Use of artificial intelligence trained by a computer-implemented method for calibrating a control device of a motor vehicle

Abstract

Use of an artificial intelligence (1) trained by means of a computer-implemented method, in which an energy-based, machine learning model (2) is implemented for calibrating a control device of a motor vehicle, for an autonomous calibration of a control device of a motor vehicle, in particular for the autonomous calibration of a control device of a drive device of the motor vehicle, wherein the computer-implemented method comprises the steps: - generating (100) a training data set containing a plurality of individual data sets (3, 4), each of the individual data sets (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, the individual data sets (3, 4) of the training data set being 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 being stored in a retrievable manner in a non-volatile storage medium, and - Training (200) the energy-based machine learning model (2) by self-supervised learning, wherein in each of the individual data sets (3, 4) of the training data set, masked areas (35, 45) are automatically generated in the pixel representations of the calibration results (34, 44), and the energy-based machine learning model (2) is trained by means of the training data set in such a way that it predicts the masked areas (35, 45) from the simulated calibration result (34, 44) using the corresponding calibration parameters (30-33, 40-43), checks the prediction of the masked areas (35, 45) using the training data set, and implements the results of the check during further training, so that the learned prediction from a given calibration result (34, 44) can be used to predict the corresponding calibration parameters (30-33, 40-43) and from given calibration parameters (30-33, 40-43) can be used to determine the calibration result (34, 44),wherein the energy-based machine learning model (2) of the artificial intelligence (1) is capable, after training, of predicting the calibration result (54) of a predetermined calibration with predetermined calibration parameters (50-53) when used in productive operation, and wherein the energy-based machine learning model (2) of the artificial intelligence (1) is capable, after training, of predicting the calibration parameters (50-53) based on a predetermined calibration result (54) when used in productive operation.

Description

[0001] The present invention relates to a use of an artificial intelligence trained by means of a computer-implemented method in which an energy-based machine learning model for calibrating a control device of a motor vehicle is implemented, for an autonomous calibration of a control device of a motor vehicle.

[0002] Control devices for drive systems of a motor vehicle, especially for internal combustion engines or electric drive systems, must first be calibrated before they can be used in regular operation. For example, the operation of an electric drive system of a battery-electric motor vehicle is controlled by a corresponding control device. For this purpose, the control device calculates values ​​that depend on parameters during the operation of the electric drive system.These parameters can include, in particular, an electrical voltage applied to the electric 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 storage device that supplies the drive device with electrical energy. The calculated values ​​can include, for example, an electrical current supplied to the drive device. For the most efficient and precise control of the drive device, it is advantageous if the calculated values ​​differ as little as possible from corresponding actual values. Therefore, calibration parameters of the control device are adjusted before regular operation of the motor vehicle.

[0003] 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 using a display device. This visualization enables a user, particularly an engineer, to evaluate the calibration result by adding appropriate annotations to the calibration results. Currently, the potential impact of missing dataset variables on the analysis resulting from latent variables (not recorded in the dataset) in the feature space, which can particularly be intervals for physical parameters such as vehicle weight, is not taken into account.

[0004] The control logic implemented in the control device is typically extremely complex, as there are many parameters that must be calibrated. It is typically particularly difficult to simulate the expected calibration results, as the control logic of the control device is often a "black box" unknown to the user. Therefore, it is desirable to be able to answer the question of which calibration interventions could have led to certain calibration results.

[0005] In the publication DAWID, Anna; LECUN, Yann: "Introduction to latent variable energy-based models: a path towards autonomous machine intelligence" Preprint [v1] Mon, 5 Jun 2023 03:55:26 UTC 2023-06-06. pp. 1-29. DOI, available at: https: / / doi.org / 10.48550 / arXiv.2306.02572, denoising autoencoders and a JEPA architecture are disclosed as examples of energy-based machine learning models that enable self-supervised learning.

[0006] The object of the present invention is to provide a use of an artificial intelligence trained by means of a computer-implemented method, in which an energy-based, machine learning model for calibrating a control device of a motor vehicle is implemented, for an autonomous calibration of a control device of a motor vehicle, wherein the artificial intelligence is trained by means of the computer-implemented method in such a way that it can autonomously predict both the calibration result for a predetermined calibration of a control device of a motor vehicle and the calibration for a predetermined calibration result without human annotations and only on the basis of pixel analyses.

