Method for determining a driving change process

Abstract

Method for determining driving change processes, with the steps: selecting first driving routes from a first database, obtaining the first driving routes as input variables, generating for the first driving routes a first driving change process belonging thereto, storing the driving routes and respectively belonging driving change processes detected in a driving operation in a second database, selecting from the second database a second driving route and respectively belonging second driving change process detected in a driving operation, obtaining as input variables a pairing of one of the first driving routes with respectively belonging first generated driving change process and a pairing of the second driving route with respectively belonging second driving change process detected in a driving operation, calculating an output from the input variables, optimizing an objective function from the output, which objective function represents a distance between a distribution with the pairing of the first generated driving change processes and a distribution with the pairing of the second driving change processes detected in a driving operation.

Description

TECHNICAL FIELD

[0001] The present application relates to a computer-implemented method for generating or evaluating a driving course of a vehicle by means of a machine learning system, a computer-implemented method for teaching such a machine learning system and a computer program and a learning system designed therefor. BACKGROUND

[0002] DE 10 2017 107 271 A1 discloses a method for determining a dominant driving cycle for carrying out a driving test to determine exhaust emissions from a motor vehicle. Here, a speed profile for different driving cycles is derived on the basis of a parameter set. The aim is to determine a dominant cycle which reflects the "maximum" emission situation within given boundary conditions as far as possible.

[0003] In some countries, it is legally prescribed that the approval of a new motor vehicle driven by means of an internal combustion engine depends on the emissions produced in actual driving operation. The English term real driving emissions is also used for this. Such motor vehicles include, for example, motor vehicles driven exclusively by means of an internal combustion engine, but also motor vehicles having a hybrid drive train.

[0004] For this purpose, an examiner drives through one or more driving cycles with the motor vehicle and measures the emissions produced in the process. The approval of the motor vehicle then depends on these measured emissions. Here, the examiner can freely select the driving cycles within wide limits. A typical duration of a driving cycle can be, for example, 90-120 minutes here.

[0005] It is therefore a challenge for the manufacturer of a motor vehicle in the development of a motor vehicle to have to foresee already early on in the development process of a new motor vehicle whether the emissions of the motor vehicle remain within the legally prescribed limits in each permissible driving cycle.

[0006] It is therefore important to provide methods and devices which can reliably predict the expected emissions of a motor vehicle already in the development phase of the motor vehicle in order to be able to make changes to the motor vehicle in the event of an expected exceedance of a limit value. Since a large variety of driving cycles can be considered, such an estimation based only on measurements on a test stand or in a motor vehicle in operation is very complex.

[0007] Therefore, in the prior art it is proposed, for example, to determine so-called dominant cycles for which it is particularly challenging to meet the emissions regulations. It is assumed here that if the emissions regulations are met for the most challenging cycles, then the emissions regulations will be met for all conceivable driving cycles.

[0008] However, in addition to the requirement that the exhaust gas regulations must be met in each conceivable or permissible driving cycle, a further important objective of vehicle development or driver development is to minimize the total emissions of the vehicle drive system in actual operation. While it can be possible to guarantee compliance with the standards in all cycles in the case of a vehicle drive system that is adapted or optimized to the most critical or particularly critical driving cycles, there is a risk of a significant deterioration in emissions in less critical cycles thereby. Thus, if the less critical cycles are also more frequent cycles in actual driving operation, which is often the case, the entire system is worsened in terms of emissions in actual operation by such an optimization. For example, optimizing emissions to a critical, but actually very rare, driving cycle with an extreme speed profile (for example, an extreme uphill drive with strong acceleration) can lead to a deterioration in emissions for less critical, but more frequent, driving cycles with a common speed profile (for example, short, city driving with traffic lights), which can lead to higher emissions in actual operation overall.

[0009] Therefore, in order to develop a vehicle with an emissions-optimized internal combustion engine, it is very advantageous to be able to automatically generate a large number of realistic speed profiles, the distribution of which corresponds to or approximates the actual expected distribution. The aim is therefore to generate speed profiles with a distribution that is representative of actual operation.

