Computer-implemented method for training artificial intelligence

Abstract

The invention relates to a computer-implemented method for training an artificial intelligence in which an energy-based machine learning model is implemented for calibrating a control device of a motor vehicle, comprising the steps: - generating a training data set, whereby individual data sets of the training data set are generated automatically in simulation loops by changing calibration parameters and subsequent simulation in a simulation environment of the control device and output of a calibration result, and - Training the energy-based machine learning model through self-supervised learning, whereby masked areas are automatically generated in the pixel representations of the calibration results in each of the individual data sets of the training data set and the masked areas are predicted from the simulated calibration result using the corresponding calibration parameters and the prediction of the masked areas is controlled using the training data set and the results of the control are implemented in further training.

Description

The present invention relates to a computer-implemented method for training an artificial intelligence.Control devices for drive devices of a motor vehicle, in particular for internal combustion engines or electric drive devices, must first be calibrated before they can be used in regular operation. For example, the operation of an electric drive device of a battery-electric motor vehicle is controlled by means of a corresponding control device. For this purpose, during the operation of the electric drive device, the control device calculates values which depend on parameters. These parameters can include, in particular, an electrical voltage applied to the electrical drive device, a temperature of the drive device, a rotational speed of the drive device, a torque generated by the drive device, a mechanical power of the drive device and / or a state of charge of a battery-electric energy store which supplies the drive device with electrical energy. The calculated values can include, for example, an electric current intensity supplied to the drive device. For the most efficient and accurate control possible of the drive device, it is advantageous if the calculated values have the smallest possible difference from corresponding real values. Calibration parameters of the control device are therefore adapted prior to regular operation of the motor vehicle.It is already known in principle to visualize the calibration results obtained during the calibration of the control device in the form of pixel representations by means of a display device. This visualization enables a user to evaluate the calibration result, in particular by an engineer, by adding annotations corresponding to the calibration results. Currently, the possible effects of missing dataset variables on analysis resulting from latent (not recorded in the dataset) variables in the feature space, which may be, in particular, intervals for physical parameters such as vehicle weight, are not considered.The control logic implemented in the controller is usually extremely complex, as there are many parameters that need calibration. It is usually particularly difficult to simulate the expected calibration results, since the control logic of the control device is often a "black box" unknown to the user. Thus, it is desirable to be able to answer the question which calibration interventions could have led to certain calibration results.The object of the present invention is to provide a method for training an artificial intelligence, in which the artificial intelligence is trained in such a way that it can predict both the calibration result during a predefined calibration of a control device of a motor vehicle and the calibration during a predefined calibration result autonomously without human annotations and only on the basis of pixel analyses.The solution of this task provides a method for training an artificial intelligence with the features of claim 1.A computer-implemented method according to the invention for training an artificial intelligence, in which an energy-based machine learning model for calibrating a control device of a motor vehicle is implemented, comprises the steps:generating a training dataset containing a plurality of individual datasets, each of the individual datasets containing a plurality of calibration parameters and a corresponding calibration result in the form of pixel representations, wherein the individual datasets of the training dataset are automatically generated in simulation loops by changing the calibration parameters and subsequent simulation in a simulation environment of the control device and outputting the calibration result and are stored in a non-volatile storage means in a retrievable manner, andtraining the energy-based machine learning model by self-supervised learning, wherein masked areas in the pixel representations of the calibration results are generated in each of the individual datasets of the training dataset in an automated manner, in particular by a method of image processing, and the energy-based machine learning model is trained by means of the training dataset in such a way that it predicts the masked areas from the simulated calibration result on the basis of the corresponding calibration parameters and controls the prediction of the masked areas on the basis of the training dataset and converts the results of the control during the further training, such that the calibration result can be deduced from a given calibration result by the learned prediction and from given calibration parameters.The energy-based machine learning model, which after training can assign a specific calibration to a given calibration result and a specific calibration result autonomously, is trained in the