EP3698036A1

EP3698036A1 - Calculation of exhaust emissions from a motor vehicle

Info

Publication number: EP3698036A1
Application number: EP18793567.1A
Authority: EP
Inventors: Heiner Markert; Martin Schiegg; Stefan ANGERMAIER
Original assignee: Robert Bosch GmbH
Current assignee: Robert Bosch GmbH
Priority date: 2017-10-16
Filing date: 2018-10-09
Publication date: 2020-08-26
Also published as: US11078857B2; WO2019076685A1; JP7119101B2; KR20200066348A; KR102538418B1; US20200240346A1; CN111201374B; DE102017218476A1; JP2020537087A; CN111201374A

Abstract

The invention relates to a method for determining emissions (50) from a motor vehicle (10) driven by means of an internal combustion engine (20) during actual driving (real driving emissions), in which method a machine learning system is trained, using measured time curves of operating variables of the motor vehicle (10) and/or of the internal combustion engine (20), to generate time curves of the operating variables and then to determine the emissions depending on these generated time curves.

Description

Description title

CALCULATION OF EXHAUST EMISSIONS OF A MOTOR VEHICLE

Technical area

The invention relates to a method and an apparatus for determining emissions, a computer program, and a machine-readable storage medium.

State of the art

From DE 10 2009 028 374 AI a method for a diagnosis of an internal combustion engine is known, which is limited by the fact that only certain load conditions of the internal combustion engine can be approached on a test bed. Furthermore, it is known from this document that starting certain operating points can only be achieved by test drives, which leads to long diagnostic times.

Advantage of the invention

The method with the features of independent claim 1 has the advantage that particularly fast and easy emissions of a driven by an internal combustion engine motor vehicle in a practical driving operation can be determined. Advantageous developments are the subject of the independent claims.

Disclosure of the invention

In some countries, the legislation provides for the approval of new motor vehicles powered by an internal combustion engine depending on to make the emissions that result in practical driving. For this purpose, the English term real driving emissions is also common. Such motor vehicles include, for example, those that are driven exclusively by an internal combustion engine, but also those with a hybridized powertrain.

For this purpose, it is provided that an inspector with the motor vehicle denies a driving cycle or several driving cycles and the resulting emissions are measured. The approval of the motor vehicle is then dependent on these measured emissions. The driving cycle can be freely selected within wide limits by the examiner. A typical duration of a drive cycle may be 90-120 min, for example.

For motor vehicle manufacturers, this raises the challenge in the development of motor vehicles of having to foresee at an early stage in the development process of a new motor vehicle whether or not the emissions of this motor vehicle remain within the legally prescribed limits in every permissible driving cycle.

It is therefore essential to provide a device that can safely predict the anticipated emissions of the motor vehicle even in the development stage of a motor vehicle in order to carry out changes to the motor vehicle in the event of an anticipated exceeding of limit values. Such an estimation based solely on measurements on a test bench or in a moving motor vehicle is extremely complicated because of the great variety of conceivable driving cycles.

In a first aspect, the invention therefore relates to a method for determining emissions of a motor vehicle driven by an internal combustion engine in a practical driving operation, wherein a machine learning system is trained thereon by means of measured time profiles (x) of operating variables of the internal combustion engine and / or of the motor vehicle Gradients (χ ') to generate operating variables of the internal combustion engine and / or the motor vehicle, and then to determine the emissions depending on these generated time courses (χ ').

That it is possible to measure only actual time histories of the operational quantities, each having their own characteristics and then having realistic timings generated by the machine learning system having a plurality of these characteristics.

The operational quantities may preferably comprise one, some or all of the sizes characterizing the following size:

an accelerator pedal position of the motor vehicle

a brake pedal position of the motor vehicle

a position of a clutch of a transmission of the motor vehicle

a gear of the transmission

a speed of the motor vehicle

a driving resistance of the motor vehicle

a tensile force of the internal combustion engine

a tensile force of an electromotive drive of the motor vehicle, a rotational speed of the internal combustion engine

a per unit time sucked air mass of the internal combustion engine, a pressure in a suction pipe of the internal combustion engine

a lot of high-pressure AG R

a lot of low-pressure AG R

a time of closing the intake valve

a time of opening the exhaust valve

a maximum valve lift of the intake valve

a maximum valve lift of the exhaust valve

a position of the system for changing the compression of the internal combustion engine

a fuel amount of injections of the internal combustion engine

an injection time of the injections a pressure in a high-pressure fuel storage, as it is also known in German, for example, the term common rail, the internal combustion engine

a coolant temperature of the internal combustion engine

a temperature in the intake system of the internal combustion engine.