[0010] In addition to the development of low-emission drive systems or the application of emissions-optimized drive systems, such generated speed profiles can also be advantageously used in predictive driving, for example in battery management for electric vehicles or electric bicycles, drive management for hybrid vehicles, regeneration management of exhaust constituents for vehicles with internal combustion engines. Speed profiles generated in this way can also make a valuable contribution to determining load spectra and load scenarios for specifying components, for example, what kind of load a particular component (for example, a pump) will experience during its service life. SUMMARY

[0011] Therefore, the computer-aided generation of speed profiles in a distribution that is representative of actual operation is an important technical task which can significantly improve the development or optimization of vehicles in various scenarios and thereby contribute to lower-emission and more efficient vehicles, in particular to lower-emission and more efficient vehicle drive systems.

[0012] Therefore, in a first aspect, a computer-implemented method for training a machine learning system for generating a driving progress of a vehicle is proposed.

[0013] The driving progress here refers to a driving characteristic of the vehicle, wherein the driving characteristic is a characteristic of the vehicle drive train which is measurable with sensors, in particular a physical or technical characteristic, which characterizes the movement of the vehicle. As the most important variant, the driving progress comprises a speed progress of the vehicle. The speed progress of the vehicle is one or more dominant variable for determining emissions, consumption, wear and similar variables for a specific driving. Here, the speed progress can be determined both by the speed value, but also by variables derived from the speed value, such as the acceleration value. Other important driving characteristics, the progress of which is important for applications such as determining emissions, consumption or wear, in particular include the position of the accelerator pedal or the transmission ratio.

[0014] The training method here comprises the following steps:

[0015] - selecting first driving routes from a first database having driving routes,

[0016] - a generator of the machine learning system obtains the first driving routes as input variables and generates a first driving progress belonging thereto for these first driving routes, respectively,

[0017] - storing the driving routes and the respectively belonging driving progress detected in driving operation in a second database,

[0018] - selecting second driving routes and the respectively belonging second driving progress detected in driving operation from the second database,

[0019] - a discriminator of the machine learning system obtains as input variables a pairing of one of the first driving routes with the respectively belonging first generated driving progress and a pairing of the second driving routes with the respectively belonging second driving progress detected in driving operation,

[0020] - the discriminator calculates an output from the input variables, which output characterizes or maps or quantifies for each pairing obtained as input variables whether this is a pairing with a first generated driving progress or a pairing with a second driving progress detected in driving operation,

[0021] - calculating, in particular optimizing, an objective function from the output of the discriminator, which objective function represents or maps or quantifies a distance or difference between a distribution of pairings with first generated driving progress and a distribution of pairings with second driving progress detected in driving operation.

[0022] Herein, the first database and the second database can be implemented as one (common) database.

[0023] Preferably, the parameters of the machine learning system are adapted according to the optimization of the objective function such that

[0024] a. the discriminator is optimized to distinguish the first generated driving change process from the second driving change process detected in the driving operation,

[0025] b. the generator is optimized to generate the first generated driving route in the first distribution, which is as difficult as possible to distinguish from the second driving change process detected in the driving operation existing in the second distribution by the discriminator.

[0026] In a preferred configuration, the parameters of the machine learning system are adapted according to the gradient of the objective function.

[0027] The proposed training method provides a computer-implemented machine learning system with which representative driving change processes can be generated, from which, in turn, measures can be taken, such as optimizing emissions or verifying systems with regard to emissions, taking into account the actual representative influence.

[0028] The objective function is preferably implemented as a static distance, such as the Jenson-Shannon distance. The objective function is preferably implemented as a Wasserstein metric, in particular the Wasserstein distance between the first distribution of the first driving change process and the second distribution of the second driving change process. The distribution of the generated data thus advantageously reflects the full variance of the distribution of the measured data, thus preventing a so-called mode collapse. Furthermore, since the objective function prevents vanishing gradients, more stable and more efficient training and better convergence can be achieved. The objective function is robust to excessive optimization steps in the discriminator. In order to optimize the use of the Wasserstein metric as an objective function, in a preferred configuration it is proposed to extend the objective function with a regularizer or to perform weight clipping.

[0029] In a preferred configuration, the route characteristics of the first driving route are generated at least partially by a machine learning system, in particular by a neural network. Thereby, a very large number of driving change processes can be generated for a large number of automatically generated routes or route characteristics. Routes or route characteristics generated in this way are subject to fewer restrictions compared to actually measured routes. Furthermore, it is complex and technically challenging to detect route data with a large number of complete route characteristics.