[0007] The solution to this problem is provided by the use of an artificial intelligence trained by means of a computer-implemented method, in which an energy-based, machine learning model for calibrating a control device of a motor vehicle is implemented, for an autonomous calibration of a control device of a motor vehicle with the features of claim 1. The subclaims relate to advantageous developments of the invention.

[0008] The invention proposes the use of an artificial intelligence trained by means of a computer-implemented method in which an energy-based machine learning model for calibrating a control device of a motor vehicle is implemented, for an autonomous calibration of a control device of a motor vehicle, in particular for the autonomous calibration of a control device of a drive device of the motor vehicle, wherein the computer-implemented method comprises the steps: - generating a training data set containing a plurality of individual data sets, each of the individual data sets containing a plurality of calibration parameters and a corresponding calibration result in the form of pixel representations, the individual data sets of the training data set being generated automatically in simulation loops by changing the calibration parameters and subsequent simulation in a simulation environment of the control device and output of the calibration result and being stored in a non-volatile storage medium in a retrievable manner, and - Training the energy-based machine learning model by self-supervised learning, wherein in each of the individual data sets of the training data set, masked areas are generated automatically in the pixel representations of the calibration results, in particular by an image processing method, and the energy-based machine learning model is trained by means of the training data set in such a way that it predicts the masked areas from the simulated calibration result based on the corresponding calibration parameters and checks the prediction of the masked areas based on the training data set and implements the results of the check during further training, so that the learned prediction can be used to infer the corresponding calibration parameters from a given calibration result and the calibration result from given calibration parameters, wherein the energy-based,machine learning model of artificial intelligence, after training, is capable of predicting the calibration result of a given calibration with given calibration parameters when used in production, and wherein the energy-based machine learning model of artificial intelligence, after training, is capable of predicting the calibration parameters based on a given calibration result when used in production.

[0009] The energy-based machine learning model, which after training can autonomously assign a specific calibration to a given calibration result and a specific calibration result to a given calibration, is trained in the method according to the invention through self-supervised learning based only on the pixel representations, using a masking technique to generate the masked areas in the pixel representations of the calibration results. Thus, no annotations are made in the training dataset by a user, especially an engineer, during training.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 when the specified calibration parameters are compatible with the calibration result, and such that the energy of the energy function is increased when the specified calibration parameters are incompatible with the calibration result. After the completion of training of the energy-based machine learning model, the artificial intelligence functions in two directions, thus enabling the prediction of calibration results based on specified calibration parameters (analysis) and, in addition, the prediction of calibration parameters based on the specified calibration result (analysis and calibration).

[0010] The calibration parameters can be provided, in particular, as two- or three-dimensional calibration diagrams or calibration maps. The calibration parameters, in particular the calibration maps, and the associated calibration results are stored together as individual data sets in the non-volatile memory device, preferably in the form of a pixel matrix, and are retrievably stored. For each calibration, the respective calibration parameters and the associated calibration results are thus available as individual data sets in the training data set.

[0011] In an advantageous embodiment, it is proposed that the pixel representations of the calibration parameters and the corresponding calibration result in the individual data sets are each displayed on the same image. This can simplify the self-supervised learning of the energy-based machine learning model.

[0012] In a preferred embodiment, it is possible to use a hardware-in-the-loop simulation environment or a software-in-the-loop simulation environment to generate the training data set. In the hardware-in-the-loop simulation environment, a real (test) control device is used. In contrast, the functions of the control software of the control device are emulated in the software-in-the-loop simulation environment.

[0013] In one embodiment, it is proposed that the calibration parameters are randomly changed when generating the training data set.

[0014] 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”).

[0015] Latent variables in the feature space, often referred to as "hidden variables," are used to account for possible variables that influence the calibration task but were not captured in the individual data sets (either because they are unknown to the engineer or, for example, because there is not enough storage space to record them). In other words, latent variables are variables whose values ​​are unknown. However, if they were known, latent variables would facilitate the actual calibration task. Examples of latent variables include the vehicle weight, the ambient temperature, or the state of charge of a vehicle's traction battery.