In addition to the time profiles of the operating variables, it is also possible, depending on parameters which characterize the internal combustion engine and / or the motor vehicle in the cycle to be completed, to train time profiles (χ ') of operating variables.

These characterizing parameters may include one, some or all of the following quantities:

a mass of the motor vehicle

a gear ratio of the motor vehicle

a maximum drive power of a drive system of the motor vehicle maximum torque of the drive system

a kind of gearbox

a fuel

a specification of hybridization

a type designation of the engine

a type designation of the vehicle

a characterization of the driven route, for example GPS records, ambient temperature, ambient pressure, etc.

The injections may in particular be multiple injections, for example a main injection, pre-injection and post-injection. In such multiple injections, those of the motor vehicle and / or the internal combustion engine preferably include quantities which characterize each fuel quantity and injection time of each of the partial injections.

The use of a machine learning system is important because the number of possible time courses is very large. With a time course of lh duration and a sampling rate of, for example, 100 ms results in 15 operating variables of the motor vehicle and / or the internal combustion engine, a 15 x 36,000 = 540,000-dimensional space in which possible temporal courses lie.

It is obvious to a person skilled in the art that only a relatively small number of points in this very high-dimensional space correspond to actual possible time profiles of the operating variables of the motor vehicle and / or the internal combustion engine, ie are realistic. The problem of selecting these realistic time courses is solved by the method having the features of independent claim 1.

In a further development, it can be provided that the machine learning system comprises a first part - an encoder - which first transforms the measured time courses (x) into first variables (z, qe (z | x)), which respectively contain latent variables (eg ), wherein a space of latent variables has a reduced dimensionality compared to the space of the measured time courses,

and wherein the machine learning system comprises a second part - a decoder - which generates second quantities (χ ', ρφ (χ' | ζ)) depending on latent variables (z), each of which generates the generated time profiles (χ ') of the operating variables characterize the motor vehicle and / or the internal combustion engine.

By mapping the actually measured trajectories to a low-dimensional space, the machine learning system is trained to extract the essential features, the realistic temporal courses - ie especially those courses which are contained in the training data of the measured time courses. The nature of realistic trajectories is therefore coded in parameters that characterize the coder or the decoder.

It is also possible that the coder transforms the measured time courses (x) not only into latent variables (z), but also into additional variables depending on the characterizing parameters.

In a particularly advantageous development, it can be provided here that the first variables which characterize the latent variables (z) are the latent variables (z) themselves and wherein the second part is one of third parameters (γ) parameterized Gaussian process model, and the third parameters (γ) and the latent variables (z) are adapted during training of the machine learning system such that a marginal probability (p (x | z)) of the reconstruction of the measured time courses (x) is maximized below these latent variables (z).

The use of the Gaussian process model has the advantage that a probabilistic statement about whether a respective second variable (z) corresponds to a course having the same characteristics as the measured courses (x) or not is possible. This makes it possible to judge for any test cycles based on the corresponding associated courses of operating variables of the motor vehicle and / or the internal combustion engine, whether these are relevant or not in view of the measured courses that have entered the training phase.

In a further development of this aspect, provision may be made for the first part to comprise a function parametrized by the fourth parameter (v), such as a neural network, for example, and adaptation of the latent variables (z) during training by adaptation of the fourth parameter (v).

This is a particularly efficient way of mapping the measured courses (x) of the operating variables of the motor vehicle and / or of the internal combustion engine to latent variables (z). In an alternative particularly advantageous development, it can be provided that the first part and the second part of the machine learning system form an auto-encoder.

An auto-encoder means that the first quantities that characterize the latent variables (z) are the latent variables themselves, and the second quantities that characterize the generated timings are the generated timings themselves. The first part and the second part may be given, for example, by neural networks. This method has the advantage that the training of the machine learning system is particularly simple. In particular, parameters which parameterize the auto-decoder can be adapted in such a way that a cost function which includes a reconstruction error, for example a standard | x-x '| ^{2 is characterized by} a difference between the measured time profiles and the time profiles (χ ') generated therefrom by means of the auto-encoder.