[0030] In addition to the first driving route and the second driving route, the input variables of the generator and the discriminator can also include additional information, in particular driver characteristics and / or vehicle characteristics, respectively. By such, in particular also non-discrete, additional information, the driving change process can also be adjusted in terms of other variables, such as the motorization of the vehicle or the experience of the driver, which enables a more precise adaptation and conclusion from the generated driving change process.

[0031] In an advantageous configuration, the generator obtains a random variable as an additional input variable from a random generator and generates a first driving change process belonging to the first driving route from the first driving route and the random variable, respectively. This increases the efficiency of the training method. Here, the random variable can be implemented as a global random vector or a temporary or local random vector, on the one hand. Alternatively, the random variable can also be implemented as a combination of a global random vector and a temporary or local random vector. It has proven particularly advantageous to combine a global random vector and a temporary / local random vector, since by this it is possible to map variations in the data on the basis of global influences as well as local or temporary influences.

[0032] In a preferred configuration, the generator and / or the discriminator are implemented as a neural network, in particular a recurrent neural network, respectively. Thereby, driving change processes of any length can also be generated or evaluated, wherein the transitions between the individual segments of these driving change processes always conform to the learned transition model. Here, the recurrent neural network can in particular be implemented as a long short-term memory (LSTM) neural network or a gated recurrent unit (GRU).

[0033] In a preferred configuration, it can also be advantageous to take into account so-called anticipatory driving. Here, the speed of the vehicle can already be adapted to future or subsequent route characteristics by the driver's behavior or by automatic vehicle interventions (for example by already visible traffic lights, known following speed limits, etc.). To this end, in a preferred configuration the generator can be implemented as a bidirectional recurrent neural network. In order to generate the driving development by means of the generator, it is also possible to supplement or extend the route characteristics of the first driving route at a particular discretization step, in particular at a particular point in time or a particular route distance, with the route characteristics of one subsequent discretization step (in particular the route characteristics of a subsequent point in time or a subsequent route distance) or with the route characteristics of a plurality of subsequent steps.

[0034] In a further aspect, the elements of the machine learning system taught using the proposed training method can be used as a computer-implemented system in order to generate driving developments (generator) or to evaluate driving developments (discriminator).

[0035] In particular, the route-specific emissions of a drive system of a vehicle can be determined from driving developments generated in this way, for example in a simulation in which a model of the drive system is calculated. Here, such a model can include submodels which describe the engine and the exhaust gas aftertreatment system of the drive system.

[0036] The route-specific emissions of a drive system of a vehicle determined using driving developments generated in this way can then be used to verify or adapt the drive system, in particular to adapt it in such a way that the emissions are minimized.

[0037] As already described, the adaptation can be made here by means of a representative distribution of the generated driving developments, which optimizes the drive system in terms of emissions not on the basis of individual driving developments or particularly critical driving developments. Rather, the optimization of the drive system is achieved in such a way that the expected emissions in actual operation are minimized overall.

[0038] Here, the optimization can be made by adapting components or parameters in the development of the drive system, by adapting data in the application of the drive system or by adapting control variables when the drive system is operated in a vehicle.

[0039] In order to carry out the described computer-implemented method, a computer program can be designed and stored in a machine-readable memory. A computer-implemented learning system comprising such a machine-readable memory can be designed to carry out the method, wherein the computations to be carried out are carried out by one or more processors of the computer-implemented learning system. BRIEF DESCRIPTION OF DRAWINGS

[0040] Figure 1 An exemplary training method for a computer-implemented machine learning system is shown. Detailed Implementation

[0041] The driving route or vehicle route is stored in database 1. Example routes in database 1 are... Figure 1 The number 11 is used to represent this. The driving route or vehicle route is stored in database 2 along with its associated driving change process. An example pairing of routes and their associated route change processes in database 2 is shown below. Figure 1 The term 21 is used to represent this. The driving change process in Database 2 corresponds here to the driving change process determined or measured during the vehicle's operation. That is, when the vehicle actually travels along its designated route, the driving change process is preferably detected and stored by the vehicle's sensors. Databases 1 and 2 are implemented throughout the system, particularly on machine-readable storage. Here, "database" refers only to data stored in the system on machine-readable storage.

[0042] In machine learning system 4, generator 41 should now be trained to generate driving variation processes for routes in database 1. These driving variation processes should preferably 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, wherein the random generator can also be a pseudo-random generator.

