Method for determining driving trend

Abstract

The present invention relates to a computer-implemented method for training a machine learning system, which is used to generate a driving trend and / or a driving route of a vehicle. The method comprises the following steps: a generator of the machine learning system obtains a first random vector as an input parameter and generates a first driving route and a first driving trend respectively for the first random vector; the driving route and the driving trend are stored in a database; a second driving route and a second driving trend are selected from the database; a discriminator obtains a first pairing and a second pairing as input parameters; the discriminator calculates an output based on the input parameters, and outputs a representation for each pairing obtained as an input parameter, wherein the pairing is a first pairing consisting of the generated first driving route and the first driving trend or a second pairing consisting of the second driving route and the second driving trend; and an objective function representing the distance between the distribution of the first pairing and the distribution of the second pairing is optimized based on the output of the discriminator.

Description

Technical Field

[0001] The present invention relates to a computer-implemented method for generating or evaluating a driving maneuver of a vehicle using a machine learning system, a computer-implemented method for learning such a machine learning system, and a computer program and a learning system provided therefor. Background Art

[0002] DE 10 2017 107 271 A1 discloses a method for determining a master driving cycle for a driving test used to determine exhaust emissions from a motor vehicle. Speed ​​profiles for different driving cycles are derived based on a parameter set. The goal is to determine a master driving cycle that reflects the "maximum" emissions scenario as closely as possible within given boundary conditions. Summary of the Invention

[0003] In some countries, legislation stipulates that new vehicles powered by internal combustion engines are approved based on the emissions generated during actual driving. For this reason, the term "real driving emissions" is also commonly used. This includes, for example, vehicles powered solely by internal combustion engines, but also vehicles with hybrid drivetrains.

[0004] To this end, it is provided that the inspector conducts one or more driving cycles with the vehicle and measures the resulting emissions. The vehicle's approval is then tied to these measured emissions. In this case, the inspector can freely select the driving cycle within wide limits. A typical driving cycle duration can be, for example, 90-120 minutes.

[0005] Therefore, a challenge for motor vehicle manufacturers during the development of a new motor vehicle is to be able to foresee in advance whether the emissions of the motor vehicle will remain within legally prescribed limits in every permitted driving cycle.

[0006] Therefore, it is important to provide methods and devices that can reliably predict the expected emissions of a motor vehicle even during its development phase, so that modifications to the vehicle can be implemented if limit values ​​are expected to be exceeded. Such estimates, which are based solely on measurements on a test bench or in a driving vehicle, are extremely complex due to the large variety of conceivable driving cycles.

[0007] Therefore, in the prior art, it is proposed, for example, to determine a so-called dominant cycle for which compliance with emission regulations is particularly challenging. It is assumed that if this is the case for the most challenging cycle, then emission regulations can be met for all conceivable driving cycles.

[0008] However, in addition to meeting the requirements of exhaust gas regulations in every conceivable or permissible driving cycle, a key goal in vehicle or drive system development is to minimize the overall emissions of the vehicle drive system during real-world operation. While adapting or optimizing a vehicle drive system to the most critical or particularly critical driving cycle may ensure compliance with standards in all cycles, there is the risk of significantly worsening emissions in less critical cycles. If the less critical cycles in real-world driving are still more frequent (which is often the case), then this optimization can lead to a decrease in emissions for the entire system during real-world operation. For example, optimizing emissions for a critical, yet infrequent, driving cycle with an extreme speed profile (e.g., driving on an extremely bumpy road with heavy acceleration) can result in worse emissions for less critical but more frequent driving cycles with a common speed profile (e.g., short city trips with traffic lights), which can ultimately lead to higher emissions in real-world operation.

[0009] Therefore, for the development of emission-optimized vehicles with internal combustion engines, it is very advantageous to be able to automatically generate a large number of real speed profiles whose distribution corresponds to or approximates the real expected distribution. Therefore, the goal is to generate speed profiles with a distribution that is representative of real operation.