method according to the invention by self-supervised learning only on the pixel representations, wherein a masking technique is used to generate the masked areas in the pixel representations of the calibration results. Thus, during training, no annotations themselves are made in the training data set by a user, in particular by an engineer. The energy-based machine learning model is preferably trained such that the energy of an energy function of the energy-based machine learning model is minimized if the predetermined calibration parameters are compatible with the calibration result, and that the energy of the energy function is increased if the predetermined calibration parameters are not compatible with the calibration result. The artificial intelligence works in two directions after completion of the training of the energy-based machine learning model and thus enables the prediction of the calibration results on the basis of predefined calibration parameters (analysis) and, moreover, also the prediction of the calibration parameters on the basis of the predefined calibration result (analysis and calibration).The calibration parameters can be present in particular as two- or three-dimensional calibration diagrams or calibration characteristic maps. The calibration parameters, in particular the calibration characteristic maps, and the associated calibration results are collectively stored as individual data sets in the nonvolatile memory device, preferably in the form of a pixel matrix, such that they can be called up. For each calibration, the respective calibration parameters and the associated calibration result are thus available as individual datasets in the training dataset.In an advantageous embodiment, it is proposed that the pixel representations of the calibration parameters and of the corresponding calibration result in the individual data sets are each represented on the same image. This makes it possible to simplify self-supervised learning of the energy-based machine learning model.In a preferred embodiment, it is possible that a hardware-in-the-loop simulation environment or a software-in-the-loop simulation environment is used for generating the training dataset. The hardware-in-the-loop simulation environment employs a real (test) controller. In contrast, the functions of the control software of the control device are emulated in the software-in-the-loop simulation environment.In one embodiment, it is proposed that the calibration parameters are randomly changed during the generation of the training dataset.In advantageous embodiments, an energy-based machine learning model is used which is based on an LVEBM architecture (LVEBM="Latent Variable Energy Based Model") or an LVGEBM architecture (LVGEBM="Latent Variable Generative Energy Based Model") or a JEPA architecture (JEPA=("Joint Embedding Predictive Architecture").Latent variables in the feature space, often referred to as "hidden variables", are used to account for possible variables that have an influence on the calibration task but have not been captured in the individual datasets (either because they are not known to the engineer or, for example, there is not enough memory space available to record them). Latent variables are thus, in other words, variables whose values are not known. However, latent variables would facilitate the actual calibration task if they were known. Examples of latent variables are the vehicle weight, the ambient temperature or the state of charge of a traction battery of the motor vehicle.As mentioned above, in one embodiment, there is a possibility that the energy-based machine learning model is based on an LVEBM architecture (LVEBM="Latent Variable Energy Based Model"). Very simply, the LVEBM architecture is based on the concept of evaluating the degree of compatibility between a variable x (in the present case the calibration) and a variable y (in the present case the predefined calibration result) with the aid of a latent variable z. The latent variable z can be considered as a parameterizing of the set of possible relationships between an x value and a set of compatible y values. Latent variables thus represent information about the variable y that cannot be extracted from the variable x. In the LVEBM architecture, a parameterized energy function E(x, y, z) is implemented with the variables x, y and with the latent variable z. An inference method determines a value for the latent variable z for a variable pair (x, y), which minimizes the energy of the energy function E (x, y, z).In an alternative embodiment, it is also possible for the energy-based machine learning model to be implemented as a (generative) LVGEBM architecture (LVGEBM="latent variable generative energy based model"). This architecture closely resembles the LVEBM architecture. However, since the LVGEBM architecture is generative, it is also advantageously possible to generate alternative calibration results for predefined calibration parameters or alternative calibration parameters for a predefined calibration result.In a further alternative embodiment, it is possible for the energy-based machine learning model to be based on a JEPA architecture (JEPA="Joint Embedding Predictive Architecture"). In general, a JEPA architecture learns to predict the embeddings of a signal y (in the present case a representation of the calibration result) from a compatible signal x (in the present case a representation of the calibration) by using a predictor network which is conditioned with additional, in particular latent, variables z in order to facilitate the prediction. The aim here too is to determine that value for the latent variable z which minimizes the energy of an energy function E (x, y, z).More details and further explanations, in particular regarding LVEBM architectures and JEPA architectures, can be found (with further evidence) for