[0043] The driving variation processes generated by generator 41 should preferably be as indistinguishable as possible, or nearly indistinguishable, from the driving variation processes determined during driving operations from database 2. To this end, discriminator 42 is trained to distinguish, as well as possible, the driving variation processes generated by generator 41 from those obtained from database 2, or to distinguish the respective pairs of driving variation processes and route characteristics. The learning system should not only generate driving variation processes that are as indistinguishable as possible from the individual driving variation processes determined during driving operations, but more importantly, the distribution of the generated driving variation processes in the parameter space of the input variables should also be as close as possible to the distribution of the driving variation processes determined during driving operations in the parameter space of the input variables; that is, it should achieve a representative distribution of the driving variation processes.

[0044] Therefore, training the machine learning system 4 includes optimizing the objective function 5 to adapt the parameters of the generator 41 and the discriminator 42 according to the objective function.

[0045] The following should be based on Figure 1 The training of the proposed machine learning system 4 is described in more detail.

[0046] The routes in the database 1 are stored in particular as sequences of discrete data points, wherein for each data point or each discretization step the route properties in this discretization step are stored.

[0047] The routes r in the database 1 have for example a length N: Each data point r t corresponds to one discretization step. It is particularly preferred that the following implementations are realized: In these implementations the discretization steps correspond to a temporal or spatial discretization. In the case of temporal discretization, the data points correspond to the time elapsed since the start of the route, respectively, so that the sequence of data points corresponds to a time-varying process. In the case of spatial discretization, the data points correspond to the distance traveled along the route, respectively.

[0048] The sampling rate is generally constant. In temporal discretization, the sampling rate can for example be defined as x seconds, in spatial discretization for example as x meters.

[0049] Each data point r t The route properties at the respective discretization step are described, i.e. D is the number of route properties, wherein each dimension of a multi-dimensional route property is counted as one dimension of a one-dimensional route property.

[0050] Such route properties can for example relate to the discretization step, in particular the time point or time interval or the location or the distance, respectively:

[0051] • Geographical properties, such as absolute height or slope

[0052] • Properties of the traffic flow, such as the time-dependent average speed of traffic

[0053] • Properties of the driving road, such as the number of lanes, the driving road type or the driving road curvature

[0054] • Properties of the traffic management, such as the speed limit, the number of traffic lights or the number of specific traffic signs, in particular stop or yield or pedestrian crossing

[0055] • Climatic properties, such as the amount of rainfall, the wind speed, the presence of fog at a predetermined time point.

[0056] A route from the database 1 is selected and transferred to the generator 41 in step 13.

[0057] Additionally, a random vector is determined, preferably in block 3, and transferred to the generator 41 in step 31. The random vector z is drawn, i.e. the random vector z is determined randomly. In this case in particular Where L can optionally depend on the length N of the route. The distribution of z drawn from this is preferably set to a simple family of distributions, such as a Gaussian distribution or an equal distribution.

[0058] Now, the input variables of generator 41 preferably consist of variables—a random vector z and a route r. Therefore, unlike inputs to generator 41 that are generated purely randomly, the generated driving variation process can be adjusted according to specific route characteristics. For example, different driving variation processes can be generated for the same pre-given route r by sampling different z. Here, the route characteristics of route r in database 1 can be actually measured route characteristics, route characteristics defined by experts, or route characteristics learned by a machine learning system (e.g., a neural network). Routes with route characteristics created from two or three of these variations can also be provided in database 1.

[0059] In an exemplary application scenario—where the generated driving variation process is used to determine the emission characteristics of a vehicle's drive system—the extent to which specific driving characteristics affect emissions during combustion can be specifically examined by selectively altering certain route characteristics, generating matching driving variation processes, and simulating emissions against these profiles. This allows for targeted optimization of the drive system parameters, particularly the drive system's control parameters, for example, within a control device, for specific route curves or routes with particular requirements.

[0060] Generator 41 now generates the driving change process based on the input variable random vector (step 31) and the selected route (step 13). To this end, generator 42 has a computer-implemented algorithm that generates the model and outputs the driving change process (step 43).

[0061] This driving change process generated by generator 41 can be, for example, as 𝑥 = (𝑥1, ..., 𝑥) 𝑁 The output is thus the same length N as the route described below, and the driving variation process is adjusted according to the route. Alternatively, however, the route characteristics can also exist discretized with respect to location, but the generated speed can be discretized with respect to time. For this purpose, after each step starting from the previous location, the next discretized time point can be calculated using the generated speed, and the route characteristics at that location can then be used as x. {t+1} Input.

