Method for training a machine learning model and method for determining parameter values of an actuator

Abstract

The invention relates to a method for training a machine learning model for use in determining parameter values of one or more parameters of an actuator, in particular as part of a quality check, the method comprising: providing (212) training data, the training data comprising: measured values (204) of one or more measurement variables relating to the actuator, the measured values having been acquired, at least in part, by means of a measuring device; and parameter information relating to the one or more parameters (208) of the actuator; adapting (216) the machine learning model on the basis of the training data such that, by means of the machine learning model, on the basis of measured values from a target actuator, parameter values of the one parameter or of at least one of the plurality of parameters of the target actuator are determined; and providing the adapted machine learning model.

Description

[0001] Description

[0002] title

[0003] Method for training a machine learning model and method for determining parameter values ​​of an actuator

[0004] The present invention relates to a method for training a machine learning model for use in determining parameter values ​​of one or more parameters of an actuator, in particular in the context of a quality inspection, a method for determining such parameter values, as well as a computing unit and a computer program for carrying out the method.

[0005] Background of the invention

[0006] Actuators, such as air system actuators, can be used in a variety of applications. Such actuators can be manufactured in large quantities, often from the same components, even if their specific applications vary. Such actuators are typically designed to be functionally tested before their final use.

[0007] Disclosure of the invention

[0008] According to the invention, various methods, in particular computer-implemented methods, as well as a computing unit and a computer program for carrying out the methods are proposed, having the features of the independent patent claims. Advantageous embodiments are the subject of the dependent claims and the following description.

[0009] The invention relates to actuators and their parameters. In particular, the invention will be explained using the example of air system actuators. In embodiments, the invention also relates to the collection of product parameters from the ongoing production of actuators or air system actuators, which can then be used, for example, as a basis for data-driven applications such as quality assurance in production or the creation of digital twins for further use cases. At this point, it should be mentioned that the invention is not only applicable to air system actuators, but also to any other actuators for which measured values ​​of measured variables can be recorded and which have concrete parameters for identification. This includes, for example, actuators with an electric motor or hydraulic drive and the like.

[0010] Air mass flows in systems such as internal combustion engines or fuel cells are typically adjusted by so-called air actuators (i.e., an air system actuator). Such air actuators can be developed for use in various applications (gasoline, diesel, gas, or hydrogen engines, stationary or mobile fuel cells, etc.). These air actuators generally have a similar functional principle and design, i.e., they are essentially assembled from the same subcomponents (e.g., DC or BLDC motor, gearbox, housing, springs, etc.) and operated by a control unit with essentially the same software (e.g., controlling the motor using a pulse-width-modulated voltage signal, reading back the actuator position as a sensor voltage).

[0011] One method that is possible for technical and economic reasons (simple and robust, short cycle time) is to check the correct function of the actuator or product by testing directly accessible measured variables such as the mechanical adjustment range, the step response times and the return times to the emergency position (if available).

[0012] The disadvantages here, however, are that the testing is carried out in a closed control loop, which can compensate for or mask negative product characteristics, making detection more difficult. For example, changes in friction can disrupt the small-signal behavior of the actuator; however, this plays a negligible role in a step response time when the device is at full capacity. The testing is carried out, for example, under boundary conditions and according to a test logic that differs from later use, for example in the vehicle, so that the transferability of the test results can be limited. For example, engine parameters can change significantly at high temperatures and impair control quality, but are unnoticeable under test conditions at room temperature.

[0013] The test can be conducted using highly aggregated metrics, making it difficult to draw detailed conclusions about the causes or sources of a deviation. An example is an increased step response time of the actuator, which could be caused by increased electrical resistance, increased mechanical friction, a weak motor, or a combination of these.

[0014] The risk of deviating or scattering product properties increases in the context of the transformation of the automotive industry through the development of new and relocation of existing production sites, diversification of supply chains and general material and design adaptations.

[0015] In summary, various deficiencies or limitations can arise from production monitoring approaches based on this or other approaches. The focus of monitoring in the context of, for example, Industry 4.0 is usually not the actual relevant actuator or product (e.g., the air actuator) and its parameters, but rather the parameters of the process and machines along the production line used to manufacture the actuator. However, from the perspective of the subsequent application and also the customer, the function and parameter tolerances of the actuator or product are (usually the only) relevant.