example in the online publication "A Path Towards Autonomous Machine Intelligence version 0.9.2, 2022-06-27" by Yann LeCu (retrievable at: https: / / openreview.net / pdf?id=BZ5a1r-kVsf).In one embodiment, there is the possibility that the energy-based machine learning model is trained by contrasting learning. The energy-based machine learning model is intended to learn to distinguish similar features from dissimilar features in the pixel representations of the training dataset.In an alternative embodiment, it is proposed that the energy-based machine learning model is trained by regularized learning. In this regularized learning method, a range of values of quantities that might have affected the calibration results but were not measured can be chosen as a regularizer for the latent variables.According to a further aspect, the present invention relates to a system comprising an electronic data memory and a digital electronic data processing unit, wherein machine-readable instructions are stored in the data memory, wherein the data processing unit is designed to read out and execute the instructions, and wherein the data processing unit is designed to execute a computer-implemented method according to one of Claims 1 to 8 when the instructions are executed.According to yet another aspect, the present invention relates to the use of an artificial intelligence trained by means of a method according to one of Claims 1 to 8 for autonomous calibration of a control device of a motor vehicle. This control device can be, in particular, a control device of a drive device of the motor vehicle.The method presented here for training an artificial intelligence and the use of the trained artificial intelligence in later productive operation offer, in particular, the advantages listed below:Calibration of the control device of the motor vehicle is possible without information about the software control logic itself, which is usually a "black box".No user interventions, such as adding annotations to the calibration parameters and the calibration results, are required, resulting in a time and cost saving both during training and during productive operation of the artificial intelligence.The energy-based machine learning model learns a feature representation of the calibration problem, wherein not every detail (i.e. not every individual pixel in the respective pixel representation) has to be learned / captured in order to assign a calibration to a predefined calibration result or a calibration result to a predefined calibration represented by the calibration parameters.A possible additional benefit can be seen in that latent variables are taken into account during training and some of them are regularized with values that also make sense in the real world. This can allow the generation of new, realistic calibrations / calibration results with predetermined calibration results / calibrations.Further features and advantages of the present invention will become apparent from the following description of a preferred embodiment with reference to the accompanying drawings. The following are shown: FIG. 1 is a schematic diagram illustrating the basic flow of a computer-implemented method for training an artificial intelligence, FIG. 2 is a schematic diagram illustrating training of an energy-based artificial intelligence machine learning model, FIG. 3 is a schematic diagram illustrating the use of the trained artificial intelligence according to a first variant, FIG. 4 is a schematic diagram illustrating use of the trained artificial intelligence according to a second variant.With reference to FIGS. 1 and 2, the basic sequence of a method for training an artificial intelligence 1 is first illustrated below. FIGS. 3 and 4 then show the use of the trained artificial intelligence 1 in the later productive operation.In the artificial intelligence 1, an energy-based machine learning model 2 for calibrating a control device of a drive device of a motor vehicle is implemented. This drive device can be, for example, an internal combustion engine or an electric machine of the motor vehicle. Details of the energy-based machine learning model 2 are explained in more detail below.In a first step 100, a training data set is first generated, which is used later for the training of the artificial intelligence 1. This training data record contains a plurality of individual data records 3, 4, of which two are schematically illustrated by way of example on the left-hand side in FIG. 2. Each of the individual data sets 3, 4 contains a plurality of calibration parameters 30- 33, 40- 43 which represent the calibration of the control device and a corresponding calibration result 34, 44 in the form of pixel representations. Preferably, the pixel representations of the calibration parameters 30- 33, 40- 43 and the corresponding calibration result 34, 44 in the individual datasets 3, 4 can each be represented on the same image. This makes it possible to simplify self-supervised learning of the energy-based machine learning model 2, which is explained in more detail below.For generating the training dataset, either a hardware-in-the-loop simulation environment or a software-in-the-loop simulation environment may be used. The hardware-in-the-loop simulation environment employs a real (test) controller. In contrast, the functions of the control software of the control device are emulated in the software-in-the-loop simulation environment. The training dataset is automatically created in simulation loops by changing the calibration parameters 30- 33, 40- 43 and subsequent simulation in the simulation environment and outputting the calibration result 34, 44. Preferably, the calibration parameters 30-33, 40-43 in the simulation loops are randomly varied. The calibration parameters 30- 33, 40- 43 can be present in particular as two- or three-dimensional