[0062] The driving variation process is adjusted based on the routes selected from database 1 and transmitted to the generator. Variations in the possible driving variation processes for the same route are mapped using a random distribution of z. This is achieved through parameters. The generative model is parameterized. For example, the architecture of the generative model can be a recurrent neural network. The computer implementation of the generator is performed by storing the algorithm implementing the generative model and the parameters of the model in a machine-readable memory, by processing the computational steps of the algorithm by a processor and storing the generated driving variation process in a machine-readable memory.

[0063] In one possible configuration, the driving variation process can be generated with a fixed length, i.e. with a set number of discretization steps or data points. Then, when generating longer driving variation processes, multiple generated short time series have to be linked together. Here, however, the transitions are usually inconsistent. The method can be extended in an alternative configuration such that driving variation processes of any length can also be generated or evaluated and the transitions are always consistent with the learned transition model. For this purpose, the generator and the discriminator are preferably both implemented as recurrent neural networks, for example as long short-term memory (LSTM) neural networks or gated recurrent units (GRU). On the architecture, the generator is preferably implemented as a sequence-to-sequence model, but can also be implemented as a vector-to-sequence model. The discriminator is preferably implemented as a sequence-to-scalar model, but can also be implemented as a sequence-to-sequence model.

[0064] For the architecture of the generative model as a recurrent neural network, various options exist.

[0065] For example, a global random vector z can be sampled for the complete driving variation process, where the term "global" can again relate to the temporal or spatial discretization. In this configuration, properties are considered or learned in the latent space that globally change the driving variation process, for example properties that are constant during the route, such as constant driver properties (e.g. age or experience), constant weather properties (e.g. continuous rain) or constant vehicle properties (e.g. maneuverability). This random vector can now be used to initialize the hidden states in the first time step, or / and can be fed to the recurrent neural network at every time step.

[0066] Local or temporary random vectors z can also be sampled, i.e. properties are considered or learned in the latent space that locally or temporarily change the driving variation process, for example short-term properties such as short-term traffic management properties or traffic flow properties (state of traffic lights, crossroad congestion, pedestrians on the lane). Here, the random vector is regenerated at intervals of M time steps and fed to the recurrent neural network, with M > 0. M can also be random, i.e. the random vector can also be changed at random intervals.

[0067] In a preferred configuration, a combination of a global random vector and a local or temporary random vector can also be implemented. Here, some dimensions of the random vector are sampled only once per driving change process, the remaining dimensions are changed every M time steps. Alternatively, for this purpose, the global random vector can also be fed to the recurrent neural network at every time step, wherein the global random vector is replaced by a local, i.e. newly sampled, random vector every M time steps.

[0068] It turns out that a combination of a global random vector and a local random vector is particularly advantageous, since by this both global influences and local or temporary influences can be mapped in the data.

[0069] Preferably, also an expected or predicted driving can be taken into account in the generative model.

[0070] Thus, in one possible configuration, the route characteristic r t The route characteristic r t+1 ,..., r t+m This configuration is particularly advantageous in the case of online calculation, i.e. in the case of limited computing resources, or in the case where the influence of future route characteristics should or can be limited to a few discretization steps.

[0071] Alternatively, a bidirectional recurrent neural network can be used as a generative model, in which additionally the hidden states of future neurons of the recurrent neural network are taken into account. By this all possible future time steps can be explicitly included.

[0072] Thus, instead of adjusting the speed generation at time point t only on the basis of the route characteristic at time point t (and, if necessary, also on the basis of the hidden state at time point t-1), future route characteristics can also be used for the speed generation at time point t. Thus, in addition to the route characteristic at time point t, the route characteristics at time points t+1, t+2,..., t+m (or a subset thereof) are also used to adjust the speed generation at time point t. This makes it possible to simulate so-called "anticipatory driving", for example an early reaction of the driver to route characteristics (e.g. traffic lights, speed limits, highway exits, etc.) that are already visible in the distance, in particular in the sense of speed adaptation. It also enables the algorithm to learn to recover to a speed of 0 at the end of the route, for example by only having a default value (e.g. 0) for future route characteristics at the end of the route.

[0073] The route in the database 2 is stored in particular as a sequence of discrete data points, wherein for each data point or each discretization step the route characteristic in this discretization step is stored.