[0010] In addition to the development of low-emission drive systems or their emission-optimized applications, speed profiles generated in this way can also be used advantageously in predictive driving, for example in battery management for electric vehicles or electric bicycles, drive management for hybrid vehicles, and regeneration management of exhaust components in vehicles with internal combustion engines. Speed ​​profiles generated in this way can also make a valuable contribution to determining load sets and load scenarios for component specification, such as the loads to which a specific component, such as a pump, is subjected during its service life.

[0011] Therefore, the computer-aided generation of speed profiles with a distribution that represents real driving is an important technical task that can decisively improve the development or optimization of vehicles in different scenarios and thus contribute to less-emission and more efficient vehicles, in particular to less-emission and more efficient vehicle drive systems.

[0012] Thus, in a first aspect, a computer-implemented method for training a machine learning system to generate driving trends of a vehicle is presented.

[0013] Driving trends refer to the course of a vehicle's driving characteristics, where driving characteristics are physical or technical properties of the vehicle's powertrain that can be measured using sensors and that characterize the vehicle's continued movement. The vehicle's speed trend falls within the scope of driving trends as a key variant. For a specific journey, the vehicle's speed trend is one or more dominant variables for determining emissions, consumption, wear, and similar quantities. The speed trend can be determined using speed values, but also variables derived therefrom, such as acceleration values. Other key driving characteristics whose trends are relevant for applications such as determining emissions, consumption, or wear include, in particular, the position of the accelerator pedal or the transmission ratio.

[0014] Here, the training method has the following steps:

[0015] a generator of the machine learning system receives a random vector as an input variable and generates a first driving route and an associated first driving trajectory for each of the random vectors;

[0016] - storing the driving routes and the corresponding associated driving trends detected during driving in a database,

[0017] - selecting a second driving route and a corresponding associated second driving trajectory detected during driving operation from the database,

[0018] the discriminator of the machine learning system receives as input variables a pair consisting of one of the first driving routes with a correspondingly associated first driving trajectory generated and a pair consisting of a second driving route with a correspondingly associated second driving trajectory detected during driving operation,

[0019] the discriminator calculates an output as a function of the input variables, which output characterizes, maps, or quantifies each pairing obtained as input variable, whether the pairing is a pairing with a generated first driving trajectory or a pairing with a second driving trajectory detected during driving operation,

[0020] Based on the output of the discriminator, a target function is calculated, in particular optimized, which represents, maps or quantifies the distance or divergence between the distribution of pairs with a generated first driving profile and the distribution of pairs with a second driving profile detected during driving operation.

[0021] Preferably, the parameters of the machine learning system are adapted according to the optimization of the objective function such that a. the discriminator is optimized for distinguishing between the generated first driving trend and the second driving trend detected during driving operation, and b. the generator is optimized for generating the generated first driving trend with the first distribution, and the discriminator makes it as difficult as possible to distinguish the generated first driving trend from the second driving trend detected during driving operation with the second distribution.

[0022] In a preferred design, the parameters of the machine learning system are matched according to the gradient of the objective function.

[0023] The proposed training method provides a computer-implemented machine learning system that can generate representative driving trends and representative driving routes, or representative pairs of driving trends and driving routes. This allows for measures such as emissions optimization or system verification with respect to emissions, while taking into account the actual representative influences. For example, a route with a characteristic gradient can be generated as a representative driving route, and a speed profile can be generated as a representative driving profile. Using these generated parameters, various optimizations or verifications of the drive system can be automatically performed.

[0024] The objective function is preferably implemented as a static distance, such as the Jenson-Schannon distance. Preferably, the objective function is implemented as a Wasserstein metric, in particular as a Wasserstein distance between a first distribution of a first driving trend and a second distribution of a second driving trend. As a result, the distribution of the generated data advantageously reflects the full variance of the distribution of the measured data, preventing so-called mode collapse. In addition, more stable and more efficient training and better convergence are achieved because the objective function prevents vanishing gradients. The objective function is robust with respect to excessive optimization steps in the discriminator. In order to optimize the use of the Wasserstein metric as an objective function, it is proposed in a preferred design to extend the objective function with a regulating element or to perform weight clipping.

