DE102022115217A1 - Method and system for determining parameterization values for a physical model of a vehicle component - Google Patents

Abstract

Method and system (1) for determining parameterization values for model parameters of a physical replacement model (13) of a vehicle component (3), comprising a measuring device (2) and a parameterization device (8). The measuring device is set up to measure measuring points of the vehicle component (3) in order to obtain measured values and transmit them to the parameterization device (8). The parameterization device (8) comprises a computing unit (11), an electronic memory (12) and an input (9) and output interface (10) and is set up to determine parameterization values for model parameters of the physical replacement model (13) through a prediction error to determine a prediction model (14), to train the prediction model (14) depending on the prediction error, to evaluate the quality of the specific parameterization values based on at least one evaluation criterion and to carry out a transfer of the parameterization values. If the quality is insufficient, a selection strategy is determined for selecting a further measuring point to be measured using a selection model (15) and the vehicle component (3) is measured again and parameterization values are determined according to the method according to the invention.

Description

Technical area

The invention relates to the machine learning-based calibration of vehicle components using physical replacement models.

background

As the complexity of controlling technical systems increases, the calibration effort also increases. In the field of vehicle control units it can be seen that more and more functions are being implemented in vehicles in order to control them. Vehicle control units largely use assignments listed in tables, for example in the form of characteristic values, characteristic curves and maps, in order to map the internal calculation of the functions to be carried out. During the control process, the vehicle control units receive input variables, such as sensor data, by means of which output variables are generated based on the assignments mentioned in order to control subsystems and individual vehicle components. To ensure that the desired control occurs, the assignments listed in the table contain specific values or calibration values that are tailored to the vehicle component. These must be initially set up during manufacture or with any change to the vehicles, control devices or other requirements and transferred to a vehicle control unit. The determination of these calibration values is partly based on experience and technical know-how. However, most of the calibration values are measured in a complex manner at surveying facilities or in field tests. Each controllable component of a vehicle is assigned at least one vehicle control unit, which includes at least one calibration value for controlling that vehicle component. The technical setup of a vehicle drive train involves several thousand or even tens of thousands of calibration values, which must be remeasured with every change during the manufacturing process, but also in the rest of the life cycle. This creates an enormous amount of time as a cost driver. To address this situation, there are various approaches to automate the calibration of technical devices or control units and reduce the time required. An alternative way of determining the calibration values for assignment listed in a table is to use physical models, or alternative models, which represent the vehicle component to be calibrated.

State of the art

In DE 10 2017 201 632 A1 A system for predicting calibration values based on existing calibrations is proposed, with training data being received for the control of an internal combustion engine and a prediction model being automatically developed with these. The development of the prediction model includes selecting one of several stored machine learning applications based on the training data, which forms the basis for the prediction model. Following the development of the prediction model, the required calibration values are generated using the prediction model and transferred to the corresponding control unit so that it can control the internal combustion engine.

In DE 10 2019 008 400 A1 A method for calibrating a drive train for a vehicle is proposed, wherein the drive train, which can be an internal combustion engine, electrified or electrical design, is operated according to a flow chart and the calibration values are determined by means of reinforced learning.

Short presentation

The state of the art is characterized by the circumstances that in order to automate the calibration process, either historical training data must be used in order to initially develop a suitable prediction model, although validation using real measured values remains neglected, or that the technical device itself follows a conventional flow plan must be operated in order to generate calibration values based on this sequence using elements of machine learning. Physical substitute models for determining calibration values can provide a solution. Substitute models include one or more parameters that represent the physical properties and behavior of the modeled vehicle component. According to conventional standards, the adaptation of the replacement models by determining the parameterization values of the model parameters requires an analogous procedure for determining the calibration values themselves.

This is addressed by the invention proposed here, the task of which is to automate the determination of the parameterization values for a replacement model of a vehicle component, to reduce the operation of the vehicle component within a measuring device according to a conventional flow chart by using machine learning, and to reduce the result of the Evaluate parameterization. This object is achieved by a method according to claim 1 and a system according to claim 10, with further variants of the invention being formed by the dependent subclaims.

The proposed invention includes a system and a method for determining parameterization values for a physical replacement model of a vehicle component, wherein one or more measuring points from a conventional measurement flow chart for operating the vehicle component to be controlled by a vehicle control unit are first measured. The measured values generated in this way are transmitted to a parameterization device and initially used to parameterize a replacement model. The parameterization of the replacement model includes the determination of parameterization values of the model parameters. The replacement model newly parameterized in this way is in turn part of a prediction model, which trains a data-based method, preferably a machine learning method, in particular preferably a neural network, and predicts an error in calibration values, or also a prediction error, which is generated by the newly parameterized replacement model. The prediction error can be used as one of other possible evaluation criteria to evaluate the quality of the parameterization. If the quality of the parameterization is insufficient compared to one or more evaluation criteria, a selection model is activated that is set up to select a new measuring point from the measurement flow plan and taking the prediction error into account. The new measuring point is then measured so that, based on the measured values determined in this way, the replacement model is parameterized again, the prediction error is determined, the neural network is trained and the quality of the parameterization is evaluated. This process sequence can be repeated until the parameterization achieves sufficient quality according to one or more evaluation criteria.