[0074] The routes r in the database 2 have e.g. a length S, e.g.: Each data point r t Corresponds to one discretization step. It is particularly preferred that the following implementations are realized: In these implementations the discretization step corresponds to a temporal or spatial discretization. In case of temporal discretization, the data points respectively correspond to the time elapsed since the beginning of the route, thus the sequence of data points corresponds to a time-varying process. In case of spatial discretization, the data points respectively correspond to the distance travelled along the route.

[0075] The sampling rate is typically constant. In case of temporal discretization, the sampling rate can e.g. be defined as x seconds, in case of spatial discretization e.g. as x meters.

[0076] Each data point r t Describes the route property at the respective discretization step, i.e. D is the number of route properties, wherein each dimension of a multi-dimensional route property counts as one dimension of a one-dimensional route property.

[0077] Such route properties can e.g. respectively relate to the discretization step, in particular a time point or time interval or a location or distance travelled:

[0078] • a geographical property like absolute height or slope

[0079] • a property of the traffic flow like traffic density or time-dependent average speed of traffic

[0080] • a property of the driving road like number of lanes, type of driving road or driving road curvature

[0081] • a property of the traffic management like speed limit, number of traffic lights or number of specific traffic signs, in particular stop or yield or pedestrian crossing

[0082] • a property of the climate like amount of rainfall, wind speed, presence of fog at a pre-given time point.

[0083] These are preferably the same type of route properties as stored in the first database for the routes.

[0084] The routes determined in this way are stored in the database 2 together with the driving variation processes measured in driving operation which belong to the routes. These pairs of route and belonging driving variation process serve as training data for the machine learning system. For this training, in particular pairs of route and belonging driving variation process are selected and transferred to the discriminator 42 in steps 23 and 24. In addition, pairs consisting of a route of the database 1 and a driving variation process generated by the generator 41 from this route are also transferred to the discriminator 42 in step 14 or 43.

[0085] The discriminator 42 has a computer-implemented algorithm with which a discrimination model is implemented. The discriminator 42 obtains as input variables pairs consisting of a route and a belonging driving variation process and decides whether the pair seen contains a driving variation process generated (by the generator 41) or a driving variation process measured (obtained from the database 2) in driving operation. The result of the decision is output in step 44. For example, the discriminator 42 can output a value > 0 for the decision "real driving variation process" and a value < 0 for the decision "generated driving variation process". Alternatively, for example, also a previously set value, like a class label, can be output. By means of parameters The discrimination model is parameterized. The output 44 of the decision in particular contains an evaluation by means of a binary decision "yes" / "no".

[0086] The computer implementation of the discriminator is carried out by storing the algorithm implementing the discrimination model and the parameters of the model in a machine-readable memory, by processing the calculation steps of the algorithm by a processor and storing the output in a machine-readable memory.

[0087] The discriminator 42 can be implemented as a recurrent neural network, for example. Thereby in particular driving variation processes of any length can be evaluated.

[0088] There are various configurations for the evaluation (decision of the generated driving variation process against the driving variation process trajectory determined in driving operation). In particular, the evaluation can be output anew after each individual time step. Then, a global evaluation of the driving variation process is the average value or the majority decision of the individual evaluations, for example. Alternatively, also only the evaluation of the entire driving variation process can be output at the last time step. The latter configuration in particular saves additional calculation steps and has the further advantage that the complete driving variation process is likewise introduced into the evaluation.

[0089] According to the output 44 of the discriminator 42, an objective function is optimized in block 5, in particular a loss function is minimized. For this purpose, the input variables of the discriminator are labeled as real samples, i.e. pairs having driving variation processes determined in a driving operation, or as generated samples, i.e. pairs having driving variation processes generated by the generator 41. The objective function here characterizes the extent to which the generated driving variation processes correspond to the actually measured driving variation processes, or the extent to which the distribution of the generated driving variation processes in the parameter space corresponds to the distribution of the measured driving variation processes in the parameter space. The parameters θ of the generator 41 are adapted according to the adaptation of the objective function G or the generated model implemented there and the parameters θ of the discriminator 42 are adapted D or the discriminative model implemented there. Here, the parameters are adapted, in particular with regard to the gradient of the objective function.