[0025] In addition, the input variables of the generator and the discriminator each include additional information, in particular driver characteristics and / or vehicle characteristics. This additional information, which is also non-discretized, can also be used to adjust the driving trajectory based on other variables, such as the vehicle's maneuverability or the driver's experience, which allows for more precise adaptation and inferences based on the generated driving trajectory.

[0026] The random vector is generated, in particular, by a random generator. The random variable can preferably be implemented as a global random vector or as a transient or local random vector. Alternatively, the random variable can also be implemented as a combination of global and transient or local random vectors. The combination of global and transient / local random vectors has proven to be particularly advantageous because it allows the variance in the data to be mapped based not only on global influences but also on local or transient influences.

[0027] In a preferred embodiment, the generator and / or the discriminator are each implemented as a neural network, in particular as a recurrent neural network. This allows the generation or evaluation of driving profiles of arbitrary length, wherein the transitions between segments of the driving profile always correspond to the learned transition model. The recurrent neural network can be implemented, in particular, as a long short-term memory (LSTM) neural network or a gated recurrent unit (GRU).

[0028] In a preferred embodiment, so-called anticipatory driving can also be advantageously taken into account. This takes into account that the vehicle's speed can already be adapted to future or subsequent route characteristics, such as already visible traffic lights, known subsequent speed limits, etc., through driver behavior or also through automatic vehicle intervention. For this purpose, the generator is implemented as a bidirectional recurrent neural network in a preferred embodiment.

[0029] In other proposed aspects, elements of a machine learning system that learns using the proposed training method can be used as a computer-implemented system to generate (generator) or evaluate (discriminator) driving maneuvers.

[0030] In particular, the route-specific emissions of the vehicle's drive system can be determined based on the driving profile generated in this way, for example in a simulation in which a model of the drive system is calculated. Such a model can include submodels that describe the engine and the exhaust gas aftertreatment system of the drive system.

[0031] Based on the route-specific emissions of the vehicle's drive system determined on the basis of the driving profile generated in this way, the drive system can then be verified or adapted, in particular adapted to minimize emissions.

[0032] As already explained, an adaptation can be performed here by generating a representative distribution of driving profiles, which does not optimize the drive system with respect to emissions for a single or particularly critical driving profile. Instead, the drive system can be optimized so that the emissions expected in real operation are minimized overall.

[0033] Optimization can be performed by adapting components or parameters during the development of the drive system, by adapting data in the application of the drive system, or by adapting control variables during operation of the drive system in the vehicle.

[0034] To perform the described computer-implemented method, a computer program can be provided and stored on a machine-readable memory. A computer-implemented learning system including such a machine-readable memory can be configured to perform the method, wherein the calculations to be performed are executed by one or more processors of the computer-implemented learning system. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 A computer-implemented training method for a machine learning system is exemplified. DETAILED DESCRIPTION

[0036] The driving routes or the routes of the vehicles together with the corresponding associated driving trends are stored in the database 2. Figure 1 In the example, an exemplary pairing of a route and an associated driving profile in database 2 is indicated by 21. The driving profile in database 2 corresponds to a driving profile determined or measured during vehicle operation. In other words, the driving profile is preferably detected and stored by the vehicle's sensors while the vehicle actually travels the associated route. In this context, the term "database" simply refers to data systematically stored in a machine-readable memory.

[0037] In the machine learning system 4, a generator 41 is now to be trained to generate driving profiles and associated driving routes. These driving profiles are to be determined based on random input variables, for which random variables, such as random vectors, can be provided in block 3. In particular, a random generator can be implemented in block 3, which can also be referred to as a pseudo-random generator.

