CALCULATION OF EXHAUST EMISSIONS OF A MOTOR VEHICLE

Abstract

A method for ascertaining emissions of a motor vehicle driven with the aid of an internal combustion engine in a practical driving operation. A machine learning system is trained to generate time curves of the operating variables with the aid of measured time curves of operating variables of the motor vehicle and/or of the internal combustion engine, and to then ascertain the emissions as a function of these generated time curves.

Claims

1-17. (canceled)

18. A method for ascertaining emissions of a motor vehicle driven using an internal combustion engine in a practical driving operation, the method comprising the following steps: training a machine learning system to generate time curves using measured time curves of the operating variables of the internal combustion engine and/or of the motor vehicle; and ascertaining the emissions as a function of the generated time curves generated by the machine learning system.

19. The method as recited in claim 18, wherein the machine learning system includes a first part which initially transforms the measured time curves into first variables, each of which characterizes latent variables, a space of latent variables having a reduced dimensionality, and the machine learning system includes a second part which generates, as a function of the latent variables, second variables, each of which characterizes the generated time curves of the operating variables.

20. The method as recited in claim 19, wherein the first variables, which characterize the latent variables, are the latent variables themselves, and the second part includes a parameterized Gaussian process model parameterized by third parameters, and the third parameters and the latent variables are adapted during the training of the machine learning system in such a way that a marginal probability of a reconstruction of the measured time curves is maximized below these latent variables.

21. The method as recited in claim 20, wherein the first part ncludes a neural network parameterized by fourth parameters, and the adaptation of the latent variables occurs during the training by adapting the fourth parameters.

22. The method as recited in claim 19, wherein the first part and the second part of the machine learning system form an autoencoder.

23. The method as recited in claim 19, wherein the first part and the second part of the machine learning system form a variational autoencoder.

24. The method as recited in claim 19, wherein the first variables, which characterize the latent variable, are the latent variables themselves, and the first part ascertains the latent variables from the measured time curves using a sparse dictionary learning method, which variables represent coefficients of the measured time curves in a representation as a linear combination of the dictionary learned using this method.

25. The method as recited in claim 18, wherein latent variables are predefined and the machine learning system generates the time curves of the operating variables as a function of the predefined latent variables, and the emissions are then ascertained as a function of the generated time curves.

26. The method as recited in claim 25, wherein the latent variables are ascertained using a method of statistical test planning.

27. The method as recited in claim 25, wherein a probability density distribution of the latent variables resulting as a function of the measured time curves is ascertained and the predefined latent variables are drawn as a random sample from the estimated probability density distribution.

28. The method as recited in claim 18, wherein the machine learning system includes a first part to which either the measured time curves of the operating variables or time curves of the operating variables generated by a second part of the machine learning system are fed, the first part being trained to decide whether it is fed a measured time curve of the operating variables or a generated time curve of the operating variables, and the second part being trained to generate the time curves of the operating variables as a function of randomly selected input variables.

29. The method as recited in claim 28, wherein the second part is trained to generate the time curves of the operating variables as the function of the randomly selected input variables in such a way that the first part is able to only poorly decide whether it is fed the measured time curve or the generated time curve of the operating variables.

30. The method as recited in claim 28, wherein randomly selected input variables are predefined and the machine learning system generates the time curves of the operating variables as a function of the randomly selected input variables, and the emissions are then ascertained as a function of the generated time curves.

31. The method as recited in claim 30, wherein at least some of the randomly selected input variables are ascertained using a method of statistical test planning.

32. The method as recited in in claim 18, wherein the motor vehicle is controlled as a function of the ascertained emissions.

33. A non-transitory machine-readable memory medium on which is stored a computer program for ascertaining emissions of a motor vehicle driven using an internal combustion engine in a practical driving operation, the computer program, when executed by a computer, causing the computer to perform the following steps: training a machine learning system to generate time curves using measured time curves of the operating variables of the internal combustion engine and/or of the motor vehicle; and ascertaining the emissions as a function of the generated time curves generated by the machine learning system.

34. A computer configured to ascertain emissions of a motor vehicle driven using an internal combustion engine in a practical driving operation, the computer configured to: train a machine learning system to generate time curves using measured time curves of the operating variables of the internal combustion engine and/or of the motor vehicle; and ascertain the emissions as a function of the generated time curves generated by the machine learning system.

Description

BRIEF DESCRIPTION OF THE DRAWINGS

[0083] FIG. 1 shows a structure of a motor vehicle.

[0084] FIG. 2 shows a device for ascertaining the emissions.

[0085] FIG. 3 shows by way of example, a structure of a device for training the machine learning system.

[0086] FIG. 4 shows by way of example a use of the machine learning system for ascertaining emissions.

[0087] FIG. 5 shows an exemplary structure of the machine learning system.

[0088] FIG. 6 shows an alternative exemplary structure of the machine learning system.

DETAILED DESCRIPTION OF EXAMPLE EMBODIMENTS

[0089] FIG. 1 shows by way of example a structure of a motor vehicle 10. The motor vehicle is driven by an internal combustion engine 20. Combustion products forming during the operation of internal combustion engine 20 are conducted through an exhaust system 30, which includes, in particular, an exhaust gas purification system 40, for example, a catalytic converter. At the end of exhaust system 30, emissions 50, in particular, nitrogen oxide, soot particles and carbon dioxide, escape into the environment.