[0090] The objective function is chosen such that it characterizes or represents the difference or distance between the distribution of the generated driving variation processes and the distribution of the driving variation processes determined in a driving operation, or the difference or distance between the distribution of the route-driving variation process pairs having generated driving variation processes and the distribution of the route-driving variation process pairs having driving variation processes determined in a driving operation. By choosing such an objective function, the machine learning system can be trained such that the distribution of the generated data reflects the overall variation of the distribution of the measured data. A so-called mode collapse is prevented. That is, a representative distribution of the driving variation processes is provided. The objective function here takes into account the variation of unobservable influences, in particular, too.

[0091] For this purpose, a loss function (loss) is preferably chosen as the objective function, which is implemented as a Wasserstein metric or the Wasserstein distance between the distributions.

[0092] In this configuration of the computer-implemented training, the discriminator should preferably be limited to functions with a Lipschitz limit. For this purpose, in a preferred configuration, a regularization term is added to the objective function, for example a gradient penalty or (i) the gradient of the real samples, i.e. pairs having driving variation processes determined in a driving operation, is centered at 0 or (ii) the gradient of the generated samples, i.e. pairs having generated driving variation processes, is centered at 0 or (iii) the gradient of a sample representing the average of the real samples and the generated samples is centered at 1. Here, the option "the gradient of the real samples is centered at 0" is particularly preferred, as this option has been found to be the fastest option and can lead to particularly fast convergence of the optimization problem. Alternatively, a weight clipping can be carried out after each gradient step.

[0093] The above described method for computer-implemented training of the entire learning system comprising 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 in such a way that the generator 41 generates driving variation processes that mislead the discriminator 41 as much as possible.

[0094] In addition to the described input variables of the generator 41, further input variables can also be provided to the generator 41 in such a way that said further input variables are stored in the database 1, for example, for each driving route, respectively. For example, in addition to the discrete route information, information about the driver's characteristics (such as experience, age, aggressiveness or driving style, etc.) or vehicle information (such as power, motorization, type of drive system, etc.) can also be stored for each driving route. The generated driving variation processes can thus also be adjusted in accordance with these additional information. In this case, for the training of the machine learning system 4, the respective information is advantageously also stored for the driving routes stored in the database 2, respectively. These additional information are provided as input variables to the discriminator both for the route information from the database 1 (step 14) and for the route information from the database 2 (step 23).

[0095] The computer-implemented machine learning system taught using the described training method is able to generate representative driving variation processes for routes. For this purpose, the generator of a machine learning system thus taught can be fed with the same type of input variables as in the training, i.e. in particular the route, if necessary also the random variables and further information, and generates the respective driving variation process. The driving variation processes generated in this way can be used to simulate the emissions of a vehicle and thus, for example, to carry out a probability assessment of whether the exhaust standards are met and the emissions are optimized. The optimized emissions can be carried out, for example, by adapting the drive system in development, by optimizing the data application of the drive system (for example control devices for drive system control) or by adapting the control variables of the drive system in the vehicle to optimize the emissions. In the last case, in particular, the driving variation processes can be generated in the vehicle.

[0096] The driving variation processes can also be used to optimize the predicted driving, for example, in the battery management of electric vehicles or electric bicycles, in the drive management of hybrid vehicles, in the regeneration management of exhaust components of vehicles with internal combustion engines. This optimization can in turn be carried out in development, by optimizing the respective control device application or by adapting the control variables of the respective systems in the vehicle.

[0097] The driving variation course generated in this way can also make a valuable contribution to determining load spectra and load scenarios for specifying components, for example, what loads a particular component, for example a pump, will experience during its service life.

Claims

1. A computer-implemented method for training a machine learning system (4), the machine learning system being used to generate a vehicle's driving change process, the driving change process being a speed change process, an accelerator pedal position change process, or a transmission ratio change process, characterized by the following steps: Select a first driving route from the first database (1) containing driving routes. The generator (41) of the machine learning system (4) obtains the first driving route as an input variable and generates the corresponding first driving change process for each of the first driving routes. The second database (2) stores the driving routes and the driving changes detected during the driving process. Select the second driving route and the corresponding second driving change process detected during driving from the second database (2). The discriminator (42) of the machine learning system (4) obtains the pairing of one of the first driving routes with the first generated driving change process to which they belong, and the pairing of the second driving route with the second driving change process detected during driving as input variables. The discriminator (42) calculates an output based on the input variables, the output representing for each pair obtained as input variables whether it is a pair with a first generated driving change process or a pair with a second driving change process detected during driving operation. The objective function (5) is optimized based on the output of the discriminator (42), which represents the distance between the paired distributions having a first generated driving change process and the paired distributions having a second driving change process detected during driving operation. The objective function is a loss function, wherein the parameters of the generator (41) and the discriminator (42) are adapted to minimize the loss function when optimizing the objective function (5).

