodiment, the above classification is omitted. Instead, the algorithm for class 4, which uses the parameter vector (d), is used. h , o h , D m , o m ) determined, always executed. In other words, the above optimization based on equation (17) can always be performed, regardless of whether hardware errors are present or not. Without hardware errors, the optimization then yields the result that the quantities d h and o hEssentially, it takes on the value zero. Nevertheless, the result is improved if the above classification is performed.

[0089] The inventors tested the method described above using suitable measurement data. The results showed that the method reliably determines the corresponding parameter vectors and is therefore very well suited for correcting measurement errors of corresponding current values ​​in a motor vehicle.

[0090] The embodiments of the invention described above offer a number of advantages. In particular, a fleet-based approach allows the learned data-driven models of several motor vehicles to be compared with one another in order to identify irregularities in the measurement of current values ​​in a high-voltage electrical system. Furthermore, suitable parameter values ​​are determined, which can then be used in the respective motor vehicle to correct measurement errors for the corresponding current values. Reference symbol list BN high-voltage electrical system HVB high-voltage battery EM electric machine HE heating CO air conditioning IN converter i1, i2, ..., i5 electric currents MO, MO' data-driven models VE, VE' motor vehicles LE Learning Methods MD measurement data sets MUD model vehicle measurement series UD Model Railway u1, u2, u3 Input variables is average electricity values CL1, CL2, CL3, CL4 classes f1, f2 functions PV, PV' parameter vectors pw1, pw2, pw' parameter values S1, S2 steps SE Backend Server H Hyperplane i s , i j i j m , current vectors HE, ME Shifts of current vectors Bibliography [1] Pfeiffer, J.; Wolf, P.; Pereira, R. A Fleet-Based Machine Learning Approach for Automatic Detection of Deviations between Measurements and Reality. 2019 IEEE Intelligent Vehicles Symposium (IV); IEEE Paris, France, 2019; Pages 2086-2092. Loi: 10.1109 / IVS.2019.8813858. [2] Ohlsson, H.; Yang, A.Y.; Dong, R.; Verhaegen, M.; Sastry, SS Quadratic Basis Pursuit. arXiv:1301.7002 [cs, math, stat] 2013.arXiv: 1301.7002. [3] DE 102 35 525 A1 [4] CN 109 784 560 A

Claims

[1] Method for computer-aided evaluation of measurements of electric currents (i1, i2, ..., i5) in a high-voltage electrical system (ES) of a given electrically powered motor vehicle (VV), wherein the high-voltage electrical system (ES) comprises a plurality of electrical components (HV, EM, HE, CO, IN) including an electric machine (EM) for driving the given motor vehicle (VV) and an electrical energy storage device (HV) for supplying electrical energy to the electric machine (EM), and each electric current (i1, i2, ..., i5) flows through the same node (N) in the high-voltage electrical system (ES), wherein: a) in the specified motor vehicle (VE), a first data-driven model (MO) is learned using a machine learning method (LE), and the learned first data-driven model (MO) is transmitted to a backend server (SE) by the specified motor vehicle (VE), wherein the first data-driven model (MO) is learned based on a plurality of measurement data sets (MD), each of which was measured during the operation of the specified motor vehicle (VE) at different times of operation, wherein each measurement data set (MD) for a respective time of operation comprises measured values ​​of input variables (u1, u2, u3) and measured values ​​of the electric currents (i1, i2, ..., i5), wherein the learned first data-driven model (MO) can predict measured values ​​of the electric currents (i1, i2, ..., i5) based on measured values ​​of the input variables (u1, u2, u3); b) in the backend server (SE) the learned first data-driven model (MO) and a plurality of further learned second data-driven models (MO'), each of which was learned in a different motor vehicle (VE') of the same type as the specified motor vehicle (VE) using the same machine learning method (LE) as the first data-driven model (MO) based on a plurality of measurement data sets measured in the respective other motor vehicle (VE') for the same input variables (u1, u2, u3) and the same electrical currents (i1, i2, ..., i5) as the first data-driven model (MO), are processed by calculating current averages (is) for the electrical currents (i1, i2, ...) for a model driving measurement series (MUD), which contains measurement values ​​for the input variables used in learning the first data-driven model (u1, u2, u3) for a plurality of model driving operating times within a specified model driving., i5) are determined for each model operation time, wherein a respective current mean value (is) for a respective electric current (i1, i2, ..., i5) and for a respective model operation time is the mean value of the measured values ​​of the respective electric current (i1, i2, ..., i5), which are predicted by the first learned data-driven model (MO) and the multitude of second learned data-driven models (MO') based on the measured values ​​of the input variables (u1, u2, u3) for the respective model operation time;. c) at least in the case that, depending on the measurement data sets (MD) measured in the