[0038] The driving profile generated by generator 41 should preferably be as indistinguishable or nearly indistinguishable as possible from driving profiles determined during driving operation and from database 2. To this end, discriminator 42 is trained to distinguish as well as possible between driving profiles generated by generator 41 and driving profiles extracted from database 2, or between corresponding pairs of driving profiles and driving routes or route characteristics. The learning system should not only generate individual driving profiles and driving routes that are as indistinguishable or nearly indistinguishable from individual driving profiles determined during driving operation. Rather, the distribution of the generated driving profiles and driving routes in the parameter space of the input variables should also be as close as possible to the distribution of the driving profiles and driving routes determined during driving operation in the parameter space of the input variables, i.e., to achieve a representative distribution of the driving profiles and driving routes.

[0039] To this end, the training of the machine learning system 4 includes the optimization of an objective function 5 , according to which the parameters of the generator 41 and of the discriminator 42 are adapted.

[0040] The following will refer to Figure 1 The proposed training of the machine learning system 4 is described in more detail.

[0041] The route is determined in particular as a sequence of discrete data points, wherein the route characteristics are stored for each data point or each discretization step in the discretization step. This applies in particular to generated routes, for example routes stored in a database.

[0042] For example, a route r has length N: r=(r1, ..., r N ). Each data point r t corresponds to a discretization step. Particularly preferred are implementations in which the discretization step corresponds to a temporal or spatial discretization. In the case of temporal discretization, the data points each correspond to the time elapsed since the start of the route, and thus the sequence of data points corresponds to the temporal progression. In the case of spatial discretization, the data points each correspond to a segment along the route.

[0043] The sampling rate is usually constant. In time discretization, the sampling rate can be defined as x seconds, for example, or in space discretization, as x meters.

[0044] Each data point r of the route t Describes the route characteristics at the corresponding discretization step, that is, r t ∈R D D is the number of route features, where each dimension of the multidimensional route feature is counted as a dimension of the one-dimensional route feature.

[0045] Such route characteristics can be expressed, for example, in each case with respect to discretized steps, in particular points in time or time intervals or positions or sections or distances:

[0046] - geographical features, such as absolute altitude or slope,

[0047] - traffic flow characteristics, such as the time-dependent average speed of traffic,

[0048] - lane characteristics, such as number of lanes, lane type or lane curvature,

[0049] - traffic guidance features such as speed limits, number of traffic lights or specific traffic signs, in particular stop or yield or pedestrian crossings,

[0050] - Weather characteristics, such as the amount of rain, wind speed, presence of fog at a pre-given point in time.

[0051] In particular, the generated route and the route stored in the database are determined here by the same route characteristics or the same type of route characteristics.

[0052] In block 3 a random vector is determined and in step 31 it is transmitted to the generator 41. The random vector z is extracted, ie randomly determined. In this case, in particular z∈R L , where L may optionally depend on the length of the route N. The distribution is preferably fixed to a simple family of distributions, such as a Gaussian distribution or a uniform distribution, from which z is extracted.

[0053] Therefore, the input variables of the generator 41 include at least the random vector z. The generator 41 now generates a driving profile and a driving route based on the input variables (step 31). For this purpose, the generator 41 has a computer-implemented algorithm that implements the generative model and outputs the driving profile and the driving route (step 43).

[0054] Such a driving profile generated by the generator 41 can be output as, for example, x=(x1, . . . , x N ), and is discretized with respect to position like the associated driving route generated. Alternatively, for example, the generated route characteristics, ie the driving route, can also be discretized with respect to position, but the generated driving trend, ie, for example, the generated speed, is discretized with respect to time.

[0055] To influence or determine the generated driving route and its length, in a preferred embodiment, the driving profile and the length N of the driving route to be generated can be additionally specified to the generator 41. This can be a fixed or configurable value, or in a preferred variant, it can be sampled from a distribution that represents the actual distribution of route lengths. For example, this distribution can be extracted from the described database or another database. If a recurrent neural network is used as the generator 41, the generation can be interrupted after N steps, or N can be provided as a further input variable to the generator 41.