[0090] FIG. 2 shows by way of example a structure of a device 200, with which emissions 50 of motor vehicle 10 may be ascertained in a practical driving operation. Device 200 in the exemplary embodiment is a computer, which includes a machine-readable memory medium 210, on which a computer program 220 is stored. This computer program is configured to carry out the method according to the present invention, i.e., computer program 220 includes instructions which prompt computer 200 to carry out the method according to the present invention when computer program 220 is executed on computer 200.

[0091] FIG. 3 shows by way of example a structure of a device for training machine learning system (M). Machine learning system (M) is fed measured time curves (x) of operating variables of motor vehicle 10 and/or of internal combustion engine 20 as input variables. These measured time curves need not originate from the same motor vehicle and may, for example, be stored in a database. As a function of parameters (v, , , , , Y), which are stored on machine-readable memory 210, machine learning system (M) generates from these an output variable, namely, either time curves (x) of the operating variable or a discrimination result (d). Measured time curves (x) and generated time curves (x) or, alternatively, discrimination result (d) are fed to a learning unit (L), which adapts parameters (v, , , , , Y), for example, with the aid of a gradient descent method, in such a way that a cost function is optimized.

[0092] FIG. 4 shows by way of example, a use of machine learning system (M) for ascertaining emissions (e). Machine learning system (M) generates time curves (x) of operating variables of motor vehicle 10 and/or of internal combustion engine 20 as a function of parameters (v, , , ). These parameters are fed to a block (E), which ascertains therefrom associated emissions (e) with the aid of a mathematical model or of real measurements in motor vehicle 10.

[0093] FIG. 5 shows in greater detail an exemplary structure of machine learning system (M). FIG. 5a shows the structure as it may be used during the training. Machine learning system (M) includes a coder (K) and a decoder (D). Coder (K) ascertains variables (z, q.sub.(z|x)), which characterize latent variables (z), from measured time curves (x) and parameters (v, ) fed to it, and variables are fed in turn to a decoder (D). In addition to latent variables (z), decoder (d) may also be fed further variables (not shown). Decoder (D) decodes generated time curves (x) from these variables (z, q.sub.0(z|x) and as a function of parameters (, ) and, if necessary, of the additional variables.

[0094] FIG. 5b shows the structure as it may be used during the generation of generated time curves (x). A block (S) generates latent variables (z) according to a predefinable distribution. A probability density is ascertained, for example, with the aid of a density estimator as a function of ascertained latent variables z as shown in FIG. 5a, from which block (S) then randomly draws a sample. These generated latent variables (z) are then fed to decoder (D), which generates generated time curves (x), in this case, as a function of parameters (, ).

[0095] Coder (K) and decoder (D) in this case may form an autoencoder, for example, or a variational autoencoder, or implement a sparse dictionary learning.

[0096] It is also possible in this case that decoder (D) includes a Gaussian process. Then it is possible either that coder (K) includes a neural network, which ascertains latent variables (z) as a function of parameters (v), these parameters (v) also being varied during the training in addition to parameters (), which characterize the Gaussian process, in such a way that a marginal probability (p(x|z)) of the reconstruction of measured time curves (x) is maximized below these latent variables (z). Or, it is possible that coder (K) is omitted and latent variables (z) are directly predefined, so that in addition to parameters (y), learning unit (L) also adapts these latent variables (z) in such a way that a cost function, which includes a reconstruction error between measured time curve (x) and associated curve (x) generated from the selected latent variables (z), is minimized.

[0097] FIG. 6 shows a more detailed alternative exemplary structure of machine learning system (M). FIG. 6a shows the structure as it may be used during the training. Machine learning system (M) includes a first block (U) and a second block (H). First block (U) is parameterized by parameters (Y), second block (H) is parameterized by parameters (). A random number generator (R) ascertains random numbers (or as is frequently common, pseudo-random numbers) (r) and feeds these to second block (H). Second block (H) may also be fed additional variables (not shown), which code characterizing parameters. Second block (H) generates in each case a generated time curve (x) from random numbers (r) and, if necessary, from the additional variables as a function of parameters ().

[0098] Such generated time curves (x) and measured time curves (x) are fed alternatingly to first block (U), i.e., first block (U) is fed either a generated time curve (x) or a measured time curve (x). It is also possible that first block (U) is fed both these curves (x, x) if first block (U) has an internal selection mechanism (not shown), which in each case selects one of these two curves (x, x).

[0099] First block (U) is trained as shown in FIG. 3 by adapting parameters (Y) which determine its behavior, to be able to preferably readily distinguish whether the variable fed to it is a measured time curve (x) or a generated time curve (x). The information whether this classification of first block (U) is correct or incorrect is coded in discrimination result (d).

[0100] First block (U) and second block (H) are then mutually trained, parameters (Y) of first block (U) being trained that the classification of first block (U) is preferably often correct and parameters () of second block (H) being trained that the classification of first block (U) is preferably often incorrect.

[0101] FIG. 6b shows the corresponding structure as it may be used for generating generated time curves (x). Random number generator (R) generates random numbers or pseudo-random numbers (r), and second block (H) generates as a function thereof and, if necessary, as a function of the additional variables, generated time curves (x) with the aid of parameters () adapted during the training.