In a further alternative particularly advantageous development, it can be provided that the first part and the second part of the machine learning system form a two-part auto-encoder. Variation Autoencoder are known in German under the English-language term \ l anational autencoder.

This means that the first quantities, which respectively characterize the latent variables (z), are a first probability distribution (q _e (z | x)) parameterized according to first parameters (Θ) and the second variables which characterize the generated time histories each one of second parameters (φ) parameterized second probability distribution (ρ _φ (χ '| ζ)) are. The training of the machine learning system then means that the first parameters (Θ) and the second parameters (φ) are varied in such a way that a cost function is minimized. The cost function advantageously comprises a reconstruction error of a generated reconstruction of the measured time courses and a Kullback-Leibler divergence KL [q (z) | | J \ f] between the first probability distributions (q _e (z | x)) and a normal distribution K normalized to mean 0 and variance 1.

As a first probability distribution (q _e (z | x)) and / or as a second probability distribution (ρψ (χ '| ζ)), a parameterizable distribution function, for example a normal distribution, can be used in each case. The first part and / or the second part can then each comprise, for example, a neural network which, depending on the input variables supplied to it, determines parameters which parameterize this parameterizable distribution function. The advantage of using the multi-part autocoder is that the first probability distribution (q _e (z | x)) allows a probabilistic statement as to whether or not a respective second quantity (x) has the same characteristics as the measured gradients.

In yet another alternative particularly advantageous development, it can be provided that the first variables which characterize the latent variables (z) are the latent variables (z) themselves and the first part from measured temporal progressions (x) by means of a method of sparse dictionary learning The sparse variables (z) determine which coefficients of the respective measured temporal courses (x) in the representation represent a linear combination of the dictionary learned by means of this method. In this way, the space of the latent variables (z) can be reduced particularly effectively.

In an advantageous development of the aforementioned aspects, it can be provided that latent variables (z) are specified and the machine learning system generates time profiles (χ ') of operating variables of the motor vehicle and / or the internal combustion engine as a function of these predetermined latent variables (z), and then the emissions are determined depending on these generated time courses.

That The machine learning system is initially trained by means of measured time courses to be able to generate realistic temporal courses. By specifying the latent variables, temporal courses are then generated, to which then, e.g. be determined with a suitable mathematical model such as a machine learning method or a physico-chemical model, emissions. It may also be provided to measure the emissions on a real system in order to train, for example, said machine learning method.

In a development of this aspect, it can be provided that the at least some, preferably all, of the latent variables (z) will be determined by means of a method of statistical experimental design. This is particularly good if, depending on the emissions determined, the mathematical model used to determine the emissions should be adapted to actual measured emission values. This makes it possible to ensure that the widest possible areas of the latent variable space are explored as efficiently as possible.

In an alternative development of this aspect, provision may be made for a probability density distribution of the latent variables (z) which is determined as a function of the measured time courses to be determined and the predetermined latent variables (z) to be drawn as a random sample from this estimated probability density distribution.

In this way, it can be achieved that the determined emissions are particularly representative of the emissions occurring in the real operation of the motor vehicle. Thus, the emissions occurring in real operation of the motor vehicle can be estimated particularly accurately.

It is now also possible to specify the additional variables. This is particularly advantageous if time gradients (χ ') are to be generated, which should be based on at least one predefinable property. For example, temporal profiles (χ ') can thus be generated, which are generated in a limited manner to the above-mentioned characterizing parameters, that is, for example, vehicle types or geographical locations, by specifying these characterizing parameters as additional variables.

In order to train such a machine learning system, the additional parameters which are to code the predefinable properties of the training course (χ ') generated during training are set to the true value during training (since this characteristic of the time course is known at training times) ,

In another aspect of the invention can be provided that the machine learning system comprises a first part - a discriminator - which either measured time courses of operating variables of the motor vehicle and / or the internal combustion engine or from a second part of the machine learning system - a generator - generated time profiles of the motor vehicle and / or the internal combustion engine, are fed

wherein the first part is trained to be able to decide as well as possible whether a measured or a generated time course of operating variables of the motor vehicle and / or the internal combustion engine is supplied, wherein the second part is trained thereon, depending on randomly selected input variables temporal courses of operating variables of the motor vehicle and / or the internal combustion engine as possible to generate such that the first part can decide as bad as possible, whether a measured or a generated time course of operating variables of the motor vehicle and / or the internal combustion engine is supplied.