specified motor vehicle (VE), abnormal operating behavior of the specified motor vehicle (VE) and the presence of measurement errors in the measurement of the electric currents (i1, i2, ..., i5) in the specified motor vehicle are detected, a parameter vector (PV) is determined from parameter values ​​(pw1) for parameters of a first function (f1) and from parameter values ​​(pw2) for parameters of a second function (f2) using the measured values ​​of the respective electric currents (i1, i2, ..., i5) predicted by the first learned data-driven model (MO) and the average current values ​​(is) for the respective electric currents (i1, i2, ..., i5), wherein the first function (f1) calculates current values ​​of the electric currents (i1, i2, ...) under normal operating behavior of the specified motor vehicle (VE)., i5) in the case of abnormal operating behavior of the specified motor vehicle (VE) and wherein the second function (f2) maps current values ​​of the electric currents (i1, i2, ..., i5) in the absence of measurement errors to current values ​​of the electric currents (i1, i2, ..., i5) in the presence of measurement errors in the specified motor vehicle (VE), wherein the parameter values ​​(pw1, pw2) are determined by means of an optimization with the optimization goal of a minimum number of parameter values ​​(pw1, pw2) not equal to zero in the parameter vector (PV), wherein the parameter values ​​(pw2) of the parameters of the second function (f2) are transmitted to the specified motor vehicle (VE) for the correction of future measured values ​​of at least a part of the electric currents (i1, i2, ..., i5). [2] Method according to claim 1, characterized by, that depending on the measurement data sets (MD) measured in the specified motor vehicle (VE), a classification is carried out into the following classes before step c) is performed: - a first class (CL1) according to which there is neither abnormal operating behavior of the specified motor vehicle (VE) nor measurement errors for the measured values ​​of the electrical currents (i1, i2, ..., i5) in the specified motor vehicle (VE); - a second class (CL2) according to which abnormal operating behavior of the specified motor vehicle (VE) is present, but no measurement errors are present for the measured values ​​of the electrical currents (i1, i2, ..., i5) in the specified motor vehicle (VE); - a third class (CL3) according to which there is no abnormal operating behavior of the specified motor vehicle (VE), but measurement errors are present for the measured values ​​of the electrical currents (i1, i2, ..., i5) in the specified motor vehicle (VE); - a fourth class (CL4), according to which abnormal operating behavior of the specified motor vehicle (VE) is present and measurement errors are present for the measured values ​​of the electrical currents (i1, i2, ..., i5) in the specified motor vehicle (VE); the procedure is terminated if the first class is present and if the second class is present. [3] Method according to claim 2, characterized by, that the classification is based on first features and a second feature, wherein each first feature is assigned to an electric current (i1, i2, ..., i5) and is a measure of the absolute deviation of the measured values ​​of the assigned electric current (i1, i2, ..., i5) predicted by the first learned data-driven model (MO) from the mean current values ​​(is) for the model operation times, and wherein the second feature is a measure of the magnitude of the sums of the measured values ​​of the electric currents (i1, i2, ..., i5) determined for the respective measurement data sets (MD) for the operation time of the corresponding measurement data set (MD). [4] Method according to claim 3, characterized by, that the classification is carried out via a threshold classifier, wherein abnormal operating behavior of the specified motor vehicle (VE) is present when a first value, which becomes larger with increasing size of the first features and smaller with increasing size of the second feature, exceeds a first threshold, and wherein measurement errors for the measured values ​​of the electrical currents (i1, i2, ..., i5) in the specified motor vehicle (VE) are present when a second value, which becomes larger with increasing size of the second feature, exceeds a second threshold. [5] Method according to one of claim 3, characterized by , that the classification is carried out via a machine-learned decision tree classifier, which has been comprehensively trained using training datasets, including first features and a second feature as well as an associated class from the first to fourth class. [6] Method according to any one of claims 2 to 5, characterized by , that in the case of the third class (CL3), instead of the parameter vector (PV) of step c), a parameter vector (PV') reduced compared to the parameter vector (PV) of step c) is determined from parameter values ​​(pw') for the parameters of the second function (f2), wherein the parameter values ​​(pw') are determined by a regression of the measured values ​​of the electric currents (i1, i2, ..., i5) predicted by the first data-driven model (MO) on the current mean values ​​(is) for the model driving operating times, wherein the parameter values ​​(pw') of the parameters of the second function (f2) are transmitted to the specified motor vehicle (VE) to correct future measured