[0056] In an alternative embodiment, a specific interruption criterion, in particular a random interruption criterion, is defined, which determines the length of the generated data (driving trajectory, driving route) during the generation. For example, a special symbol (e.g., $, -1, NAN) can be introduced to signal the end of the route. The route ends as soon as the generator 41 generates this symbol for the first time.

[0057] The generative model is generated by parameter θ G Parameterization. For example, the architecture of the generative model can be a recurrent neural network. The generator is implemented by storing the algorithm for implementing the generative model and the model parameters in a machine-readable memory, processing the algorithm's computational steps via a processor, and storing the resulting driving patterns in the machine-readable memory.

[0058] In one possible design, driving trends can be generated with a fixed length, i.e., with a fixed number of discretized steps or data points. When generating longer driving trends, the generated multiple short time sequences must be linked together. However, the transitions are usually inconsistent here. In an alternative design, the method can be expanded in such a way that driving trends of arbitrary length can also be generated or evaluated, and the transitions are always consistent with the learned transition model. To this end, both the generator and the discriminator are preferably implemented as recurrent neural networks, for example as long short-term memory (LSTM) neural networks or gated recurrent units (GRUs). Architecturally, the generator is preferably designed as a vector-to-sequence model, but can also be implemented as a sequence-to-sequence model. The discriminator is preferably designed as a sequence-to-scalar model, but can also be implemented as a sequence-to-sequence model.

[0059] There are different options for the architecture of generative models as recurrent neural networks.

[0060] For example, a global random vector z can be sampled for the entire driving trajectory, where the term "global" can also refer to a temporal or spatial discretization. In this design, globally varying driving trajectory properties are taken into account or learned in the latent space, such as properties that are constant over the route, such as constant driver characteristics (e.g., age or experience), constant weather characteristics (e.g., continuous rain), or constant vehicle characteristics (e.g., maneuverability). This random vector can then be used to initialize the hidden state (hidden state) in the first time step and / or fed into the recurrent neural network at each time step.

[0061] Alternatively, local or transient random vectors z can be sampled. This means that in the latent space, characteristics that can locally or transiently change driving behavior, such as short-term characteristics such as short-term traffic guidance or traffic flow characteristics (traffic light status, congestion at intersections, pedestrians on the roadway), can be considered or learned. Here, a random vector is regenerated at time steps separated by M, where M > 0, and fed into the recurrent neural network. M can also be random, meaning the random vector can also change at random intervals.

[0062] In a preferred embodiment, a combination of global and local or transient random vectors is also possible. Here, some dimensions of the random vector are sampled only once for each driving trajectory, while the remaining dimensions are changed every M time steps. Alternatively, in principle, a global random vector can also be fed to the recurrent neural network at each time step, with the global random vector being replaced every M time steps by a local (i.e., resampled) random vector.

[0063] The combination of global and local random vectors has proven to be particularly advantageous, since in this way the variance in the data can be mapped not only on the basis of global influences but also on the basis of local or transient influences.

[0064] Preferably, anticipatory or predictive driving can also be taken into account in the generative model.

[0065] Thus, a bidirectional recurrent neural network can be used as a generative model in which the hidden states of future units of the recurrent neural network are additionally taken into account. This allows all possible future time steps to be explicitly included.

[0066] That is, instead of generating a speed or route characteristic at time t and adjusting only the internal state at the current time t (and, if necessary, the hidden state at time t-1), future internal states and, therefore, also indirectly, the future route characteristic and speed profile used to generate the speed at time t can also be included. This allows for the simulation of so-called "anticipatory driving," for example, where the driver reacts in advance to route characteristics that are already visible from a distance (e.g., traffic lights, speed limits, highway exits, etc.), particularly in terms of speed adaptation. Furthermore, the algorithm can learn to return to a speed of zero at the end of a route, for example, by only having a default value, such as zero, for the future route characteristic at the end of the route.