By using a selection model that takes the current progress of the prediction model into account, individual operating points of the vehicle component are selected one after the other as measuring points, so that the entire measurement flow plan for operating the vehicle component does not have to be measured. The number of measuring points is therefore effectively reduced, which also reduces the time required for measuring. By updating the parameterization values of the model parameters of the replacement model after each new measurement, the determination of the parameterization values becomes increasingly precise in an advantageous manner over the course of the method. By evaluating the parameterization values determined in this way by determining a prediction error after each new measurement, the quality of the parameterization can advantageously be determined in each iteration of the method according to the invention, so that the process sequence can be ended when defined evaluation criteria are met. The prediction model can, for example, be trained using reinforcement learning on the behavior of the specific application and during the application of the method according to the invention, neither the presence of historical training data nor a prior adaptation to the selected technical device being required. The proposed invention therefore combines the advantages of existing parameterization methods using a prediction model and reinforced learning and expands it to include a reduction in measuring points and an evaluation of the calibration itself.

Brief description of the drawings

• 1 shows a block diagram that represents a system for determining parameterization values according to an exemplary embodiment for calibrating an electric drive.
• 2 shows a process flow diagram illustrating a method for determining parameterization values according to some embodiments of the invention.
• 3 shows the operating map and a torque characteristic of an electric machine at an intermediate state within the method according to the invention.
• 4A and 4B show graphic representations of the determination of the prediction error according to the method according to the invention, using the exemplary embodiment of the electrical machine.

Detailed description

Some embodiments of the systems and methods described here relate to determining parameterization values for an electric machine as a vehicle component, in particular for the control unit of an electric drive and preferably an asynchronous machine (ASM), which is arranged within a vehicle. However, the systems and methods may also be used to determine other types of parameterization values for surrogate models for determining calibration values for vehicle control units, vehicle components, or vehicles in general and are not intended to be limited by the embodiments described herein. All systems presented, part Systems and components can be implemented both as hardware and as software, include one or more modules and one or more transmission interfaces, which are not shown individually.

1 shows a system (1) for determining parameterization values for a physical replacement model (13) of a component of a vehicle, not shown, comprising a measuring device (2), which is essentially set up to generate measured values for parameterization values and a parameterization device (8) , which is essentially designed to determine corresponding parameterization values. The measuring device (2) is to be understood as the entirety of individual technical components and devices which are required to generate measured values for parameterization values of the model parameters of the physical replacement model (13) of the vehicle component (3). In general, the measuring device (2) comprises at least one vehicle control unit (6), one vehicle component (3) and one measuring device (7).

The vehicle component (3) is to be understood as that technical system or subsystem which or whose control is calibrated and can accordingly comprise an entire vehicle, an individual vehicle component (3) or a vehicle control unit (6) itself. The vehicle control unit (6) is to be understood as the technical system or subsystem that controls the vehicle component (3) by transmitting control commands. The measuring device (7) is to be understood as the technical system or subsystem which records the current operating state of the vehicle component (3), for example by recording measured values. In one embodiment, the measuring device (2) is a vehicle, the vehicle component (3) to be calibrated is a drive unit of the vehicle, the vehicle control unit (6) is a drive control device for the drive unit and the measuring device (7) is a computer comprising at least one data interface for data from the drive unit and/or the drive control unit. The embodiment described is neither intended to limit the system according to the invention nor to limit it to that embodiment. Rather, this is intended to reveal the functional structure and the qualitative, task-related relationships of the individual components of the measuring device (2), such as the vehicle control unit (6), vehicle component (3) and measuring device (7) to one another.

In a preferred embodiment of the system (1) according to the invention, the measuring device (2) is a test bench, in particular a drive train test bench for a vehicle, preferably a component test bench, wherein the vehicle component (3) has an electric drive, in particular an electric machine (EM) (4). and a frequency converter (5), the vehicle control unit (6), a drive control device which is set up to control the EM (4) and / or the frequency converter (5), and the measuring device (7) test bench measurement technology, which is at least set up to to measure the current strength, speed, torque and slip of the EM (4).

In some embodiments, the vehicle control unit (6) comprises one or more electrical and electronic components, such as its own electronic computing unit, such as a microprocessor or other electronic processors, non-volatile memories, input and output interfaces as well as control and data buses. Control data is arranged on the vehicle control unit (6) or on its electronic memories in order to control the vehicle component (3). The calibration values are arranged within the control data, which include, for example, tabular assignments in the form of characteristic values, characteristic curves or maps, but also parameters for arithmetic operations and other control functions, so that a change in the calibration values correspondingly causes a changed control behavior of the vehicle component (3). In a preferred embodiment, the control data includes at least one operating map for the EM (4), the individual entries of the map corresponding to the calibration values of the EM (4).