[0016] The result of monitoring is usually a pass / fail classification or anomaly detection, which, however, is unable to resolve trends or batch differences below the diagnostic threshold. However, particularly with increasing diversification (e.g., changed / different suppliers) and relocations (e.g., changed production lines), detailed and early detection of differences is of great importance. Monitoring product functionality can often be characterized either by a significantly abbreviated test, which, with its short runtime optimized for mass production, is cost-effective but functionally incomplete (production line), or by a time- and technically complex test that validates the product's function in detail but, due to its runtime and effort, is unsuitable for mass production (line release).

[0017] Within the scope of the present invention, possibilities are proposed to improve the monitoring or testing of the actuator or product by recording and evaluating the parameters relevant to function (system identification), e.g., from existing test sequences in the production lines. In contrast to conventional approaches to system identification, however, no mathematical model (i.e., equations) is created to describe the actuator and adapted to the measured values ​​(or measurement data). Instead, a generalized approach is used, independent of the domain and the underlying system.

[0018] One embodiment relates to the generation of training data for training a machine learning model for use in determining parameter values ​​of one or more parameters of an actuator, in particular in the context of a quality inspection. Measured values ​​of one or more measured variables relating to the actuator are received (e.g., in a computing unit) or provided, wherein the measured values ​​have been acquired at least partially by means of a measuring device. The measured values ​​can be present, for example, as one or more time series (e.g., one time series per measured variable). Such measured values ​​can later serve as model inputs or features, e.g., so-called step response and decay curves.

[0019] In addition, parameter information relating to one or more parameters of the actuator is received (e.g., in the computing unit) or provided. Such parameters correspond to the parameters (or system parameters) of the product used to record the measured values ​​that are to be determined (or estimated) later in the application; in the case of a throttle valve as an actuator, these can be, for example, mechanical or electrical parameters of the motor and / or the springs. The parameter information is, for example, information or data that can be processed by a computing unit and indicates which parameters are involved and, in particular, which specific values ​​they have.

[0020] Training a machine learning model typically requires that the training data contain both inputs (in this case, the measured values) and target variables to be estimated later (in this case, the parameters). Quantitative values ​​or examples of such parameters are therefore required, such as concrete values ​​for electrical resistance (e.g., 1.5 ohms, 1.2 ohms, 1.3 ohms, etc.). However, it is usually not sufficient to simply have metadata such as "resistance."

[0021] While the parameters considered include, for example, the aforementioned mechanical stop or adjustment range, the step response time, and a return time to the emergency position (or in an emergency situation), these parameters are particularly preferred, including those that cannot conventionally be read or measured, such as actuator constants such as a motor constant, an electrical resistance, a spring constant, and a friction coefficient.

[0022] The measured values ​​and the parameter information are then provided as training data. The training data can be generated in various ways; the parameters can, for example, also be obtained from measuring the throttle valve or its components (or the corresponding actuator) using other measuring methods on other test devices. However, a particularly preferred approach when a system identification method is known is if at least some of the measured values ​​are or have been recorded as part of an identification method for identifying the product. Then, for example, one or at least one of the several parameters can be or have been determined as part of the identification method. For this purpose, a (known or already used) identification method can, for example, be extended to include the test steps of the production line.The actuator parameters (system parameters) can then be determined or identified from the original portion of the identification process as target variables or labels (y) of the training data. The portion of the identification process expanded in this way (i.e., the production line test steps) can be used as input variables or features (x) of the training data.

[0023] A testing procedure at the so-called selection test station is particularly suitable for this purpose, which provides for a very detailed but correspondingly complex test for initial parts at the start of the line or a statistical sample selection.

[0024] One embodiment relates to a method for training a machine learning model for use in determining parameter values ​​of one or more parameters of an actuator, in particular in the context of a quality inspection. For this purpose, training data is provided, wherein the training data comprises: measured values ​​of one or more measured variables relating to the actuator, wherein the measured values ​​have been acquired at least partially by means of a measuring device; and parameter information relating to the one or more parameters of the actuator. This can, in particular, be the training data obtained in the manner explained above.

[0025] The machine learning model is then adapted based on the training data, so that, using the machine learning model, parameter values ​​of one or at least one of the multiple parameters of the target actuator are determined based on measured values ​​from a target actuator. This adaptation may, for example, include adjusting the weights between neurons of the machine learning model, which is designed as a neural network, e.g., until the neural network delivers the desired output data.