This method has the advantage that the thus generated courses of operating variables of the motor vehicle and / or the internal combustion engine are particularly realistic.

The meaning of the word "random" may include that the variables thus selected are determined by means of a true random number generator or by means of a pseudo-random number generator.

The training of the first part and the second part can advantageously be carried out alternately to ensure that the training is as effective as possible. In this case, the training can expediently be continued until the discriminator is no longer able to differentiate with predeterminable accuracy, whether the temporal courses supplied to him are measured or generated by the generator over time.

In a further development of this aspect, it can be provided that randomly selected input variables are specified and the machine learning system generates temporal profiles (χ ') of the operating variables of the motor vehicle and / or the internal combustion engine as a function of these predefined input variables, and then the emissions as a function of these generated temporal Gradients are determined. That is, the machine learning system is initially trained by means of measured time courses to be able to generate realistic temporal courses. By specifying the randomly selected input variables (z), actual time histories are then generated, for which emissions are then determined, for example using a suitable mathematical model such as, for example, a machine learning method or a physicochemical model. It may also be provided to measure the emissions on a real system in order to train, for example, said machine learning method. By specifying the additional variables, it is possible to generate profiles that correspond to the given properties.

In particular, it may be provided that the randomly selected input variables are determined by means of a method of statistical experimental design.

This is particularly good if, depending on the emissions determined, a mathematical model, with which the emissions are determined, should be adapted to actual measured emission values. This makes it possible to ensure that as far as possible areas of the space of the latent variables are explored as quickly as possible.

Finally, it is also possible to carry out the determination of the emissions by means of a trained model while the motor vehicle is being operated. It is then possible that the motor vehicle is driven depending on the detected emissions. The determination of the emissions by means of the machine learning system then acts as a virtual sensor, which saves costs, is particularly fail-safe and, above all, can be used predictively. It is then possible, for example, to throttle a power of the internal combustion engine or to change a fuel / air ratio depending on the determined emissions, for example when the motor vehicle enters a zone in which there are particular, in particular stricter, limits for the emissions. It is conceivable, for example, in inner cities.

In other aspects, the invention relates to a computer program which is adapted to carry out one of the aforementioned methods, if it is on a Computer is running, a machine-readable storage medium on which this computer program is stored (this storage medium may of course be spatially distributed, eg distributed in parallel execution across multiple computers), and a device, in particular a monitoring unit, which is set up, one of these methods execute (for example, by playing the aforementioned computer program).

Hereinafter, embodiments of the invention will be explained in more detail with reference to the accompanying drawings. In the drawings show:

Figure 1 shows a structure of a motor vehicle;

Figure 2 shows a device for determining the emissions;

FIG. 3 shows by way of example a construction of a device for training the machine learning system;

FIG. 4 shows by way of example a use of the machine learning system for determining emissions;

FIG. 5 shows an exemplary construction of the machine learning system;

FIG. 6 shows an alternative exemplary construction of the machine learning system.

Description of the embodiments

Figure 1 shows an example of a structure of a motor vehicle (10). The motor vehicle is driven by an internal combustion engine (20). Combustion products produced during operation of the internal combustion engine (20) are conducted through an exhaust tract (30), which in particular comprises an exhaust gas purification system (40), for example a catalytic converter. At the end of the exhaust tract (30) emissions (50) escape into the environment, in particular nitrogen oxides, soot particles and carbon dioxide. Figure 2 shows an example of a structure of a device (200), with the emissions (50) of the motor vehicle (10) can be determined in practical driving. In the exemplary embodiment, the device (200) is a computer which comprises a machine-readable storage medium (210) on which a computer program (220) is stored. This computer program is set up to carry out one of the methods according to the invention, ie the computer program (220) contains instructions which cause the computer (200) to carry out the method according to the invention when the computer program (220) is executed on the computer (200).