calibration diagrams or calibration maps. The calibration parameters 30- 33, 40- 43, in particular the calibration characteristic maps, and the associated calibration results 34, 44 are stored as individual data sets, preferably in the form of a pixel matrix, in a non-volatile memory device in such a way that they can be used during the later training of the energy-based machine learning model 2.After the training data set with the plurality of individual data sets 3, 4 has been created, the actual training of the energy-based machine learning model 2 takes place in a next step 200. In all architectures, latent variables in the feature space are preferably used to take into account possible variables which have an influence on the calibration task but have not been recorded in the individual datasets 3, 4 (either because they are not known to the engineer or because, for example, there is not sufficient memory space available to record them). Latent variables are thus, in other words, variables whose values are not known. They are therefore frequently also referred to as "hidden variables". Latent variables would facilitate the actual calibration task if they were known. Examples of latent variables are the ambient temperature or the state of charge of a traction battery of the motor vehicle.In one embodiment, the energy-based machine learning model 2 may be based on an LVEBM architecture (LVEBM="Latent Variable Energy Based Model"). Very simply, the LVEBM architecture is based on the concept of evaluating the degree of compatibility between a variable x (in the present case the calibration) and a variable y (in the present case the calibration result) with the aid of (at least) one latent variable z. The latent variable z can be considered as a parameterizing of the set of possible relationships between an x value and a set of compatible y values. Latent variables z thus represent information about the variable y that cannot be extracted from the variable x. In the LVEBM architecture, a parameterized energy function E(x, y, z) is implemented with the variables x, y and with the latent variable z. An inference method determines a value for the latent variable z for a variable pair x, y, which minimizes the energy of the energy function E (x, y, z).In an alternative embodiment, it is possible for the energy-based machine learning model 2 to be based on a JEPA architecture (JEPA="Joint Embedding Predictive Architecture"). Generally, such an architecture learns to predict the embeddings of a signal y (in the present case the calibration result) from a compatible signal x (in the present case the calibration) by using a predictor network which is conditioned by additional, in particular latent, variables z in order to facilitate the prediction. The aim here too is to determine that value for the latent variable z which minimizes the energy of an energy function E (x, y, z).In a further embodiment, it is also possible for the energy-based machine learning model 2 to be implemented as a generative LVGEBM (LVGEBM=latent variable generative energy based model). Since this architecture is generative, it is also advantageously possible to generate alternative calibration results 34, 44 or alternative calibration parameters 30- 33, 40- 43.The training of the energy-based machine learning model 2 is carried out by self-supervised learning, wherein a masking technique is used, by means of which at least one masked area 35, 45 is generated in each case in an automated manner, in particular by a method of image processing, in the pixel representations of the calibration results 34, 44 which are contained in the individual data sets 3, 4. The energy-based machine learning model 2 is trained by means of the training dataset in such a way that it predicts the masked regions 35, 45 from the calibration result 34, 44 obtained by the simulation on the basis of the calibration parameters 30- 33, 40- 43 used in the simulation and controls the prediction of the regions on the basis of the training dataset by self-supervised learning and converts the results of the control during further training, so that the calibration result 34, 44 can be deduced from both a predefined calibration result 34, 44 to the corresponding calibration parameters 30- 33, 40- 43 and from the predefined calibration parameters 30- 33, 40- 43 to the learned calibration result 34, 44. In this case, the energy-based machine learning model 2 is trained in such a way that the energy of the energy function E (x, y, z) of the energy-based machine learning model 2 is minimized if the predefined calibration parameters 30- 33, 40- 43 are compatible with the calibration result 34, 44, and that the energy of the energy function E (x, y, z) is increased if the predefined calibration parameters 30- 33, 40- 43 are incompatible with the calibration result 34, 44. Referring to the illustration in FIG. 2, the energy-based machine learning model 2 must be trained such that the pixel representation on the right side must be compatible with the two partially masked pixel representations on the left side.The training method used may be either a contrasting or regularized method. In the contrastive method, it is important to use negative images. In this case, the energy-based machine learning model 2 is intended to learn to distinguish similar features from dissimilar features in the pixel representations of the training dataset. In the regularized learning method, a range of values of variables which could have influenced the calibration results 34, 44 but were not measured can be selected as a regularizer for the latent variables z.FIGS. 3 and 4 show the trained artificial intelligence 1 when used in productive operation for autonomous calibration of a control device of a motor vehicle, which can be, in particular, a control device of a drive apparatus of the motor vehicle.As