[0067] Each route determined in this manner is stored in database 2 along with the driving profiles actually measured during driving and assigned to the route. The pairs of these routes and the associated driving profiles serve as training data for the machine learning system. For this training, pairs of routes and associated driving profiles are selected and transmitted to discriminator 42 in steps 23 and 24. Furthermore, pairs of the generated driving route and the driving profile generated from the route by generator 41 are also transmitted to discriminator 42 in steps 13 or 43.

[0068] The discriminator 42 has a computer-implemented algorithm with which the discrimination model is implemented. The discriminator 42 receives as input a pair consisting of a route and an associated driving trend and determines whether the pairing seen contains a driving trend generated (by the generator 41) or actually measured (obtained from the database 2). The result of this determination is output in step 44. For example, the discriminator 42 can output a value > 0 for the determination "real driving trend" and a value < 0 for the determination "generated driving trend". Alternatively, for example, a previously determined value, such as a class label, can also be output. The discrimination model is determined by the parameter θ D The output 44 of the decision comprises in particular an evaluation of a binary decision “yes” / “no”.

[0069] The computer implementation of the discriminator is achieved by storing an algorithm implementing the discrimination model and parameters of the model in a machine-readable memory, processing the computational steps of the algorithm by a processor and storing the output in the machine-readable memory.

[0070] Discriminator 42 can be implemented, for example, as a recurrent neural network. In particular, driving profiles of arbitrary length can be evaluated.

[0071] There are various design options for evaluating the generated driving profile and route against the driving profile and route characteristics measured during driving. In particular, the evaluation can be re-output after individual time steps. The global evaluation of the driving profile and route can then be, for example, an average of the individual evaluations or a majority decision. Alternatively, the entire driving profile and route can be evaluated only in the last time step. This latter design option particularly saves on additional computational steps and has the further advantage that the entire driving profile and route are also included in the evaluation.

[0072] Based on the output 44 of the discriminator 42, an objective function is optimized in block 5, in particular, a loss function is minimized. To this end, the input variables of the discriminator are in particular labeled as real samples (i.e., pairs with driving profiles determined during driving operation) or as generated samples (i.e., pairs with driving profiles generated by the generator 41). The objective function characterizes the extent to which the generated driving profile corresponds to the actually measured driving profile, or the extent to which the distribution of the generated driving profile in parameter space corresponds to the distribution of the measured driving profile in parameter space. Depending on the adaptation of the objective function, the parameters θ of the generator 41 or the generative model implemented therein are G and the parameters θ of the discriminator 42 or of the discriminator model implemented therein D In this case, the parameters are adapted in particular with respect to the gradient of the objective function.

[0073] The objective function is selected so that it characterizes or represents the difference or distance between the distribution of the generated driving trends and driving routes and the distribution of the driving trends and driving routes measured during driving operation, or the difference or distance between the distribution of route-driving trend pairs with the generated driving trends and the distribution of route-driving trend pairs with the driving trends measured during driving operation. By selecting such an objective function, the machine learning system can be trained so that the distribution of the generated data reflects the full variance of the distribution of the measured data. This prevents so-called mode collapse. This means that a representative distribution of driving trends and driving routes is provided. In particular, the objective function also takes into account the variance of unobservable influences.

[0074] For this purpose, a loss function (loss) is preferably chosen as the objective function, which loss function is implemented as a Wasserstein metric or a Wasserstein distance between distributions.

[0075] In computer-implemented training schemes, the discriminator should preferably be restricted to the Lipschitz limit In a preferred embodiment, the objective function is extended for this purpose by a regularization term, for example by a gradient penalty (gradient penalty) or by centering (i) the gradient of the real samples (i.e., the pairings with the driving trends determined during driving operation) to 0 or (ii) the gradient of the generated samples (i.e., the pairings with the generated driving trends) to 0 or (iii) the gradient of the sample representing the average of the real samples and the generated samples to 1. Particularly preferably, the option "centering the gradient of the real samples to 0" is used here, as this has proven to be the fastest option and leads to a particularly fast convergence of the optimization problem. Alternatively, weight clipping can be performed after each gradient step.