FIG. 3 shows by way of example a structure of a device for training the machine learning system (M). The machine learning system (M) are supplied as input variables of measured time profiles (x) of operating variables of the motor vehicle (10) and / or of the internal combustion engine (20). These measured time profiles do not have to originate from the same motor vehicle, and can for example be stored in a database. Depending on parameters (v, y, θ, φ "Γ, Y) stored on the machine-readable storage medium (210), the machine learning system (M) generates therefrom an output quantity, namely either time histories (χ ') of the operating quantities or a result of discrimination (d). The measured time courses (x) and the generated time courses (χ ') or alternatively the discrimination result (d) are fed to a learning unit (L), which uses, for example, by means of a gradient descent method the parameters (v, y, θ, φ, Γ, Y ) so that a cost function is optimized.

FIG. 4 shows by way of example a use of the machine learning system (M) for determining emissions (e). Depending on the parameters (v, y, θ, φ), the machine learning system (M) generates time profiles (χ ') of operating variables of the motor vehicle (10) and / or the internal combustion engine (20). These become one

Block (E) fed, which determines therefrom by means of a mathematical model or real measurements on the motor vehicle (10) the associated emissions (e). Figure 5 shows in more detail an exemplary construction of the machine learning system (M). FIG. 5a shows the structure how it can be used during training. The machine learning system (M) comprises an encoder (K) and a decoder (D). The coder (K) determines from the measured temporal progressions (x) and parameters (v, Θ) fed to it quantities (z, q _e (z | x)) which characterize the latent variables (z) and which in turn is a decoder (D) are supplied. In addition to the latent variables (z), the decoder (D) can also be supplied with further variables (not shown). The decoder (D) from these quantities (z, qe (z | x)) and, depending on parameters (γ, φ) and possibly the other variables, the generated time profiles (χ ').

FIG. 5b shows the structure as it can be used when generating generated time profiles (χ '). A block (S) generates latent variables (z) according to a predefinable distribution. For example, a probability density is determined by means of a density estimator as a function of the latent variables z determined as shown in FIG. 5a, from which the block (S) now randomly draws a random sample. These generated latent variables (z) are fed to the decoder (D), which generates the generated time profiles (χ ') as a function of parameters (γ, φ).

In this case, coders (K) and decoders (D) can, for example, form an auto-decoder, or implement a partial auto-encoder, or a sparse dictionary learning.

It is also possible here for the decoder (D) to include a Gaussian process. Then either it is possible for the encoder (K) to comprise a neural network which determines the latent variables (z) as a function of parameters (v), and in addition to the parameters (γ) characterizing the Gaussian process, this also applies during training Parameters (v) are varied such that a marginal probability (p (x | z)) of the reconstruction of the measured time courses (x) among these latent variables (z) is maximized. Or it is possible that the encoder (K) is omitted and latent variables (z) are specified directly, such that the learning unit (L) adjusts these latent variables (z) in addition to the parameters (γ) in such a way that a cost function which causes a reconstruction error between the measured time course (x) and the associated course generated from the selected latent variables (z) (χ ') is minimized.

Figure 6 shows in more detail an alternative exemplary structure of the machine learning system (M). FIG. 6a shows the structure how it can be used during training. The machine learning system (M) comprises a first block (U) and a second block (H). The first block (U) is parametrized by parameters (Y), the second block (H) by parameters (Γ). A random number generator

(R) determines random numbers (or, as is often the case, pseudo-random numbers) r and delivers them to the second block (H). Likewise, the second block (H) further variables (not shown) are supplied, the characterizing parameters encode the second block (H) generated from the random numbers (r) and possibly the other variables depending on the parameters (Γ) each one generated time course (χ ').

Thus generated timings (χ ') and measured timings (x) are alternately fed to the first block (U), i. the first block (U) is either a generated time course (χ ') or a measured time course (x) supplied. It is also possible for the first block (U) to be supplied with these two courses (x, x ') if the first block (U) has an internal selection mechanism (not shown), each of which has one of these two courses (x, x). x ').

As illustrated in FIG. 3, the first block (U) is trained by adapting its behavior determining parameter (Y) to be able to differentiate as well as possible whether the variable supplied to it is a measured time course (x) or a generated time course (χ '). ). The information as to whether this classification of the first block (U) is true or false is encoded in the discrimination result (d). The first block (U) and the second block (H) are now trained alternately, the parameters (Y) of the first block (U) are trained so that the classification of the first block (U) is often correct and the parameters (Γ ) of the second block (H) that the classification of the first block (U) is as often as possible wrong.