Referring to 1 The parameterization device (8) correspondingly comprises at least one input (9) and one output interface (10), which are set up to establish communication connections with other components and to transmit data. The parameterization device (8) further comprises at least one computing unit (11) and an electronic memory (12). The computing unit (11) can include an electronic processing device, such as a microprocessor or other types of processors. The electronic memory (12) is designed to store at least one software-implemented replacement model (13), one prediction model (14) and one selection model (15). The replacement model (13) is essentially designed to represent a parameterizable physical model of the vehicle component (3) to be calibrated and to generate model values. The prediction model (14) is essentially set up to implement a data-based method, preferably a method machine learning, in particular preferably to train a neural network, to determine parameterization values for the replacement model (13) and to generate a prediction error of the model prediction, or in other words of the replacement model (13) at the individual measurement points. The selection model (15) is essentially set up to select one or more measuring points to be measured in each iteration of the method according to the invention. The embodiments described are neither intended to limit the system (1) according to the invention nor to restrict it to that embodiment. Rather, this is intended to determine the functional structure and the qualitative, task-related relationships of the individual components of the parameterization device (8), such as input (9) and output interfaces (10), computing unit (11), memory (12), replacement model (13), prediction model ( 14) and selection model (15) are revealed to each other. In a preferred embodiment, the parameterization device (8) comprises a computer, with input (9) and output interfaces (10), computing unit (11) and memory (12) implemented in hardware, as well as a replacement model (13), prediction model (14) and selection model (15). are implemented in software.

2 shows the higher-level process flowchart, which represents the method according to the invention for determining parameterization values of a physical replacement model (13) for a vehicle component (3) according to some embodiments of the invention. In a first step S1, the selection strategy for selecting the measuring point to be measured is determined by the selection model (15). The selection model (15) includes algorithms for selecting the measuring point to be measured in the best possible way using a spatial coverage method or depending on a prediction error generated by the replacement model (13). However, in the first run of the process flow plan of the method according to the invention, the selection of one or more initial starting measuring points is required, from which the selection model (15) selects the further measuring point or points within one or more further iterations of the process flow. The initial measurement of measuring points includes the measurement of at least one operating point as a starting measuring point for the subsequent method steps. In the conventional calibration process, the general conditions for the vehicle component (3) to be calibrated are usually already known qualitatively and also quantitatively and/or are physically defined based on the technical design of the vehicle component (3). The speed-torque map of an EM (4), for example, has a characteristic appearance, with the detailed calibration values arising due to the technical design and the associated physical properties of the EM (4). In order to calibrate corresponding vehicle components (3) using conventional methods, grids within the maps, characteristic curves and characteristic values, or even the calibration values, are defined based on empirical values or purely statistical characteristics, whereby each individual operating point of the grid must be measured. However, for the initial measurement of measuring points, at least characteristic operating points, such as corner points of torque maps or physical limits of characteristic curves as well as maxima and minima of characteristic values, can be taken from this grid or the characteristic properties and defined as (starting) measuring points to be measured. The selection of the number and properties of the starting measuring points can have an influence on the parameterization method according to the invention in that the determination of the parameterization values can converge more precisely and/or faster with a higher number of starting measuring points. However, the intention of the method according to the invention is advantageously to reduce the number of measuring points to be measured, which creates an application-specific trade-off between the number of starting measuring points to be measured and reducing the total number of measuring points to be measured.

In an embodiment in which the vehicle component (3) to be calibrated comprises an EM (4), in the second step S2 of the method according to the invention, 6 starting measuring points are selected which correspond to the qualitatively characteristic corner points of the speed-torque map of an EM (4). Which quantitative values the associated operating points have is determined by the technical design and physical properties of the EM (4), can be specified by technical expertise or can be determined from an equivalent circuit diagram of the equivalent model (13). Alternatively, the initial selection of more or fewer starting measuring points than the number of 6 is possible and typically depends on the type and structure of the vehicle component (3) to be calibrated. For example, in the speed-torque maps of internal combustion engines, 3 or 4 starting measuring points are sufficient to cover the corner points of that speed-torque map. According to the application-specific dependence of the number of starting measuring points to be measured, at least one measuring point is selected in the second step S2 and within the first run of the method according to the invention using a selection model (15) or manually.