can be seen in FIG. 3, after training, when used in productive operation, the energy-based machine learning model 2 of the artificial intelligence 1 is capable of predicting the calibration result 54 of a predefined calibration with predefined calibration parameters 50- 53.As can be seen in FIG. 4, moreover, after training, when used in productive operation, the energy-based machine learning model 2 of the artificial intelligence 1 is also capable of predicting the calibration parameters 50- 53 on the basis of a predefined calibration result 54.A variation of the variant inscriptions makes it possible to change the generated pixel representations and thus to generate new calibration results for a given calibration. By suitable regularization, it is possible to bring the values of the latent variables z within a certain range corresponding to variables from the real world that could have influenced calibration, such as regularization of the vehicle speed in an interval [0, 300 (km / h)], the ambient temperature between -40° C. and +60° C., etc.References included in the specificationThis list of documents cited by the applicant has been produced in an automated manner and is only included for the better information of the reader. The list is not part of the German patent application or utility model application. The DPMA does not take any adhesion for any faults or omissions.Cited Non-Patent LiteratureA Path Towards Autonomous Machine Intelligence version 0.9.2, 2022-06-27" from Yann LeCu (retrievable at: https: / / openreview.net / pdf?id=BZ5a1r-kVsf

[0018]

Claims

Computer-implemented method for training an artificial intelligence (1) in which an energy-based machine learning model (2) for calibrating a control device of a motor vehicle is implemented, comprising the steps of: - generating (100) a training dataset containing a plurality of individual datasets (3, 4), each of the individual datasets (3, 4) containing a plurality of calibration parameters (30-33, 40-43) and a corresponding calibration result (34, 44) in the form of pixel representations, wherein the individual datasets (3, 4) of the training dataset are generated automatically in simulation loops by changing the calibration parameters (30-33, 40-43) and subsequent simulation in a simulation environment of the control device and output of the calibration result (34, 44) and are stored in a non-volatile storage means, and - training (200) the energy-based, retrievable, or non-transitory storage means, Machine learning model (2) by self-supervised learning, wherein masked areas (35, 45) in the pixel representations of the calibration results (34, 44) are automatically generated in each of the individual datasets (3, 4) of the training dataset and the energy-based machine learning model (2) is trained by means of the training dataset in such a way that it predicts the masked areas (35, 45) from the simulated calibration result (34, 44) on the basis of the corresponding calibration parameters (30-33, 40-43) and the prediction of the masked areas (35, 45) is checked on the basis of the training dataset and converts the results of the control during the further training, so that the calibration result (34, 44) can be deduced from a given calibration result (34, 44) to the corresponding calibration parameters (30-33, 40-43) and from given calibration parameters (30-33, 40-43) to the calibration result (34, 44) by means of the learned prediction.The computer-implemented method of claim 1, characterized in that the energy-based machine learning model (2) is trained such that an energy of an energy function of the energy-based machine learning model (2) is minimized if the predetermined calibration parameters (30-33, 40-43) are compatible with the calibration result (34, 44), and that the energy of the energy function is increased if the predetermined calibration parameters (30-33, 40-43) are incompatible with the calibration result (34, 44).Computer-implemented method according to one of Claims 1 or 2, characterized in that the pixel representations of the calibration parameters (30-33, 40-43) and of the associated calibration result (34, 44) in the individual datasets (3, 4) are each represented on the same image.Computer-implemented method according to one of Claims 1 to 3, characterized in that a hardware-in-the-loop simulation environment or a software-in-the-loop simulation environment is used for generating the training dataset.Computer-implemented method according to one of Claims 1 to 4, characterized in that, when generating the training dataset, the calibration parameters (30-33, 40-43) are randomly modified.Computer-implemented method according to one of Claims 1 to 5, characterized in that an energy-based, machine learning model (2) is used which is based on an LVEBM architecture or an LVGEBM architecture or a JEPA architecture.Computer-implemented method according to one of Claims 1 to 6, characterized in that the energy-based, machine learning model (2) is trained by contrasting learning.Computer-implemented method according to one of Claims 1 to 6, characterized in that the energy-based, machine learning model (2) is trained by regularized learning.A system comprising an electronic data memory and a digital electronic data processing unit, wherein machine-readable instructions are stored in the data memory, wherein the data processing unit is configured to read out and execute the instructions, and wherein the data processing unit is configured to perform a method according to any one of claims 1 to 8 when the instructions are executed.Use of an artificial intelligence (1) trained by means of a computer-implemented method according to one of Claims 1 to 8 for autonomous calibration of a control device of a motor vehicle, in particular for autonomous calibration of a control device of a drive device of the motor vehicle.

Images