FIG. 6b shows the corresponding structure as it can be used to generate generated time profiles (χ '). The random number generator (R) generates random numbers or pseudo-random numbers (r), and the second block (H) generates the generated temporal courses (Γ ') depending on the further variables and, if necessary, on the parameters adapted in training (Γ). )

Claims

claims

1. A method for determining emissions (E) of a motor vehicle (10) driven by an internal combustion engine (20) in a practical driving mode (English: real driving emissions), wherein a machine learning system (M) by means of measured time courses (x) operating variables of the internal combustion engine (20) and / or of the motor vehicle (10) is trained to generate temporal profiles (χ ') of, and then to determine the emissions (e) as a function of these generated time profiles (χ').

The method of claim 1, wherein the machine learning system (M) comprises a first part (K) which first transforms the measured time courses (x) into first quantities (z, qe (z | x)), each latent variable (z), where a space of latent variables has a reduced dimen- sionality,

and wherein the machine learning system (M) comprises a second part (D) which, depending on latent variables (z), generates second quantities (χ ', ρφ (χ' | ζ)) which respectively generate the generated time profiles (χ ') characterize the farm sizes.

The method of claim 2, wherein the first quantities (z, q _e (z | x)) characterizing the latent variables (z) are the latent variables (z) themselves, and wherein the second part is one of third parameters (γ ) parameterized Gaussian process model, and the third parameters (γ) and the latent variables (z) during training of the machine learning system (M) are adapted such that a marginal probability (p (x | z)) the reconstruction of the measured time courses (x) under these latent variables (z) is maximized.

The method of claim 3, wherein the first part (K) comprises a neural network parameterized by fourth parameters (v), and adaptation of the latent variables (z) during training is done by adjusting the fourth parameter (v).

5. The method of claim 2, wherein the first part (K) and the second part (D) of the machine learning system (M) form an auto-encoder.

6. The method of claim 2, wherein the first part (K) and second part (D) of the machine learning system (M) form a variational auto-encoder (English: variational auto-encoder).

The method of claim 2, wherein the first quantities characterizing the latent variables (z) are the latent variables (z) themselves, and the first part (K) is from measured timings (x) by a method of sparse dictionary learning (English : sparse dictionary learning) the latent variables (z) determines which coefficients of the respective measured temporal courses (x) represent in the representation as a linear combination of the dictionary learned by means of this method.

8. The method according to any one of claims 1 to 7, wherein latent variables (z) are given and the machine learning system (M) depending on these predetermined latent variables (z) generates time profiles (χ ') of the operating variables, and then the emissions ( e) are determined as a function of these generated time profiles (χ ').

9. The method of claim 8, wherein the latent variables (z) are determined by a method of statistical experimental design.

10. The method of claim 8, wherein a probability density distribution of the latent variable (z), which is determined as a function of the measured time courses (x), and the predetermined latent variables (z) are drawn as a sample from this estimated probability density distribution.

11. The method according to claim 1, wherein the machine learning system (M) comprises a first part (U) to which either measured time profiles (x) of the operating variables or of a second part (H) of the machine learning system (M) are generated ( χ ') of the operating variables, are fed

wherein the first part (U) is trained to be able to decide as well as possible whether a measured time course (x) or a generated time course (χ ') of the operating variables is supplied to it,

wherein the second part (H) is trained, depending on randomly selected input variables (r) to generate temporal profiles (χ ') of the operating variables as possible so that the first part (U) can decide as badly as possible, if he has a measured time course (x) or a generated time course (χ ') of the operating variables is supplied.

12. The method of claim 11, wherein randomly selected input variables (r) are specified and the machine learning system depending on these predetermined input variables temporal courses (χ ') generates the, and then the emissions are determined depending on these generated time courses

13. The method of claim 12, wherein at least some of the randomly selected input variables are determined by a method of statistical experimental design.

14. The method according to any one of the preceding claims, wherein the motor vehicle is driven depending on the determined emissions.

A computer program (220) arranged to perform the method of any one of the preceding steps.

16. A machine-readable storage medium (210) on which the computer program according to claim 15 is stored.

17. A computer (200) arranged to perform the method of any one of claims 1 to 14.