According to the application-specific dependency of the number of starting measuring points to be measured, in the third step S3 of the method according to the invention, at least one measuring point is measured using a measuring device (2) in order to obtain at least one measured calibration value, or also measured value. The measurement of measuring points includes the recording of measured values by the measuring device (7) via selected operating parameters of the vehicle component (3) to be calibrated at one or more defined operating points or measuring points of the vehicle component (3). The measurement of measuring points according to the third step S3 of the method according to the invention further includes the transmission of the generated measured values to the parameterization device (8). The transmission of measured values can take place via an electronic data connection, such as data buses, wireless communication connections or all other known transmission methods. The manner or design of the data transmission is not the subject of the invention; rather, the communication between the measuring device (2) and the parameterization device (8) is intended to be clarified. In general, it will be understood by the person skilled in the art that components of the measuring device (2), such as parts of the measuring device (7) are directly in electronic connection with components of the parameterization device (8), but can also be software and/or hardware implemented modules of the same computer-implemented system can.

Referring to 2 In a fourth step S4 of the method according to the invention, parameterization values of the model parameters for the replacement model (13) are determined. The replacement model (13) comprises a physical model of the vehicle component (3) to be calibrated, which is at least set up to generate the required model values. In a preferred embodiment, the equivalent model (13) comprises an equivalent circuit diagram of the EM (4), wherein the model parameters can each include one or more electrical resistances, capacitances and inductances. According to the invention, the parameterization values of the model parameters are determined depending on the measured values that have already been measured. In one embodiment, a measured value for an EM (4) comprises a measured operating point, or in other words a measuring point, each consisting of a value for the slip, a current amplitude and a speed of the EM (4). The entirety of the calibration values therefore comprises a large number of measuring points for an operating map of the EM (4), with slip frequencies associated with them preferably being plotted at reference points specified over a grid, comprising a current amplitude and a speed.

3 shows an example of the grid for a slip characteristic map (16) of an EM (4) (top) and the associated slip characteristic curve (17) (bottom) for a selected support point (19) of the grid, with individual steps of the method according to the invention for a preferred embodiment are shown. The grid of the slip map (16) comprises a large number of support points (18), consisting of a current amplitude (I) and a speed (n), whereby the individual support points (18) can be statistically distributed and/or predetermined by specialist knowledge. Empty support points (18) (white) represent measurement points that have not yet been measured and filled support points (19) (black) represent measurement points that have already been measured. In one embodiment of the method according to the invention, the corner points of the slip map (16) of the EM (4) are used as starting measurement points. measured, which correspond to the filled support points (19). During the measurement of the measuring points according to the third step S3 of the method according to the invention, the slip characteristic curve (17) is measured at the corresponding support point (19) in order to determine the point (20) of the maximum torque (M) as a function of the slip frequency (f). . A measured measuring point, comprising a support point (19), consisting of a current amplitude (I) and a speed (n), as well as the associated slip frequency (f) for a determined maximum torque (M), corresponds to a measured value and thus a calibration value for the Slip map (16) and for the EM (4) or the vehicle component to be calibrated (3) according to a calibration value from conventional calibration methods.

In a preferred embodiment, the measurement and thus the determination of the point (20) of the maximum torque (M) of a support point (19), consisting of a current amplitude (I) and a speed (n), is carried out using a mathematical optimization method, in particular preferably using " Golden Section Search (GSS), which is known to those skilled in the art as a method for determining an extreme point within a defined interval of a given function. In contrast to the conventional approach, in which the slip characteristic curve (17) is also measured over a fixed grid of support points, using the GSS method can advantageously determine a maximum torque (M) faster than with other optimization or measurement methods be, in particular because the interval for the slip characteristic (f) is already specified by the replacement model (13) and / or specialist knowledge about the physical properties of the EM (4) to be calibrated. The fourth step S4 of determining the parametrics The model parameters for the replacement model (13) can be determined automatically and within the prediction model (14), the measured values being transmitted from the measuring device (7) to the parameterization device (8), which are recorded by an input interface (9) and sent to the electronic memory (12), in particular to the replacement model (13) and / or to the prediction model (14).

According to the invention, parameterization values of the model parameters of the replacement model (13) are determined by the prediction model (14) using measured values generated in this way. In some embodiments, the prediction model (14) includes elements of machine learning, where the machine learning can be carried out using various types of methods, where the prediction model (14) includes neural networks, logic programming, support vector machines, clustering, Bayesian networks, and others , as well as different embodiments thereof. The detailed structure of the prediction model (14) should not be limited by the listed embodiments. Rather, it should be indicated that the prediction model (14) carries out mathematical operations and is at least set up to generate the parameterization values of the model parameters, taking into account a physical replacement model (13) and already generated measured values. The prediction model (14) itself is not the subject of the invention and can accordingly take on an application-specific design.