[0016] As already mentioned, in one embodiment, the energy-based machine learning model can be based on an LVEBM architecture (LVEBM = "Latent Variable Energy Based Model"). Very simply, the LVEBM architecture is based on the idea of ​​assessing the degree of compatibility between a variable x (in this case, the calibration) and a variable y (in this case, the specified calibration result) using a latent variable z. The latent variable z can be viewed as the parameterization 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 the latent variable z.Using an inference procedure, a value for the latent variable z is determined for a pair of variables (x, y) that minimizes the energy of the energy function E (x, y, z).

[0017] In an alternative embodiment, it is also possible to implement the energy-based machine learning model as a (generative) LVGEBM architecture (LVGEBM = "Latent Variable Generative Energy Based Model"). This architecture is very similar to the LVEBM architecture. However, since the LVGEBM architecture is generative, it is also advantageously possible to generate alternative calibration results for given calibration parameters or alternative calibration parameters for a given calibration result.

[0018] In another alternative embodiment, the energy-based machine learning model may be based on a JEPA (Joint Embedding Predictive Architecture) architecture. Generally speaking, a JEPA architecture learns to predict the embeddings of a signal y (in this case, a representation of the calibration result) from a compatible signal x (in this case, a representation of the calibration) by using a predictor network conditioned with additional, particularly latent, variables z to facilitate the prediction. The goal here, too, is to determine the value for the latent variable z that minimizes the energy of an energy function E(x, y, z).

[0019] Further details and further explanations, especially on LVEBM architectures and JEPA architectures, can be found (with further references) for example in the online publication "A Path Towards Autonomous Machine Intelligence Version 0.9.2, 2022-06-27" by Yann LeCun (available at: https- / / openreview.net / pdf?id=BZ5a1r-kVsf).

[0020] In one embodiment, it is possible for the energy-based machine learning model to be trained using contrastive learning. The energy-based machine learning model should learn to distinguish similar from dissimilar features in the pixel representations of the training dataset.

[0021] In an alternative embodiment, it is proposed that the energy-based machine learning model be trained using regularized learning. In this regularized learning method, a range of values ​​of variables that could have influenced the calibration results but were not measured can be chosen as regularizers for the latent variables.

[0022] 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 upon execution of the instructions.

[0023] The method presented here for training artificial intelligence and the use of the trained artificial intelligence in subsequent productive operation offer the following advantages in particular: Calibration of the motor vehicle's control system is possible without information about the software control logic itself, which is usually a "black box".

[0024] No user intervention, such as adding annotations to the calibration parameters and calibration results, is required, resulting in time and cost savings both during training and during the productive operation of the artificial intelligence.

[0025] The energy-based machine learning model learns a feature representation of the calibration problem, whereby not every detail (i.e. not every single pixel in the respective pixel representation) has to be learned / captured in order to assign a calibration to a given calibration result or a calibration result to a given calibration represented by the calibration parameters.

[0026] A potential additional benefit is that latent variables are considered during training and some of them are regularized with values ​​that also make sense in the real world. This can enable the generation of new, realistic calibrations / calibration results for given calibration results / calibrations.

[0027] Further features and advantages of the present invention will become clear from the following description of a preferred embodiment with reference to the accompanying drawings. Fig. 1 is a schematic diagram illustrating the basic process of a computer-implemented method for training an artificial intelligence, Fig. 2 a schematic diagram illustrating the training of an energy-based machine learning model of artificial intelligence, Fig. 3 is a schematic diagram illustrating the use of the trained artificial intelligence according to a first variant, Fig. 4 a schematic diagram illustrating the use of the trained artificial intelligence according to a second variant.

[0028] With reference to Fig. 1 and Fig. 2, the basic process of a method for training an artificial intelligence 1 will be illustrated below. Fig. 3 and Fig. 4 then show the use of the trained artificial intelligence 1 in later productive operation.

[0029] The artificial intelligence 1 implements an energy-based machine learning model 2 for calibrating a control device of a drive device of a motor vehicle. This drive device can be, for example, an internal combustion engine or an electric motor of the motor vehicle. Details of the energy-based machine learning model 2 are explained in more detail below.