[0026] In principle, as mentioned, weights can be adjusted in or by the machine learning model. One aspect of the model architecture used here is that adjusting the weights not only optimally adapts the machine learning model to the training data, but also allows a selection of different possible paths within the machine learning model. In other words, not only is a predefined path used to optimally train the model, but rather, a decision can also be made initially for one of several paths—i.e., from a set of paths—in order to find the optimal path for a current problem.

[0027] Thus, there is no need to make a (usually non-trivial) decision about which model architecture should be used to solve the specific problem – and this is precisely what is part of model training for the concept presented here. The same applies, in particular, to feature engineering, which is usually a preceding step, as will be explained in more detail below.

[0028] The adapted machine learning model is then made available, for example, for later use in a quality inspection. The parameter values ​​here can include, for example, the aforementioned parameter information, but especially regarding the specific target actuator. Preferably, however, the parameter values ​​include information about the quality of the target actuator; this then enables a quality inspection or other assessment of the target actuator—i.e., a specific, tested actuator.

[0029] One embodiment relates to a method for determining parameter values ​​of one or more parameters of an actuator, in particular within the scope of a quality inspection. For this purpose, measured values ​​of one or more measured variables relating to the actuator are provided, wherein the measured values ​​have been recorded at least partially by means of a measuring device. Based on a machine learning model and the measured values, parameter values ​​of the one or at least one of the plurality of parameters of the actuator are then determined. These parameter values ​​are then provided, e.g., within the scope of a quality inspection. The machine learning model can, in particular, be the training data obtained or trained in the manner explained above. One idea here is that the behavior of a system (the function of the actuator orProduct) is completely determined by its parameters (system parameters) and is observable through the temporal progression (the test sequence) of the relevant measured variables (state variables) such as spatial position, electrical voltage, electrical current, and the like – regardless of whether the description of the system in the form of mathematical equations is known or even solvable. This means that in existing test sequences, the information about the desired system parameters is encoded and can, in principle, be extracted using a suitable (but generally unknown) method.

[0030] For this purpose, a machine learning model or, more generally, artificial intelligence is used. The usual procedure in machine learning for finding a suitable model is to transform the initial values ​​or initial data into so-called features ( ), which are then trained with the target variables or labels (y) as a data set ( ,y) using a model approach (f) and tested for statistical significance. If the result is unsatisfactory, other model approaches (e.g., other architectures or architectural characteristics) or other feature combinations or features (e.g., through other transformations of the initial data) must be tried, and the process repeated.

[0031] In contrast, the approach proposed here can be trained directly on the source data and requires neither a prior transformation of the measured values ​​or input data (time series) into suitable features, nor the use of a specific mathematical model in the form of equations, as already mentioned above.

[0032] For example, from "Christian Szegedy et al: Going Deeper with Convolutions, 2014. http: / / ar-xiv.org / abs / 1409.4842," a model approach in the form of a CNN (Convolutional Neural Network) is known as a machine learning model for classification and object recognition in images based on 2D convolution. Simply put, object recognition occurs by extracting features of increasing complexity from the original image through the various layers of the neural network (e.g., curves, then circles, then a pair of eyes, then a face) and finally linking them to the target variable (a human portrait).

[0033] Similarly, in this case, a 1D convolution can be applied to time series data (the available measurement data) to generate derived features (e.g., velocities and accelerations from position signals). Since the parameters of the convolution layers are part of the parameter training, the generation of suitable features occurs automatically, and no explicit selection or decision is required by the user.

[0034] In one embodiment, the machine learning model comprises an extractor component and a regressor component. Using the extractor component, the measured values ​​obtained as one- or multi-dimensional, time-series-based input values ​​can be transformed into one or more intermediate variables (so-called "latent space"). Using the regressor component, a relationship between the one or at least one of the multiple intermediate variables and the parameter values ​​to be output as output values ​​can be determined or established. Optionally, the machine learning model can comprise a validator component, which can be used to determine one or more characteristic values ​​that indicate whether the measured values ​​lie within a predetermined (or known) range. This allows, for example, a prediction of the machine learning model not to be rejected per se.

[0035] The extractor component can, for example, be implemented as a CNN model with one or more levels (with respect to the depth of the model) and, for example, several parallel branches of different shapes (hyperparameterization). A depth with multiple levels enables the model to learn increasingly non-linear relationships. The parallel branches enable the model to select the most suitable from several possible operations and to further adapt them to the task through parameter training. If several such sub-models are connected in parallel, the information content of several, even temporally unrelated, test profiles can be processed simultaneously. The regressor component can, for example, be implemented as an MLP model (multi-layer perceptron, conventional neural network) in one or more levels (with respect to the depth of the model), and links the intermediate variables calculated in the extractor component with the model output variables as a regression.