Referring to 3 According to the invention, model values are formed by the replacement model (13) for each predetermined measurement point, for example according to an existing grid or a measurement flow chart. The result of the model value formation is, for example, an assignment of calibration values to support points (18), (19). In a preferred embodiment, the result of the model value formation includes an assignment of predicted maximum torques (M) to the individual support points (18), (19), consisting of a current amplitude (I) and a speed (n), according to a predetermined slip grid -map of an EM (4). Alternatively, depending on the technical design of the vehicle component (3) to be calibrated, further assignments are possible in the form of characteristic values, characteristic curves or characteristic maps. In some exemplary embodiments, where the vehicle component (3) to be calibrated is a fuel injection system of an internal combustion engine, the result of the prediction includes, for example, a table-listed assignment of predicted fuel injection quantities to each of the individual support points (18), (19), consisting of a torque (M) and a speed (n), a speed-torque map of the internal combustion engine. Alternatively, any other technical system and subsystem of a vehicle or other technical device can be measured in the manner according to the invention and model values can be generated using a replacement model (13). In general, the method according to the invention measures individual measuring points from a predetermined grid, with the prediction model (14) determining parameterization values using the measured values and taking into account the replacement model (13).

According to 2 In a fifth step S5 of the method according to the invention, the prediction error is determined by the prediction model (14) in order to evaluate the accuracy of the calibration values generated by the replacement model (13) and thus the quality of the specific parameterization values. In a preferred embodiment, the determination of the prediction error includes forming the difference between the measured values associated with the individual measuring points and the model values generated by the replacement model (13) parameterized in accordance with the parameterization values. The amount or the square of the difference between the measurement and the model can be used as a measure of the prediction error. According to the invention, the replacement model (13) advantageously generates model values for the already measured measuring points of a previous iteration, so that the difference between the measured value and the model value can also be formed at these points in the current iteration. This advantageously results in the aspect that the more measuring points are measured, the more precisely the prediction error can be formed over the entire grid of the vehicle component (3) to be calibrated. Referring to 3 In one embodiment, in each iteration of the method according to the invention, prediction errors are determined for all measuring points (20) of the slip characteristic curve (17) of a measuring point (18, 19) of the slip characteristic map (16) to be calibrated and result in an overall prediction error for the current measuring point (18, 19) merged. In an advantageous embodiment, the squared errors between the measured value and the model value of the equivalent model (13) of the torque (M) are determined and summed up at each measured slip frequency.

In a preferred embodiment, the prediction errors are interpolated within the predetermined grid of measurement points, so that the prediction error according to 4A can be represented, this being plotted qualitatively and by way of example via the speed-current amplitude map in the third dimension and, where dark Areas of quality represent a high model error and bright areas represent a low model error. In the illustrated exemplary embodiment for an EM (4), the areas around the corner points of the already measured measuring points, including the support points (19), (21), are subject to a small prediction error, since their measured values are used when parameterizing the prediction model (14). underlying replacement model (13) were used according to the fourth step S4 of the method according to the invention. Areas with high interpolated prediction errors refer to the areas in which no measured values are yet available, especially when there are a small number of measured measuring points.

Referring to 2 In a sixth step S6 of the method according to the invention, the training of the prediction model (14) takes place, with the aim of selecting the measuring point to be measured in the next iteration of the method according to the invention in the best possible way and thus optimizing the determination of the parameterization values of the model parameters of the replacement model (13). In some embodiments, the prediction model (14) includes one or more neural networks. In an advantageous embodiment, the prediction model (14) comprises one or more elements of machine learning, preferably elements of reinforcement learning, or in other words “reinforcement learning” (RL). Reinforcement learning is known to those skilled in the art as a method of machine learning, whereby the neural network learns a strategy independently and without the use of training data from an existing process flow in order to maximize at least one or more rewards. The learning of the neural network refers in particular to changing or parameterizing the individual weights of the network so that changed output variables are generated based on the same input variables. The structure and functionality of the neural network used in the individual case are not the subject of the invention, especially since this design is purely mathematical in nature. Rather, it should be made clear that the neural network and thus the prediction model (14) is generally set up to use the prediction error as a reward criterion in order to learn to select the new measuring point in such a way that its measurement is based on it Newly parameterized replacement model (13) based on information generates the smallest possible prediction error.

In an advantageous embodiment, the determination of the updated parameterization values according to the fourth step S4, the determination of the prediction error according to the fifth step S5 and the training of the prediction model (14) according to step S6 take place simultaneously, by training and using the one or more neural networks of the prediction model (14).

Referring to 2 In a seventh step S7, the quality of the parameterization values of the model parameters of the replacement model (13) determined by the prediction model (14) is evaluated based on at least one evaluation criterion, and the specific parameterization values are transferred in a further step S8, depending on the evaluation. The step S7 of evaluating the quality of the specific parameterization values represents an abort criterion within the higher-level process flow according to the method according to the invention, the process flow jumping to step S8 if fulfilled and to the first step S1 if not fulfilled, the first step S1 being a further iteration loop of the invention Procedure initiates in order to further improve the result of the prediction model (14) and thus the quality of the specific parameterization values. In some embodiments of the method according to the invention, the prediction error determined is used as an evaluation criterion as a measure of the achieved quality of the specific parameterization values. For this purpose, an overall prediction error can be determined from the respective prediction errors of the individual support points (18), (19), for example by means of arithmetic or weighted averaging, median or other mathematical methods that are suitable for determining an overall prediction error from the individual prediction errors. In one embodiment, the overall prediction error is compared in terms of magnitude against a threshold value, and depending on this comparison, sufficient quality of the parameterization values can be determined. In this way, the required minimum quality of the parameterization values can be determined in an advantageous manner by adjusting the threshold value, but the number of iterations of the method according to the invention to be carried out can also be influenced, whereby the time required for calibrating the vehicle component (3) can ultimately be reduced, which is the stated goal of the method according to the invention.