values ​​of at least a part of the electric currents (i1, i2, ..., i5). [7] Method according to any one of the preceding claims, characterized by, that the input variables (u1, u2, u3) include an acceleration of the specified motor vehicle (VE) and a speed of the specified motor vehicle (VE), wherein the input variables (u1, u2, u3) preferably also include the inclination of the specified motor vehicle (VE) relative to the horizontal and / or the gear engaged in the specified motor vehicle (VE) and / or a power request and / or allocation to the electric machine (EM) and / or a speed of the electric machine (EM). [8] Method according to any one of the preceding claims, characterized by , that the high-voltage electrical system (BN) includes, in addition to the electric machine (EM) and the electric energy storage (HVB), a heater (HE) and / or an air conditioner (CO) and / or an inverter (IN) for voltage conversion to a low-voltage electrical system. [9] Method according to any one of the preceding claims, characterized by, that the first data-driven model (MO) and the majority of second data-driven models (MO') each represent a neural network structure, in particular an LSTM network structure. [10] Method according to any one of claims 1 to 8, characterized by , that the first data-driven model (MO) and the majority of second data-driven models (MO') each represent a state-space model. [11] System for computer-aided evaluation of measurements of electric currents (i1, i2, ..., i5) in a high-voltage electrical system (ES) of a given electrically powered motor vehicle (VV), wherein the high-voltage electrical system (ES) includes a plurality of electrical components (HV, EM, HE, CO, IN) comprising an electric machine (EM) for driving the given motor vehicle (VV) and an electrical energy storage device (HV) for supplying electrical energy to the electric machine (EM), and each electric current (i1, i2, ..., i5) flows through the same node (N) in the high-voltage electrical system (ES), wherein the system is set up to carry out a method in which: a) in the specified motor vehicle (VE), a first data-driven model (MO) is learned using a machine learning method (LE), and the learned first data-driven model (MO) is transmitted to a backend server (SE) by the specified motor vehicle (VE), wherein the first data-driven model (MO) is learned based on a plurality of measurement data sets (MD), each of which was measured during the operation of the specified motor vehicle (VE) at different times of operation, wherein each measurement data set (MD) for a respective time of operation comprises measured values ​​of input variables (u1, u2, u3) and measured values ​​of the electric currents (i1, i2, ..., i5), wherein the learned first data-driven model (MO) can predict measured values ​​of the electric currents (i1, i2, ..., i5) based on measured values ​​of the input variables (u1, u2, u3); b) in the backend server (SE) the learned first data-driven model (MO) and a plurality of further learned second data-driven models (MO'), each of which was learned in a different motor vehicle (VE') of the same type as the specified motor vehicle (VE) using the same machine learning method (LE) as the first data-driven model (MO) based on a plurality of measurement data sets measured in the respective other motor vehicle (VE') for the same input variables (u1, u2, u3) and the same electrical currents (i1, i2, ..., i5) as the first data-driven model (MO), are processed by calculating current averages (is) for the electrical currents (i1, i2, ...) for a model driving measurement series (MUD), which contains measurement values ​​for the input variables used in learning the first data-driven model (u1, u2, u3) for a plurality of model driving operating times within a specified model driving., i5) are determined for each model operation time, wherein a respective current mean value (is) for a respective electric current (i1, i2, ..., i5) and for a respective model operation time is the mean value of the measured values ​​of the respective electric current (i1, i2, ..., i5), which are predicted by the first learned data-driven model (MO) and the multitude of second learned data-driven models (MO') based on the measured values ​​of the input variables (u1, u2, u3) for the respective model operation time;. c) at least in the case that, depending on the measurement data sets (MD) measured in the specified motor vehicle (VE), abnormal operating behavior of the specified motor vehicle (VE) and the presence of measurement errors in the measurement of the electric currents (i1, i2, ..., i5) in the specified motor vehicle are detected, a parameter vector (PV) is determined from parameter values ​​(pw1) for parameters of a first function (f1) and from parameter values ​​(pw2) for parameters of a second function (f2) using the measured values ​​of the respective electric currents (i1, i2, ..., i5) predicted by the first learned data-driven model (MO) and the average current values ​​(is) for the respective electric currents (i1, i2, ..., i5), wherein the first function (f1) calculates current values ​​of the electric currents (i1, i2, ...) under normal operating behavior of the specified motor vehicle (VE)., i5) in the case of abnormal operating behavior of the specified motor vehicle (VE) and wherein the second function (f2) maps current values ​​of the electric