[0036] The validator part can be implemented as an autoencoder (e.g. CNN for reconstructing the input signals), as a clustering model or as a density model.

[0037] In summary, this offers several advantages. For example, it allows monitoring the product instead of the process. The described approach is capable of capturing the system parameters relevant for assessing the product's subsequent function.

[0038] It is possible to detect changes below the error threshold. The system parameters determined using the described procedure, as continuous (or quasi-continuous) values, allow for the early detection of even minor changes in the product that fall below the usual diagnostic thresholds with a pure pass / fail assessment.

[0039] It is also possible to identify the source of the error. Using domain knowledge, a deviation in a system parameter can be mapped to product components or associated production steps and correlated manually or programmatically with process or machine parameters. This allows conclusions to be drawn about a production step or component (which may be a supplier product).

[0040] In terms of cost efficiency and risk minimization, the described approach is backward compatible with a conventional production line setup and does not require risky changes or uneconomical extensions of the test routines. Generalizability and scalability are ensured. The described approach requires the acquisition of the state variables (measured variables) relevant for the system description in suitable test profiles as input data, but does not require an explicit description of the system (e.g., using differential equations that are completely or partially unknown or unsolvable) or the explicit extraction of key figures from the input data.

[0041] A computing unit according to the invention, e.g. an engine control unit or a computer, is configured, in particular in terms of programming, to carry out a method according to the invention.

[0042] The implementation of a method according to the invention in the form of a computer program or computer program product with program code for carrying out all method steps is also advantageous, since this entails particularly low costs, in particular if an executing control unit is also used for other tasks and is therefore already present. Finally, a machine-readable storage medium is provided with a computer program stored thereon, as described above. Suitable storage media or data carriers for providing the computer program are, in particular, magnetic, optical, and electrical memories, such as hard disks, flash memories, EEPROMs, DVDs, and others. Downloading a program via computer networks (Internet, intranet, etc.) is also possible. Such a download can be wired or cable-based or wireless (e.g., via a WLAN network, a 3G, 4G, 5G, or 6G connection, etc.).

[0043] Further advantages and embodiments of the invention will become apparent from the description and the accompanying drawings.

[0044] The invention is illustrated schematically in the drawing using exemplary embodiments and is described below with reference to the drawing.

[0045] Brief description of the drawings Figure 1 shows schematically a test setup to explain the invention.

[0046] Figures 2a, 2b schematically show a sequence of a method in one embodiment.

[0047] Figure 3 shows schematic diagrams to explain the invention.

[0048] Figures 4a, 4b schematically show a machine learning model in one embodiment.

[0049] Embodiment(s) of the invention

[0050] Figure 1 schematically illustrates a test setup 100 for explaining the invention. An actuator 110 designed as a throttle valve is shown as an example, as well as an engine control unit 120, by means of which the throttle valve 110 can be controlled. Three control lines (representing control signals) are shown as examples. In this respect, this can be, for example, a conventional test setup for testing and / or checking the throttle valve 110.

[0051] Furthermore, the test setup 100 comprises a measuring device or, more generally, measuring technology 130, by means of which measured values ​​of measured variables of the throttle valve 110 can be recorded. In addition, a computing unit 140 is shown, in which the recorded measured values ​​can be received and then processed and / or evaluated.

[0052] Figure 2a schematically shows a flow of a method in one embodiment, which here includes, by way of example, the generation of training data, the training of the machine learning model, and its application. Figure 2b shows an illustration of specific steps.

[0053] In step 200, a time series 202 of one or more measured variables relating to the actuator is acquired. A portion of the time series can be used, e.g., initially, as measured values ​​204 for the present procedure. In step 206, these measured values ​​204 can then be provided or, e.g., received in the computing unit, as mentioned in Figure 1. The measured values ​​are acquired at least partially using a measuring device.

[0054] Figure 2a shows the time series 202 with the measured values ​​204, with the measured values ​​204 being enlarged and shown as three sections 204a, 204b, and 204c of the time series. Each section shows exemplary measured values ​​for three different measured variables. These three sections can, for example, each comprise different phases.

[0055] These are, in particular, different test steps from the overall test sequence (e.g., step response in the opening direction, section 204a, step response in the closing direction, section 204b, fallback curve in the emergency position, section 204c). This does not involve a long, continuous profile or a long, continuous series of measurements, but rather several independent parts. Accordingly, the machine learning model should also have several independent inputs.