[0030] In a first step 100, a training data set is first generated, which is later used for training the artificial intelligence 1. This training data set contains several individual data sets 3, 4, of which Fig. 2, two examples are schematically shown on the left. 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 data sets 3, 4 can each be displayed on the same image. This can simplify self-supervised learning of the energy-based machine learning model 2, which is explained in more detail below.

[0031] Either a hardware-in-the-loop simulation environment or a software-in-the-loop simulation environment can be used to generate the training data set. In the hardware-in-the-loop simulation environment, a real (test) control device is used. In contrast, the functions of the control software of the control device are emulated in the software-in-the-loop simulation environment. The training data set is created automatically in simulation loops by changing the calibration parameters 30-33, 40-43 and subsequently simulating in the simulation environment and outputting the calibration results 34, 44. Preferably, the calibration parameters 30-33, 40-43 are changed randomly in the simulation loops. The calibration parameters 30-33, 40-43 can, in particular, be present as two- or three-dimensional calibration diagrams or calibration maps.The calibration parameters 30-33, 40-43, in particular the calibration maps, as well as 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 so that they can be used in the subsequent training of the energy-based machine learning model 2.

[0032] After the training dataset with the plurality of individual datasets 3, 4 has been created, the next step 200 involves the actual training of the energy-based machine learning model 2. Energy-based machine learning models 2 having different architectures can be used. In all architectures, latent variables in the feature space are preferably used to account for possible variables that have an influence on the calibration task but were not recorded in the individual datasets 3, 4 (either because they are unknown to the engineer or because, for example, there is not enough storage space available to record them). In other words, latent variables are variables whose values ​​are unknown. They are therefore often referred to as "hidden variables." If latent variables were known, they would facilitate the actual calibration task.Examples of latent variables are the ambient temperature or the state of charge of a vehicle’s traction battery.

[0033] 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 idea of ​​assessing the degree of compatibility between a variable x (in this case, the calibration) and a variable y (in this case, the calibration result) using (at least) one latent variable z. The latent variable z can be viewed as a parameterization 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 the latent variable z.Using an inference procedure, a value for the latent variable z is determined for a pair of variables x, y, which minimizes the energy of the energy function E (x, y, z).

[0034] 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"). In general, such an architecture learns to predict the embeddings of a signal y (in this case, the calibration result) from a compatible signal x (in this case, the calibration) by using a predictor network conditioned by additional, in particular latent, variables z to facilitate the prediction. The goal here, too, is to determine the value for the latent variable z that minimizes the energy of an energy function E(x, y, z).

[0035] In a further embodiment, it is also possible for the energy-based machine learning model 2 to be implemented as a generative LVGEBM architecture (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.

[0036] The training of the energy-based machine learning model 2 is carried out by self-supervised learning, using a masking technique by means of which at least one masked area 35, 45 is initially generated in the pixel representations of the calibration results 34, 44 contained in the individual data sets 3, 4, in an automated manner, in particular by an image processing method.The energy-based machine learning model 2 is trained using the training data set in such a way that it predicts the masked areas 35, 45 from the calibration result 34, 44 obtained by the simulation using the calibration parameters 30-33, 40-43 used in the simulation and controls the prediction of the areas using the training data set using self-supervised learning and implements the results of the control during further training, so that the learned prediction can be used to infer the corresponding calibration parameters 30-33, 40-43 from a given calibration result 34, 44 as well as from the given calibration parameters 30-33, 40-43 to the calibration result 34, 44.The energy-based machine learning model 2 is trained such that the energy of the energy function E(x,y,z) of the energy-based machine learning model 2 is minimized when the specified calibration parameters 30-33, 40-43 are compatible with the calibration result 34, 44, and such that the energy of the energy function E(x,y,z) is increased when the specified calibration parameters 30-33, 40-43 are incompatible with the calibration result 34, 44. Reference is made to the illustration in . Fig. 2, this means that the energy-based machine learning model 2 must be trained in such a way that the pixel representation on the right side must be compatible with the two partially masked pixel representations on the left side.

