EP3953908A1

EP3953908A1 - Method for determining a load prediction for a component of a motor vehicle

Info

Publication number: EP3953908A1
Application number: EP20719969.6A
Authority: EP
Inventors: Stéphane Foulard; Rafael Fietzek; Ousama ESBEL
Original assignee: Compredict GmbH
Current assignee: Compredict GmbH
Priority date: 2019-04-12
Filing date: 2020-04-09
Publication date: 2022-02-16
Also published as: CN113785333A; US12044598B2; US20220178789A1; LU101181B1; WO2020208182A1

Abstract

The invention relates to a method (1) for determining a load prediction for a component of a motor vehicle. In a model creation process (3), a load model is created using an identification method, based on a training journey data collective, a training vehicle data collective and a training load collective. The training journey data collective includes a plurality of training journey data sets, each including certain route course data and/or accompanying journey data for a training journey of a training motor vehicle on a training route. The training vehicle data collective comprises vehicle data of the training motor vehicle used on the respective training journey. The training load collective includes training load data. The training load data includes a load of the component corresponding with a training journey data set and measured or calculated on the basis of measurement variables. The load model estimates a probability distribution of the occurring load on a predefined training route and/or with predefined accompanying journey data and/or according to predefined vehicle data. In a model evaluation process (4), the load prediction is determined using the load model for a predefined prediction route and/or for predefined prediction accompanying journey data and/or according to predefined prediction vehicle data of a prediction vehicle.

Description

Method for determining a load forecast for a component of a motor vehicle

The invention relates to a method for determining a load prediction for a component of a motor vehicle.

When designing automotive components

usually a load to be expected for each motor vehicle component over a specified service life

determined and the respective component designed taking this load into account. It is known from the prior art to determine the expected load on the basis of

Training trips with motor vehicles

To determine the type of vehicle, the type of vehicle corresponding to the type for which the new component is to be designed. It is also known from the publication WO 2018/178 196 A1 that is used to determine the load

to determine the necessary information during the operation of an entire vehicle fleet. In particular with such a procedure, numerous different route courses with varying roadway properties and weather conditions as well as different driver properties can be determined for the determination of the load and when determining the load to be expected

be taken into account.

The methods known from the prior art are basically suitable for determining a specific load for a specific component, provided all of them

Model input variables are specified. In order to be able to simulate different loads, it is necessary to for example, to specify several speed profiles for a given route and to determine corresponding loads on this basis with the known load models.

To design vehicle parts, it is desirable to determine the load on the component during a typical journey. However, in reality, training drives on the basis of which the load models are created, even if they are carried out by a professional driver on one and the same training route, differ slightly from one another and lead to different load profiles and, accordingly, loads. In this case, the load and the load profile are subject to a probability distribution that the

Describes the probability of a certain load condition depending on the training vehicle used and the course of the training route. Such influences are not taken into account in the simulation of loads and load profiles in the methods known from the prior art.

The load is the total load such as B. a load spectrum that has acted on the component during a trip, while the

Load history around a time series of individual

Is load values that acted on the component while driving. The load can be determined from the load profile.

For the purposes of the following description of the invention, data that are required for teaching in the load model are used, denoted by the prefix "training". Data that are generated with the stress model carry the prefix "prediction".

It is seen as an object of the invention to further develop the methods known from the prior art for determining a load profile or a load.

According to the invention, this object is achieved by a method for determining a load prediction for a component of a motor vehicle, wherein in one

Model creation process starting from a

Training trip data collective, one

Training vehicle data collective, and one

Training load collective, a load model is created with an identification method, with the

Training trip data collectively a variety of

Contains training trip data sets that each contain route data and / or trip accompanying data for a training trip of a training motor vehicle on a training route, with the

Training vehicle data collective Vehicle data of the used in the respective training drive

Includes training vehicle, wherein the

Training load collective has training load data, the training load data each corresponding to a training trip data set

have metrologically recorded or calculated on the basis of measured variables load on the component, the load model having a probability distribution of the load that has occurred for a given one

Training route and / or given specified journey data and / or as a function of specified vehicle data

approximated so that the stress model is a model of the probability distribution, and where in a

Model evaluation process the load forecast with the help of the load model for a given

Prediction route and / or for predetermined

Prediction trip accompanying data and / or depending on predefined prediction vehicle data one

Prediction vehicle was determined. Advantageously, numerous in the model evaluation process

Simulations of the load model for a

Prediction vehicle carried out. This way you can better cover the scattering terms of use, so that you probably get a better picture of the in the end

received minimum and maximum loads occurring.