In a preferred embodiment of the method according to the invention, the change in the parameterization values of the model parameters of the replacement model (13) within the iterations of the method according to the invention is used as an evaluation criterion as a measure of the quality of the specific parameterization values. The background is that the method according to the invention derives the time saving potential in particular from a reduction in measuring points, whereby not the entirety of the time saved by a grid or measurement The measuring points specified in the flowchart are measured, but rather the entirety of the measuring points are determined by a selection model (15). This can, for example, lead to a higher uncertainty about the generated model values than calibration values and thus a higher prediction error in some areas of measured maps, but at the same time these areas may have little practical relevance in the use of the end application of the vehicle component to be calibrated (3). , so that in some areas higher deviations in calibration values are permitted than in others. If the process sequence of the method according to the invention has now already gone through several iterations and it can be determined that the parameterization values of the replacement model (13) only change to a small extent, then a sufficient quality of the approximation to the measured values is already acceptable. Accordingly, in one embodiment, the absolute changes in the parameterization values, for example as a difference from the previous iteration, can be used as a measure of the quality of the determination of those parameterization values and evaluated by comparing them against a threshold value.

In an alternative embodiment, the termination criterion can be firmly determined as an evaluation criterion by a fixed number of iterations that the process flow of the method according to the invention must go through. In some embodiments, specialist knowledge and historical experience regarding the physical properties of the vehicle component (3) to be calibrated can be used, so that the relationship between a required minimum number of iterations of the process flow and the quality of determining the parameterization values for the replacement model that can be achieved in this way (13) is known. Through such a procedure, on the one hand, historical empirical values can be included in the method according to the invention and, on the other hand, the time required for determining the parameterization values can advantageously be estimated in advance. For this purpose, in some embodiments, data about the characteristic and physical properties of the vehicle component (3) to be calibrated, the structure and arrangement of the measuring device (2) used, the vehicle control unit (6) and the measuring device (7) and the data about the The process flow of the calibration process can be stored in a database to be used as additional information when determining the number of iterations of future procedures.

If it is now determined in the seventh step S7, for example after the first iteration of the process flow, that the quality of the specific parameterization values is to be assessed as insufficient, for example measured against one of the evaluation criteria mentioned, comprising a prediction error, a parameter change of the replacement model (13) and/or a number of iterations already completed, the process flow jumps to the first step S1, according to the process flow 2 a selection strategy for selecting the new measuring point is determined by a selection model (15).

Referring to 4A In some embodiments, the prediction model (14) is trained to reduce the prediction error generated by the replacement model (13), so that, for example, in a first iteration, the measuring point 21 of the grid of the speed-torque map of the EM (4) was measured as the current measuring point, the determination of the prediction error according to step S5 from the prediction error map 4A generated. According to the previous description of the method according to the invention, the prediction error at the newly measured measuring point 21, which was generated by the replacement model (13), is low (white) and dark areas, for example (4), continue to form (black), which have a high prediction error exhibit. According to step S6, the prediction model (14) is trained to minimize the prediction error and to select the new measurement point with high probability in one of the (4) black areas, which is shown by the white dashed arrow in the exemplary embodiment shown, according to step S2 , selecting the new measuring point. According to the invention, training a neural network for the prediction model (14) using RL opens up the advantage that new measurement points are selected at the points where the greatest prediction error, or in other words the greatest optimization potential of the replacement model (13), is present, so that the change in the parameterization of the replacement model (13) converges relatively quickly, although the uniform measurement of the entire space of the map is uncertain.

In an alternative embodiment of the method according to the invention, a space coverage method, or in other words a “space filling design”, is used within the selection model (15) instead of the neural network of the prediction model (14) trained using RL to take the prediction error into account. Space coverage methods are known to those skilled in the art as methods for generating points in a space in order to best cover the space. In an advantageous embodiment, within the off choice model (15) applied the Halton sequence method as the spatial coverage method. Alternatively, any other known space coverage method can be used insofar as a homogeneous filling can be achieved by selecting corresponding measuring points from a map of the measured values of the vehicle component (3) to be calibrated. The use of spatial coverage methods within the selection model (15) offers the advantage that a uniform scanning of the specified grid or measurement schedule is achieved by selecting and measuring corresponding measuring points.