[0037] The training method used can be either a contrastive or a regularized method. In the contrastive method, it is important to use negative images. The energy-based machine learning model 2 should learn to distinguish similar from dissimilar features in the pixel representations of the training dataset. In the regularized learning method, a range of variables that could have influenced the calibration results 34, 44 but were not measured can be chosen as regularizers for the latent variables z.

[0038] Fig. 3 and Fig. 4 show the trained artificial intelligence 1 being used in productive operation for the autonomous calibration of a control device of a motor vehicle, which may in particular be a control device of a drive device of the motor vehicle.

[0039] As in Fig. 3, the energy-based machine learning model 2 of the artificial intelligence 1, after training, is able to predict the calibration result 54 of a given calibration with given calibration parameters 50-53 when used in production.

[0040] As furthermore in Fig. 4, the energy-based machine learning model 2 of the artificial intelligence 1 is, after training, also able to predict the calibration parameters 50-53 based on a given calibration result 54 when used in production.

[0041] Varying the variant labels allows the generated pixel representations to be modified, thus generating new calibration results for a given calibration. Through appropriate regularization, it is possible to bring the values ​​of the latent variables z into a specific range corresponding to real-world variables that may have influenced the calibration, such as regularizing the vehicle speed in an interval [0, 300 (km / h)], the ambient temperature between -40°C and +60°C, etc.

Claims

[1] Use of an artificial intelligence (1) trained by means of a computer-implemented method, in which an energy-based, machine learning model (2) is implemented for calibrating a control device of a motor vehicle, for an autonomous calibration of a control device of a motor vehicle, in particular for the autonomous calibration of a control device of a drive device of the motor vehicle, wherein the computer-implemented method comprises the steps: - generating (100) a training data set containing a plurality of individual data sets (3, 4), each of the individual data sets (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, the individual data sets (3, 4) of the training data set being 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 being stored in a retrievable manner in a non-volatile storage medium, and - Training (200) the energy-based machine learning model (2) by self-supervised learning, wherein in each of the individual data sets (3, 4) of the training data set, masked areas (35, 45) are automatically generated in the pixel representations of the calibration results (34, 44), and the energy-based machine learning model (2) is trained by means of the training data set in such a way that it predicts the masked areas (35, 45) from the simulated calibration result (34, 44) using the corresponding calibration parameters (30-33, 40-43), checks the prediction of the masked areas (35, 45) using the training data set, and implements the results of the check during further training, so that the learned prediction from a given calibration result (34, 44) can be used to predict the corresponding calibration parameters (30-33, 40-43) and from given calibration parameters (30-33, 40-43) can be used to determine the calibration result (34, 44),wherein the energy-based machine learning model (2) of the artificial intelligence (1) is capable, after training, of predicting the calibration result (54) of a predetermined calibration with predetermined calibration parameters (50-53) when used in productive operation, and wherein the energy-based machine learning model (2) of the artificial intelligence (1) is capable, after training, of predicting the calibration parameters (50-53) based on a predetermined calibration result (54) when used in productive operation. [2] Computer-implemented method according to claim 1, characterized bythat 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). [3] Computer-implemented method according to one of claims 1 or 2 characterized by that the pixel representations of the calibration parameters (30-33, 40-43) and the associated calibration result (34, 44) in the individual data sets (3, 4) are each displayed on the same image. [4] Computer-implemented method according to one of claims 1 to 3, characterized bythat a hardware-in-the-loop simulation environment or a software-in-the-loop simulation environment is used to generate the training dataset. [5] Computer-implemented method according to one of claims 1 to 4, characterized by that when generating the training dataset the calibration parameters (30-33, 40-43) are changed randomly. [6] Computer-implemented method according to one of claims 1 to 5, characterized by that an energy-based machine learning model (2) is used that is based on an LVEBM architecture or an LVGEBM architecture or a JEPA architecture. [7] Computer-implemented method according to one of claims 1 to 6, characterized by that the energy-based machine learning model (2) is trained by contrastive learning. [8] Computer-implemented method according to one of claims 1 to 6, characterized bythat the energy-based machine learning model (2) is trained by regularized learning. [9] 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 carry out a method according to one of claims 1 to 8 when executing the instructions.

Images