The modeling of the probability distribution represents the generation of non-deterministic or statistical models on the basis of training data that

capable of extrapolation with regard to new ones

Parameter combinations or sets or

Vehicle configurations are or are capable of extrapolation with regard to new routes, although there are no comparative measurements in both cases described. The term probability distribution plays a role here

Output variables play an important role because the trained model depicts the probabilities of occurrence of a certain load (e.g. the probability that a real load assumes a certain value at a certain point in time due to a current state). Specifically, it means that when simulating such a load model, every new simulation of the trained Model leads to a different result. For example, if the task is to simulate a time series, each simulation based on the same parameters would deliver different time series that deviate from one another but remain within a realistic framework due to the probability distribution (e.g. each

Model evaluation process, i.e. every simulation run of a trained model generates a different result or an individual time series, even if the

Parameter input remains the same). A real example of such behavior in a real system is the

Carrying out complete vehicle tests under driving specifications: Measurement drives differ fundamentally

from each other, as there are not very many influencing factors

can be influenced, even if the test drivers try to drive the maneuver as reproducibly as possible. This leads to an ordinary or usual and

unavoidable scatter. Exactly this problem is addressed with the approach according to the invention in which this

Scatter is modeled very specifically and then also simulated. This is an essential difference to common modeling approaches, which are static or

generate deterministic models.

If such a non-deterministic or statistical model has been generated, a large number of simulations with the same parameters can be carried out (e.g. 10,000 simulations) in order to map the complete scatter with good granularity, with each simulation being one different result leads and with all

Results put together the modeled

Map the probability distribution. In this way the range of possible system responses can be covered and more realistic results or

Simulations are generated.

The load model according to the invention therefore does not represent a model which - as usual modeling approaches - depicts reality deterministically “as well as possible”, but an AI-based data-driven model that does not reflect the spread of the load due to the

learns about measurable influencing variables and depicts reality with these undetectable uncertainties.

This approach cannot be used with regulatory approaches either

which can take into account uncertainties, but which can only work with feedback to the real system. In the approach according to the invention, only pure simulation models are used that learn from real systems, not previously

to be able to map existing systems. To this

Regulation systems are not capable of predictions.

Model accuracy and

Probability distribution of the model outputs are not confused. The model accuracy is that

Deviation between a model, which for each

Simulation delivers the same result with the same parameters and input variables, and one or more

Measurements. However, an assessment of the probability distribution is only possible if the model, like the load model according to the invention, delivers different results with the same parameters and input variables. Because real measurements are never 100% repeatable and always Show deviations, the procedure described offers a more comprehensive and realistic representation of the

Reality. Consequently, the method according to the invention can increase the accuracy of the load prediction by covering the possible uncertainties on the basis of the modeling of the probability distribution.

According to the invention it is advantageously provided that the load prediction is a

Stress history prediction, where the

Training load collective has training load time series, the training load time series each having a load profile of the component that is measured or calculated on the basis of measured variables and corresponds to a training drive data set, the load model approximating the probability distribution of the load profile that has occurred, and where in the model evaluation process the

Stress history prediction is determined. In which

In the model evaluation process, a large number of prediction load time series are advantageously determined, the individual load states calculated in each case for the prediction load time series again

Follow probability distribution. In this way, for example, you can easily move from each other

differentiating load profiles during training drives of a training driver with a predetermined one

Training vehicle can be simulated on a specific training route. The load profiles determined are time series. Since a model of the probability distribution of the The load curve, namely the load model used, differs from the simulated ones

Load profile predictions from one another, although the same input variables for the simulation, namely the prediction path and / or the

Prediction trip accompanying data and / or the

Prediction vehicle data of a prediction vehicle

be used.

The training drives are advantageously carried out as planned test drives. However, it is also possible and provided according to the invention that the

Training trips are normal road trips by any motor vehicle, with the data required for the modeling being recorded in the motor vehicle and, for example, wirelessly to a central server

be transmitted.

It is advantageously provided according to the invention that the training trip data sets are

Training trip time series. The different

Training trip time series can according to the invention

Time series sections of a single training run

represent.

The training drive data sets can be specified or

are recorded by measurement. For example, it is possible to determine the course of the route from an existing map of the training route and to specify it or to determine the course of the route during the training drive, for example using GPS. According to the invention, the load model can be determined on the basis of a probabilistic model structure. Various model structures are known from the prior art which approximate

Allow probability distributions.

The invention advantageously provides that the load model is based on a generative

Model structure is determined. In this way, the calculation effort for creating the load model can be reduced considerably. Completely visible Bayesian networks or a directional model such as a variational autoencoder are advantageously used as the model structure. With a particularly advantageous one

Design of the method according to the invention is

intended that as a model structure for the

Load model a recurrent conditional generative

Adverserial Network is used. The use of these model structures for probabilistic models are known to the person skilled in the art and are for example in Ian

Goodfellow, Yoshua Bengio, Aaron Courville: Deep Learning, mitp Verlag Frechen, ISBN: 9783958457003.

In order to check a model accuracy of the determined load model, it is provided according to the invention that

first the model evaluation process with a

predetermined prediction route and / or for predetermined prediction travel accompanying data and / or depending on predetermined prediction vehicle data one

Prediction vehicle multiple times, that is, for example 10,000 times, is performed. A classic deterministic model based on the simulated load predictions are determined, for example, using a neural network approach. The accuracy of this deterministic model can then be validated against real trip and load data, as known from classical deterministic modeling. The mean least squares error between the simulation and the real measurement data is often used for this purpose. Insofar as adequate accuracy can be achieved here, it can be assumed that the

Probability distribution of the totality of the

simulated load predictions of the real

Corresponds to the probability distribution of the load and the model quality of the load model is sufficient.