currents (i1, i2, ..., i5) in the absence of measurement errors to current values ​​of the electric currents (i1, i2, ..., i5) in the presence of measurement errors in the specified motor vehicle (VE), wherein the parameter values ​​(pw1, pw2) are determined by means of an optimization with the optimization goal of a minimum number of parameter values ​​(pw1, pw2) not equal to zero in the parameter vector (PV), wherein the parameter values ​​(pw2) of the parameters of the second function (f2) are transmitted to the specified motor vehicle (VE) for the correction of future measured values ​​of at least a part of the electric currents (i1, i2, ..., i5). [12] System according to claim 11, characterized by that the system is designed to carry out a method according to one of claims 2 to 10. [13] Backend server configured as a backend server (SE) for a method according to any one of claims 1 to 10, wherein the backend server (SE) is configured such that - in the backend server (SE) a learned first data-driven model (MO) for a given vehicle (VE) as well as a plurality of further learned second data-driven models (MO'), each of which is in a different vehicle (VE') of the same type as the given vehicle (VE) with the same machine learning method (LE) as the first data-driven model (MO) based on a multitude of measurement data sets measured in the respective other vehicle (VE') for the same input variables (u1, u2, u3) and the same electrical currents (i1, i2, ..., i5) how the first data-driven model (MO) was learned, are processed by determining current mean values ​​(is) for the electric currents (i1, i2, ..., i5) for each model-driven operating time within a given model drive for a model drive measurement series (MUD) which contains measured values ​​for the input variables used in learning the first data-driven model (u1, u2, u3) for a multitude of model-driven operating times within a given model drive, wherein a respective current mean value (is) for a respective electric current (i1, i2, ..., i5) and for a respective model-driven operating time is the mean value of the measured values ​​of the respective electric current (i1, i2, ..., i5) which are predicted by the first learned data-driven model (MO) and the multitude of second learned data-driven models (MO') based on the measured values ​​of the input variables (u1, u2, u3) for the respective model-driven operating time;. - at least in the case that, depending on the measurement data sets (MD) measured in the specified motor vehicle (VE), an abnormal operating behavior of the specified motor vehicle (VE) and the presence of measurement errors in the measurement of the electric currents (i1, i2, ..., i5) in the specified motor vehicle are determined, a parameter vector (PV) is determined from parameter values ​​(pw1) for parameters of a first function (f1) and from parameter values ​​(pw2) for parameters of a second function (f2) using the measured values ​​of the respective electric currents (i1, i2, ..., i5) predicted by the first learned data-driven model (MO) and the average current values ​​(is) for the respective electric currents (i1, i2, ..., i5), wherein the first function (f1) calculates current values ​​of the electric currents (i1, i2, ...) under normal operating behavior of the specified motor vehicle (VE)., i5) in the case of abnormal operating behavior of the specified motor vehicle (VE) and wherein the second function (f2) maps current values ​​of the electric currents (i1, i2, ..., i5) in the absence of measurement errors to current values ​​of the electric currents (i1, i2, ..., i5) in the presence of measurement errors in the specified motor vehicle (VE), wherein the parameter values ​​(pw1, pw2) are determined by means of an optimization with the optimization goal of a minimum number of parameter values ​​(pw1, pw2) not equal to zero in the parameter vector (PV), wherein the parameter values ​​(pw2) of the parameters of the second function (f2) are transmitted to the specified motor vehicle (VE) for the correction of future measured values ​​of at least a part of the electric currents (i1, i2, ..., i5). [14] Electrically powered motor vehicle configured as a predefined motor vehicle (VV) for a method according to any one of claims 1 to 10, wherein the motor vehicle (VV) is designed such that - in the motor vehicle (VE), a first data-driven model (MO) is learned using a machine learning method (LE), and the learned first data-driven model (MO) is transmitted to a backend server (SE) by the specified motor vehicle (VE), wherein the first data-driven model (MO) is learned based on a multitude of measurement data sets (MD), each of which was measured during the operation of the specified motor vehicle (VE) at different operating times, wherein each measurement data set (MD) for a respective operating time comprises measured values ​​of input variables (u1, u2, u3) and measured values ​​of the electrical currents (i1, i2, ..., i5), wherein the learned first data-driven model (MO) can predict measured values ​​of the electrical currents (i1, i2, ..., i5) based on measured values ​​of the input variables (u1, u2, u3); - the motor vehicle receives parameter values ​​(pw2) from parameters of a second function (f2) and uses the received parameter values ​​(pw2) to correct measured values ​​of at least some of the electrical currents (i1, i2, ..., i5).

Images