[0056] Based on the time series 202, one or more parameters 208 of the actuator or corresponding parameter information about them can also be determined or identified, which are then provided in step 210 or, for example, received in the computing unit.

[0057] As already mentioned, the time series recorded as part of a (e.g., existing) system identification can be extended by the relevant measured values ​​204, which then results in time series 202. However, if the desired system parameters from a system identification are already available, time series 202 is not needed at all. The measured values ​​204 (X) and these system parameters (y) are then sufficient to obtain a data set (X, y) for training—i.e., the training data. Time series 202 is therefore only required if parameters 208 are not available from another source.

[0058] In step 212, the measured values ​​and the parameter information for the parameters 208 are then provided as training data 214. It should be noted at this point that the training data does not only include the measured values ​​and parameter information of a single actuator, but typically of a plurality of actuators and / or test procedures for them.

[0059] In a step 216, the machine learning model 218 is then trained. Here, the machine learning model 218 is adapted based on the training data, so that, using the machine learning model, parameter values ​​of one or at least one of the multiple parameters of the target actuator are determined based on measured values ​​from a target actuator. In step 220, the machine learning model 222 adapted (or trained) in this way is then made available so that it can be used later.

[0060] The (adapted) machine learning model can then be applied, for example, using a computing unit 224, on which the machine learning model is then stored and executed. The computing unit 224 can also be, for example, the computing unit 140 shown in Figure 1. The application of the machine learning model then allows the determination of parameter values ​​for one or more parameters of an actuator, particularly in the context of a quality inspection.

[0061] For this purpose, in step 226, measured values ​​of one or more measured variables relating to the actuator (to be tested) are provided, wherein the measured values ​​have been recorded at least partially by means of a measuring device, as also shown, for example, in Figure 1. In step 228, parameter values ​​230 of one or at least one of the several parameters of the actuator are then determined based on the machine learning model and the measured values. In step 232, the parameter values ​​are then provided; this is, for example, an estimate or prediction. The parameter values ​​can, for example, in addition to information about the parameters themselves, also include information about a quality of the target actuator, in particular related to the parameters.

[0062] Figure 3 shows diagrams of various variables and parameters to explain the invention, wherein in each diagram a position signal 302 of the actuator is plotted against time 300. The position signal of the actuator is, for example, physically a voltage that is proportional to the opening angle of a flap of the actuator. It should be mentioned at this point that Figure 3 is for explanatory purposes and shows test profiles typically used today with the parameters determined from them. For this purpose, values ​​and time intervals are read off, for example, in certain situations.

[0063] In contrast to this, the procedure proposed within the scope of the present invention also determines parameters such as the motor constant, the electrical resistance, the spring constant, or the friction coefficient, which generally cannot be derived from the test profiles. Without the use of the invention, with precise knowledge of a model (differential equations), this model would be adapted to the observed measurement curves by adjusting the parameters, thus obtaining the parameters. In contrast, the presented method represents a possibility of identifying the parameters even without any knowledge of the model. This is particularly advantageous when modeling is too complex or a model is generally unknown.

[0064] The upper diagram illustrates the position of the actuator or one of its components using position signal 310. 312 represents a lower mechanical stop, and 314 an upper mechanical stop. These two parameters, 312 and 314, thus indicate a limit to the actuator's adjustment range.

[0065] The middle diagram illustrates a step response time of the actuator or its components with the position signal 320. 322 indicates a step response time in a first adjustment direction, and 324 indicates a step response time in a second adjustment direction.

[0066] The lower diagram illustrates, with position signal 330, a release time of the actuator or its components in an emergency situation. This release time is indicated by 332. Figure 4a schematically shows a machine learning model 400 in one embodiment. The machine learning model 400 includes an extractor portion 410, a regressor portion 420, and optionally a validator portion 430.

[0067] The extractor portion 410 can be implemented, for example, as a CNN model with one or more levels (regarding the depth of the model) and, for example, several parallel branches of different shapes (hyperparameterization). Three parallel branches, each with a certain depth, are shown as an example. Each of the three branches receives measurement values ​​404a, 404b, 404c from one of the phases (as shown in Figure 2b with 204a, 204b, 204c).