The method according to the invention can be used to calculate realistic loads and load profiles under specified conditions, that is to say for one

Route and a vehicle type to determine. In this way, for example, the

Probability distribution that affects the driving behavior of the training driver or the training driver

due to be mapped.

However, it is also possible and provided according to the invention that different training trip data sets

Include route data as well as trip accompanying data that were carried out under differing training conditions. It is advantageously possible to carry out training drives with different training vehicles of the same vehicle type on a training route

carry out, for example, by a

Manufacturing variance or other influences To determine the probability distribution for the load profile.

According to the invention it is provided that the different training conditions are that the

Training rides with at least two different

Training vehicles were carried out and / or on at least two different training routes

have been carried out and / or at least two

different environmental conditions. It is also possible and provided according to the invention that various test parameters are varied in order to obtain a generally valid model for the

Probability distribution among different

Determine conditions.

The vehicle data advantageously include a

Vehicle mass, a drag coefficient and a

Tire diameter of the training vehicle. The vehicle data are basically predetermined parameters of the training vehicle. Both

Vehicle data, on the other hand, are not parameters that change during the operation of the motor vehicle, such as a vehicle speed.

According to the invention, it is advantageously provided that the training vehicles are of a different vehicle type than the prediction vehicle. In this way, with the help of the load model, different

Load profiles are determined for different vehicle types that follow a realistic load distribution. In order to take into account a load dynamic in the load model, it is also provided according to the invention that a dynamic of a component load over the driven or predetermined route is taken into account in the load model.

Further advantageous embodiments of the method according to the invention are illustrated in the drawing

illustrated embodiment explained in more detail.

It shows:

Figure 1 is a schematically represented flow chart of the method according to the invention,

Figure 2 is a schematic representation of a

Load collective consisting of load time series and

Figure 3 is a schematic representation of a

Load collective plotted over cumulative

Load cycles.

In the method 1 shown schematically in FIG. 1, a large number of training drives by a single training driver or by different training drivers with a

Training vehicle or with various

Training vehicles carried out on a single training route or on different training routes and which occurred during the various training drives Load time series determined. In addition, for each

Training drive route data determined.

On the basis of the load time series collective formed from this time series data and

Collective training trip data and vehicle data from the training vehicle or training vehicles

comprehensive training vehicle data collective are in a model creation process 3 a

Probability distributions of the load profiles approximating load model determined. Based on this model, a

Model evaluation process 4 determines numerous loads and for a simulation, for example, a load on a certain component of a still in the

Motor vehicle under development can be used. The determined load profiles in turn follow those approximated with the model

Probability distributions of the

Training drives determined training data.

A load spectrum 5 is shown schematically in FIG. The load collective 5 consists of a large number of load time series that one

Follow probability distribution. In the drawing, a most likely load profile 6 and scatter range limits 7 are shown schematically.

FIG. 3 shows a schematic of a load spectrum 5. In the illustration, loads are plotted against cumulative load cycles. In the representation, larger loads occur less often than small loads. In addition, one is through the variation range limits 7 marked variation of the characteristics of large loads greater than the small loads.

Claims

PATENT CLAIMS

1. Method (1) for determining a load forecast for a component of a motor vehicle, wherein in one

Model creation process (3) starting from a

Training trip data collective, one

Training vehicle data collective, and one

Training trip data collectively a variety of

Includes training vehicle, wherein the

Training route and / or approximated with predetermined trip accompanying data and / or as a function of predetermined vehicle data, so that the load model is a model of the probability distribution, and in one

Model evaluation process (4) the load prediction with the help of the load model for a given

Prediction route and / or for predetermined Prediction trip accompanying data and / or depending on predefined prediction vehicle data one

Prediction vehicle is determined.

2. The method (1) according to claim 1, characterized in that the load forecast is a

Stress history prediction, where the

Training load collective has training load time series, the training load time series each having a load profile (6) of the component that corresponds to a training drive data set or that is calculated on the basis of measured variables, the load model the

Probability distribution of the occurred

Load curve (6) approximated, and in the model evaluation process (4) the

Stress history prediction is determined.

3. The method (1) according to claim 1 or claim 2, characterized in that it is the

Training trip data sets are training trip time series.

4. The method (1) according to one of the preceding claims, characterized in that the load model is determined on the basis of a generative model structure.

5. The method (1) according to claim 4, characterized in that a model structure for the load model

Recurrent Conditional Generative Adverserial Network is used.

6. The method (1) according to any one of the preceding claims, characterized in that different

Training trip data sets, route data and

Include trip-related data that were carried out under differing training conditions.

7. The method (1) according to claim 6, characterized in that the different training conditions consist in that the training trips were carried out with at least two different training vehicles and / or on at least two different ones

Training courses were carried out and / or at

at least two different environmental conditions

were carried out.

8. The method (1) according to any one of the preceding claims, characterized in that the vehicle data a

Include vehicle mass, a drag coefficient and a tire diameter of the training vehicle.

9. The method (1) according to any one of the preceding claims, characterized in that the training vehicles are of a different vehicle type than that

Prediction vehicle.