[0076] The method described above for computer-implemented training of the entire learning system, including the generator 41 and the discriminator 42, can be described as a min-max training objective. Here, the discriminator 42 maximizes its correct classification rate, while the generator 41 minimizes the correct classification rate by generating driving patterns and routes that are as misleading as possible to the discriminator 42.

[0077] In addition to the input variables described for generator 41, other input variables can also be provided to the generator. For example, for each driving route, in addition to the random vector, information about driver characteristics (such as experience, age, aggressiveness, or driving style) or vehicle information (such as power, maneuverability, drive system type, etc.) can also be stored. The generated driving profile can thus also be adjusted based on this additional information. In this case, for training machine learning system 4, corresponding information is also advantageously stored for each driving route stored in database 2. This additional information is provided as input variables to the discriminator along with the route information and driving profile from database 2 (step 23).

[0078] A computer-implemented machine learning system that learns using the described training method can generate representative driving trends and driving routes, or representative pairs of driving trends and driving routes. To this end, random variables and, if necessary, other information can be fed to a generator of the machine learning system that has learned in this way, and the generator generates the corresponding driving trends and driving routes. The driving trends and / or driving routes generated in this way can therefore be used to simulate vehicle emissions and thus for probabilistic evaluation to comply with exhaust gas standards and for emission optimization. Emission optimization can be performed, for example, by adapting the drive system during development, by optimizing in the data application of the drive system, for example, in the data application of a controller for drive system control, or by adapting the control parameters of the drive system in the vehicle to optimize emissions. In the latter case, the generation of the driving trends and driving routes can be performed, in particular, in the vehicle.

[0079] Driving profiles and / or driving routes can also be used for predictive driving optimization, for example in battery management for electric vehicles or electric bicycles, drive management for hybrid vehicles, or regeneration management of exhaust components in vehicles with internal combustion engines. This optimization can, in turn, be performed during development by optimizing the corresponding controller applications or by adapting the control parameters of the corresponding systems in the vehicle.

[0080] Likewise, driving profiles and / or driving routes generated in this way can make a valuable contribution to determining load sets and load scenarios for component specification, for example, what loads a specific component, such as a pump, is subject to during its service life.

Claims

1. A computer-implemented method for training a machine learning system (4) for generating a driving trend and / or a driving route of a vehicle, wherein: The driving trend includes the speed trend of the vehicle, the trend of the accelerator pedal position or the trend of the transmission ratio. It is characterized by the following steps: - a generator (41) of the machine learning system (4) receives a first random vector as an input variable and generates a first driving route and an associated first driving trend for the first random vector, - storing the driving routes and the corresponding associated driving trends detected during driving in a database (2), - selecting a second driving route and a corresponding associated second driving trajectory detected during driving operation from the database (2), - a discriminator (42) of the machine learning system (4) receives as input variables a first pairing consisting of a generated first driving route and a correspondingly associated first driving trend generated, and a second pairing consisting of a second driving route and a correspondingly associated second driving trend detected during driving operation, the discriminator (42) calculates an output based on the input variables, which output characterizes for each pairing obtained as input variable whether the pairing is a first pairing consisting of a generated first driving route and a corresponding associated first generated driving trend or a second pairing consisting of a second driving route and a corresponding associated second driving trend detected during driving operation, - optimizing an objective function (5) based on the output of the discriminator (42), the objective function representing the distance between the distribution of the first pair and the distribution of the second pair.

2. The method according to claim 1, characterized in that The parameters of the machine learning system (4) are adapted according to the optimization of the target function (5) so that the discriminator (42) is optimized for distinguishing between a generated first driving trend and driving route and a second driving trend and driving route detected during driving operation, The generator (41) is optimized for generating a first driving trend and driving route generated with a first distribution, which is difficult to distinguish by the discriminator (42) from a second driving trend and driving route detected during driving operation and present with a second distribution.