Referring to 2 In the second step S2 of the current iteration, the new measurement point is selected by the selection model (15), preferably advantageous according to the invention by a combination of a neural network trained using RL, which reduces the prediction error of the previous iteration, and with a spatial coverage method, which ensures the homogeneous filling of the specified grid or measurement flow plan is taken into account. In a preferred embodiment, the selection of the measuring point to be newly measured is based on a new decision in each iteration as to whether the currently newly selected measuring point is determined using a neural network, taking into account the prediction error, or using spatial coverage methods. For this purpose, a random drawing between the two selection options can be used, with each of the two selection options being assigned a probability with which the associated selection option will be carried out. Accordingly, the sum of both probabilities is 1. In an advantageous embodiment, both the selection using a neural network and the selection using the spatial coverage method are assigned a probability of 0.5, so that each of the two variants is carried out within one iteration with a probability of 50% . Alternatively, any further division can be used, so that the application-appropriate division can be carried out in an advantageous manner by the user of the method according to the invention. This implies the assignment of a probability of 1 or 0.

Referring to 4A In a first iteration of the process flow of the method according to the invention, the measuring point 21 was selected and measured as an example, so that model values in the form of an associated speed-torque map of the EM (4) were predicted by the substitute model (13) and the prediction error was predicted by the computing unit (11). , as shown, is determined. According to one embodiment, the area around the currently measured measuring point 21 has a low prediction error and there are areas (black) with a high prediction error. Within the current iteration of the process flow, the selection of the new measuring point 22 on the basis of the neural network, taking into account the prediction error, has now been carried out, for example, by the selection model (15) within the second step S2. Now that this new measuring point 22 has been measured according to the third step S3, the replacement model (13) has been parameterized again according to the fourth step S4, the prediction errors have been determined again according to the fifth step S5 and the prediction model (14) has been trained again according to the sixth step S6, The prediction error map results, for example, according to 4B , whereby, depending on the defined probability, for example, a first potential new measuring point 23' is selected according to the selection using a neural network, taking into account the prediction error (S2'), or a second potential new measuring point 23" is selected according to the selection using the spatial coverage method (S2") becomes. It should be noted that the method according to the invention should not be defined or limited by the exemplary embodiments described; in particular, no quantitative statements about individual exemplary embodiments can be derived from this. Rather, the qualitative characteristics of the individual designs should be highlighted in order to underpin the general understanding of the causal relationships.

Referring to 2 In an eighth step S8, the parameterization values are transferred if one or more of the evaluation criteria described in the seventh step S7 is met as a termination criterion, comprising the achievement of a sufficiently low prediction error, a small change in the parameterization of the replacement model (13), compared to a previous iteration, or a fixed number of previous iterations. In a preferred embodiment, transferring the parameterization values includes transferring calibration values generated by the replacement model (13) to the vehicle control unit (6) of the measuring device (2), which are coupled to the output interface (10) of the parameterization device (8) via a data connection can. Alternatively, all other known methods for data transmission can be used, provided they are set up to transmit predicted calibration values from the parameterization device (8), in particular from its electronic memory (12), to the vehicle control unit (6). In some embodiments, the calibration values are alternatively or additionally transmitted to an external data server from which the calibration values can be further distributed. For example, updates can the software of vehicle control units (6) can be transmitted directly to all vehicles in a vehicle fleet via OTA software updates.

The advantages of the method according to the invention result from the fact that all steps of the method according to the invention can be automated, so that the entire process of determining parameterization values for a vehicle control unit (6) of a vehicle component (3) to be calibrated can run automatically. Through the serial measurement of measuring points within the individual iterations of the method according to the third step S3, specific selected measuring points are measured, so that the need for a grid-based complete measurement of entire characteristic curves or maps is eliminated, which results in enormous time savings. By cyclically adapting the model parameters of the physical replacement model (13), taking into account the current measurement data, according to the fourth step S4, the prediction model (14) is continuously improved within the application of the method. By evaluating the quality of the prediction model (14) by determining a prediction error according to the fifth step S5, the current state of the determination of the parameterization values can be quantitatively assessed at any time in the process sequence of the method according to the invention. By evaluating the parameterization values determined in this way according to the seventh step S7 both directly, through the prediction error and indirectly, by changing the parameterization of the physical replacement model (13) and / or statically, by determining the number of iterations, are advantageous and manner, several termination criteria are given, by means of which the method according to the invention can be adapted to a ratio between the achievable quality of the determination of the parameterization values and the required time period. By combining the consideration of the prediction error in a neural network trained using reinforcement learning and a space filling method according to the first step S1, the measurement point reduction can be flexibly designed for rapid convergence or homogeneous sampling. Through the specific selection of measuring points to be measured according to the second step S2, an effective reduction in measuring points and thus a time saving when calibrating vehicle control units (6) can be achieved. The electronic transmission of the predicted and well-tested calibration values according to the eighth step S8 makes it possible to update the software or control functions of vehicle control units (6) in an application-oriented and rapid manner on the test bench, but also directly in the field.