[0068] Multi-level depth allows the model to learn increasingly nonlinear relationships. Parallel branches allow the model to select the most suitable operations from several possible operations and further adapt them to the task through parameter training. If several such submodels are connected in parallel, the information content of several, even temporally unrelated, test profiles can be processed simultaneously.

[0069] The regressor portion 420 can, for example, be implemented as an MLP model (multi-layer perceptron, conventional neural network) in one or more layers (relative to the depth of the model) and links the intermediate variables calculated in the extractor portion 410 with the model output variables as a regression. The parameter values ​​408 are output.

[0070] The validator portion 430 can be implemented, for example, as an autoencoder (e.g., a CNN for reconstructing the input signals), as a clustering model, or as a density model. Thus, output data 406a, 406b, 406c are output for each branch at each phase.

[0071] Figure 4b shows a portion of the machine learning model 400 in a further embodiment, namely a subunit of the extractor portion 410. Here, for example, there is an input layer 450, followed by several parallel paths 454, 452, 456, and 458. These are linked in layer 460 and then output in layer 462.

[0072] The individual paths differ here, for example, in the layers selected within them. For example, a signal can be sampled in different step sizes, operations can be performed with different sampling lengths, etc. By connecting these possible layers (i.e. the different paths) in parallel, the machine learning model can select the most suitable layers when merging in step or level 460 or when reducing in step or level 462. If, on the other hand, the model were purely sequential (the output of one layer corresponds to the input of the next layer, only one path), each layer could and would have to be traversed. Therefore, the machine learning model could not make a selection during training, and unsuitable layers would then impair the quality of the machine learning model.

Claims

Claims 1 . A method for training a machine learning model for use in determining parameter values ​​of one or more parameters of an actuator (110), in particular in the context of a quality inspection, comprising: Providing (212) training data (214), wherein the training data comprises: Measured values ​​(204) of one or more measured variables relating to the actuator, wherein the measured values ​​have been recorded at least partially by means of a measuring device, Parameter information relating to the one or more parameters (208) of the actuator; Adapting (216) the machine learning model (218) based on the training data, such that parameter values ​​of the one or at least one of the plurality of parameters of the target actuator are determined by means of the machine learning model based on measured values ​​from a target actuator; and providing (220) the adapted machine learning model (222).

2. The method according to claim 1, wherein the one or more measured variables comprise at least one of the following measured variables: an actuator voltage; an actuator current; a position of a component of the actuator and / or a voltage relating to the position.

3. The method according to claim 1 or 2, wherein the one or more parameters comprise at least one of the following parameters: a constant of the actuator, in particular a motor or spring constant; an electrical resistance of the actuator; a friction coefficient of the actuator.

4. Method according to one of the preceding claims, wherein at least some of the measured values ​​are or have been recorded as part of an identification method for identifying the actuator.

5. The method according to claim 4, wherein the one or at least one of the plurality of parameters is or has been determined as part of the identification method.

6. Method for determining parameter values ​​of one or more parameters of an actuator, in particular in the context of a quality inspection, comprising: Providing (226) measured values ​​of one or more measured variables relating to the actuator, wherein the measured values ​​have been recorded at least partially by means of a measuring device; Determining (228), based on a machine learning model and the measured values, parameter values ​​(230) of the one or at least one of the plurality of parameters of the actuator; and Providing (232) the parameter values.

7. The method of claim 6, wherein the machine learning model has been trained according to a method of any one of claims 1 to 5.

8. Method according to one of the preceding claims, wherein the parameter values ​​comprise information about a quality of the target actuator.

9. Method according to one of the preceding claims, wherein the machine learning model (400) comprises an extractor part (410), by means of which the measured values ​​obtained as one- or multi-dimensional, time series-based input values ​​are transformed into one or more intermediate variables, and wherein the machine learning model comprises a regressor part (420), by means of which a relationship between the one or at least one of the several intermediate sizes and the parameter values ​​to be output as output values.

10. The method according to claim 9, wherein the machine learning model further comprises a validator portion (430) by means of which one or more characteristic values ​​are determined which indicate whether the measured values ​​are within a predetermined range.

11. Method according to one of the preceding claims, wherein the actuator (110) is designed as an air system actuator, in particular for an internal combustion engine and / or a fuel cell.

12. Computing unit (140) configured to carry out all method steps of a method according to one of the preceding claims.

13. A computer program which causes a computing unit to carry out all method steps of a method according to one of claims 1 to 11 when executed on the computing unit.

14. A machine-readable storage medium having a computer program according to claim 13 stored thereon.

Images