3. The method according to claim 2, characterized in that Parameters of the machine learning system (4) are matched according to the gradient of the objective function (5).

4. The method according to claim 1, wherein The statistical distance between the first distribution of the first driving trajectory and the driving route and the second distribution of the second driving trajectory and the driving route is implemented as a target function (5).

5. The method according to claim 4, characterized in that The objective function (5) is implemented as the Jensen-Shannon distance or as the Wasserstein metric.

6. The method according to claim 5, characterized in that The objective function (5) is extended to adjust the elements or perform weight clipping.

7. The method according to any one of claims 1 to 6, characterized in that The first driving route and the second driving route are data present in temporal or spatial discretization steps, wherein route characteristics are stored in each discretization step for each route.

8. The method according to claim 7, characterized in that The route characteristics include geographical characteristics, traffic flow characteristics, lane characteristics, traffic guidance characteristics and / or weather characteristics of the route.

9. The method according to any one of claims 1 to 6, characterized in that The input variables of the generator (41) and the discriminator (42) include driver characteristics and / or vehicle characteristics.

10. The method according to any one of claims 1 to 6, characterized in that The first random vector is implemented either as a global random vector or as a local random vector, wherein the global random vector can be sampled for the entire driving profile.

11. The method according to any one of claims 1 to 6, characterized in that The first random vector is implemented as a combination of a global random vector and a local random vector, wherein the global random vector can be sampled for the entire driving profile.

12. The method according to any one of claims 1 to 6, characterized in that The generator (41) and / or the discriminator (42) are each implemented as a neural network.

13. The method according to claim 12, characterized in that The generator (41) is implemented as a bidirectional recurrent neural network.

14. The method according to any one of claims 1 to 6, characterized in that The length of the generated driving trajectory and the generated driving route is determined as a function of predetermined or configurable input variables or as a function of an abort criterion.

15. The method according to claim 4, characterized in that The target function (5) is implemented as a Wasserstein distance between a first distribution of a first driving trajectory and a driving route and a second distribution of a second driving trajectory and a driving route.

16. The method according to any one of claims 1 to 6, characterized in that The generator (41) and / or the discriminator (42) are each implemented as a recurrent neural network.

17. A method for generating a driving trend and / or driving route of a vehicle, wherein: The driving profile is generated by a computer-implemented generator (41) of a machine learning system (4), which learns using a method according to any one of claims 1 to 16.

18. A method for evaluating a vehicle's driving trend and / or driving route, wherein: The driving profile and / or driving route is evaluated by a computer-implemented discriminator (42) of a machine learning system (4), which is learned using a method according to any one of claims 1 to 16.

19. A method for determining route-specific emissions of a drive system of a vehicle, wherein the route-specific emissions are determined based on a driving profile and / or a driving route generated using the method according to claim 17.

20. The method according to claim 19, characterized in that The emissions are determined in a simulation in which a model of the drive system is calculated.

21. The method according to claim 20, characterized in that The model of the drive system includes sub-models that describe the engine and the exhaust gas aftertreatment system of the drive system.

22. A method for matching a drive system of a vehicle, characterized in that: The drive system of the vehicle is adapted according to route-specific emissions determined using the method according to one of claims 19 to 21 .

23. The method according to claim 22, characterized in that The adaptation is performed by adapting components or parameters during the development of the drive system, by adapting data during the use of the drive system, or by adapting control variables during the operation of the drive system in the vehicle.

24. A computer program product configured to perform the method according to any one of claims 1 to 23. 25 . A machine-readable storage medium having a computer program stored thereon, the computer program being configured to execute the method according to claim 1 .

26. A computer-implemented machine learning system (4) having a machine-readable storage medium according to claim 25.

Images