Reference symbol list

1

system

2

Surveying facility

3

Vehicle component

4

Electric machine

5

frequency converter

6

Vehicle control unit

7

Measuring device

8th

Calibration facility

9

Input interface

10

Output interface

11

Computing unit

12

Electronic storage

13

Replacement model

14

Prediction model

15

Selection model

16

Engine map

17

Slip characteristic curve

18

empty measuring point

19

measured measuring point

20

maximum torque

21

current measuring point

22

new measuring point

23

potentially new measuring point

f

Slip frequency

n

number of revolutions

I

Current amplitude

M

Torque

QUOTES INCLUDED IN THE DESCRIPTION

This list of documents listed by the applicant was generated automatically and is included solely for the better information of the reader. The list is not part of the German patent or utility model application. The DPMA assumes no liability for any errors or omissions.

Cited patent literature

• DE 102017201632 A1 [0003]
• DE 102019008400 A1 [0004]

Claims (10)

Computer-implemented method for determining at least one parameterization value of at least one model parameter of a physical replacement model (13) of a vehicle component (3), the method comprising carrying out the following steps using a parameterization device (8): • (S1) determining the selection strategy for selecting one or more measuring points to be measured using a selection model (15); • (S2) selection of at least one measuring point by the selection model (15); • (S3) measuring the at least one selected measuring point using a measuring device (2) in order to obtain at least one measured value; • (S4) determining at least one parameterization value for the replacement model (13) depending on the at least one measured value, by a prediction model (14) and generating at least one model value associated with the measured values; • (S5) determining a prediction error by the prediction model (14) by comparing at least one measured value with at least one model value; • (S6) training the prediction model (14), depending on the prediction error determined; • (S7) evaluating the quality of the specific parameterization values based on at least one evaluation criterion and transferring those parameterization values (S8) if the quality is sufficient or re-executing the previous steps (S1 - S7) otherwise. Procedure according to Claim 1 , characterized in that the evaluation of the quality of the specific parameterization values involves the evaluation of one or more evaluation criteria, comprising at least one prediction error, the change of at least one model parameter of the replacement model (13), based on the previous iteration of the method and the number of executed iterations of the Procedure includes. Procedure according to Claim 2 , characterized in that the evaluation of the quality of the specific parameterization values includes comparing the numerical value of one or more evaluation criteria with a threshold value. Method according to one of the preceding claims, characterized in that the prediction model (14) comprises a neural network and that training the prediction model (14) comprises reinforcement learning, with the aspect of minimizing the prediction error, the prediction error being determined by forming the difference between at least a measured value and at least one model value can be determined. Procedure according to Claim 4 , characterized in that the selection of a new measurement point by the selection model (15) includes the application of the neural network or a spatial coverage method, with each iteration of the method deciding on the selection either by means of the neural network or the spatial coverage method, or, within During the first run of the method according to the invention, one or more measuring points can be preselected as starting measuring points. Procedure according to Claim 5 , characterized in that the application of the neural network or the spatial coverage method is decided at each iteration of the method using random drawing, with each of the two use cases being assigned a defined probability. Method according to one of the preceding claims, characterized in that the vehicle component (3) comprises an electric drive, comprising at least one electric machine (4) and/or a frequency converter (5), wherein the measurement of measuring points includes the measurement of the slip characteristics ( 16) at support points (18, 19) of the current amplitude-speed map of the electrical machine (4). Procedure according to Claim 7 , characterized in that measuring the slip characteristic curve (16) determines the maximum torque (M) of the electrical machine (4) as a function of the slip frequency (f) to the respective support point (20) of the current amplitude-speed characteristic map of the electrical machine (4 ) of the current measuring point. Procedure according to Claim 8 , characterized in that measuring the slip characteristic curve (16) includes the application of an optimization method using “Golden Section Search” (GSS). System (1) for determining at least one parameterization value of at least one model parameter of a physical replacement model (13) of a vehicle component (3), which is set up to carry out a method according to one of the preceding claims, comprising: a measuring device (2) which is set up, in order to • measure at least one measuring point of a vehicle component (3) in order to at least one to obtain measured values (S3) and • to transmit measured values to a parameterization device (8); a parameterization device (8), comprising at least one computing unit (11), an electronic memory (12), an input (9) and output interface (10), which is set up to use the computing unit (11) • at least one parameterization value to determine (S4) a model parameter for a replacement model (13) stored on the electronic memory (12) as a function of at least one measured value transmitted by the measuring device (2); • to determine a prediction error of the prediction model (14) by comparing at least one measured value with at least one model value (S5); • train the prediction model (14) depending on the prediction error (S6); • to evaluate the quality of the specific parameterization values based on at least one evaluation criterion (S7) and to carry out a transfer of the specific parameterization values depending on the evaluated quality (S8); • a selection strategy for selecting a new measuring point, taking into account the prediction error, to be determined by a selection model (15) stored on the electronic memory (12) (S1), and • to select a further measuring point using the selection model (15) (S2).

Images