2. The method according to claim 1, characterized in that, The parameters of the generator (41) and the discriminator (42) are adapted according to the optimization of the objective function (5), so that... The discriminator (42) is optimized to distinguish the first generated driving change process from the second driving change process detected during driving operation. The generator (41) is optimized to generate a first generated driving route in a first distribution, the first generated driving route being as difficult as possible to distinguish by the discriminator (42) from a second driving change process detected during driving in a second distribution.

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

4. The method according to any one of the preceding claims, characterized in that, The statistical distance between the first distribution of the first driving change process and the second distribution of the second driving change process is realized as the objective function (5).

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

6. The method according to claim 5, characterized in that, The objective function (5) is extended with a regularization term or weights are pruned.

7. The method according to claim 1, characterized in that, The first and second travel routes are data stored in time or space discretization steps, where route characteristics are stored for each route in each discretization step.

8. The method according to claim 7, characterized in that, The route characteristics include geographical characteristics, traffic flow characteristics, road characteristics, traffic management characteristics, and / or the route's climatic characteristics.

9. The method according to any one of claims 7 or 8, characterized in that, The route characteristics of the first driving route are generated at least in part by the machine learning system.

10. The method according to claim 1, characterized in that, In addition to the first and second driving routes, the input variables of the generator (41) and the discriminator (42) also include additional information.

11. The method according to claim 1, characterized in that, The generator (41) obtains random variables from the random generator (3) as other input variables, and generates a first driving change process belonging to the first driving route based on the first driving route and the random variables.

12. The method according to claim 11, characterized in that, The random variable is implemented as a global random vector, or as a temporary or local random vector.

13. The method according to claim 11, characterized in that, The random variable is implemented as a combination of a global random vector and a temporary or local random vector.

14. The method according to any one of claims 1 to 3, characterized in that, The generator (41) and / or the discriminator (42) are respectively implemented as recurrent neural networks.

15. The method according to claim 14, characterized in that, The generator (41) is implemented as a bidirectional recurrent neural network.

16. The method according to any one of claims 1 to 3, characterized in that, In order to generate the driving change process by the generator (41), the route characteristics of the first driving route are supplemented or extended in a specific discretization step to the route characteristics of subsequent discretization steps.

17. The method according to claim 5, characterized in that, The objective function (5) is realized as the Wasserstein distance between the first distribution of the first driving change process and the second distribution of the second driving change process.

18. The method according to claim 9, characterized in that, The machine learning system is a neural network.

19. The method according to claim 10, characterized in that, The additional information includes driver characteristics and / or vehicle characteristics.

20. The method according to claim 16, characterized in that, In order to generate the driving change process by the generator (41), at a specific time point or at a specific distance, the route characteristics of the first driving route are supplemented or extended to the route characteristics of subsequent time points or subsequent distances.

21. A method for generating the driving change process of a vehicle, wherein, The driving change process is generated by a generator (41) implemented by a machine learning system (4), which is taught using the method according to any one of claims 1 to 20.

22. A method for evaluating the driving changes of a vehicle, wherein, The driving change process is evaluated by a discriminator (42) implemented by a machine learning system (4), which is taught using the method according to any one of claims 1 to 20.

23. A method for determining route-specific emissions of a vehicle's drive system, wherein the route-specific emissions are determined based on a driving variation process generated using the method of claim 21.

24. The method according to claim 23, characterized in that, The emissions are determined in the simulation, and a model of the drive system is calculated in the simulation.

25. The method according to claim 24, characterized in that, The model of the drive system includes sub-models describing the engine and the exhaust aftertreatment system of the drive system.

26. A method for adapting a drive system to a vehicle, characterized in that, The vehicle's drive system is adapted to route-specific emissions, which are determined using the method according to any one of claims 23 to 25.

27. The method according to claim 26, characterized in that, The adaptation is performed by adapting components or parameters during the development of the drive system, by adapting data during the application of the drive system, or by adapting control variables when the drive system is running in the vehicle.

28. A computer program product designed to perform the method according to any one of the preceding claims.

29. A machine-readable storage medium having a computer program product according to claim 28 stored thereon.

30. A computer-implemented machine learning system (4) having a machine-readable storage medium according to claim 29.

Images