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

By generating non-deterministic load prediction based on the probability distribution model based on the training driving data, the problem of dispersed load curve simulation results in the prior art is solved, and more accurate and real load prediction is achieved.

CN113785333BActive Publication Date: 2024-12-06COMPEDI LTD
View PDF 3 Cites 0 Cited by

Patent Information

Application Number
CN202080028256.7
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Priority Date
2019-04-12
Filing Date
2020-04-09
Publication Date
2024-12-06
Estimated Expiration
2040-04-09

AI Technical Summary

Technical Problem

The prior art fails to effectively consider driver individual differences and path curve uncertainty when determining the load of motor vehicle components, resulting in dispersed load curve simulation results and lack of true and reliable probability distribution modeling.

Method used

By creating a probability distribution model based on training driving data, using recognition methods to generate load models, simulate predicted paths and vehicle data, perform multiple simulations to cover the possible load range, and use recursive conditional generation serial network and other model structures to generate non- Deterministic load forecast.

Benefits of technology

It improves the accuracy and authenticity of load prediction, can better map the impact of uncertainty in reality, and provides more comprehensive and credible load curve prediction results.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN113785333B_ABST
    Figure CN113785333B_ABST
Patent Text Reader

Abstract

The invention relates to a method (1) for determining a load forecast of 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 driving data set, a training vehicle data set, and a training load set. The training driving data set contains a large number of training driving data sets, each of which contains specific route curve data and / or driving accompanying data for a training vehicle on a training route. The training vehicle data set contains vehicle data of the training vehicle used in the corresponding training route. The training load data set contains training load data. The training load data respectively contain the load of the component acquired by measurement technology or the load of the component calculated based on the measured quantity corresponding to the training driving data set. The load model approximates the probability distribution of the load occurring under a predetermined training route and / or under predetermined driving accompanying data and / or based on predetermined vehicle data. In a model evaluation process (4), a load forecast is determined for a predetermined forecast route and / or for predetermined forecast accompanying data and / or based on predetermined forecast vehicle data of the forecast vehicle using the load model.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The invention relates to a method for determining a load forecast for a component of a motor vehicle. Background Art

[0002] When designing motor vehicle components, the loads to be expected within a predetermined service life are usually determined for each motor vehicle component, and the corresponding components are designed taking this load into account. It is known from the prior art to determine the expected loads based on training runs with motor vehicles of a motor vehicle type, wherein the motor vehicle type corresponds to the type for which the new component is to be designed. Furthermore, it is known from document WO 2018 / 178 196 A1 that the information required for determining the loads is determined during the operation of the entire motor vehicle fleet. In particular, in this method, a large number of different route curves with different lane characteristics and weather conditions as well as different driver characteristics can be determined for the determination of the loads and taken into account when determining the expected loads.

[0003] The methods known from the prior art are basically suitable for determining the actual load of a specific component, provided that all model input variables are predetermined. In order to be able to simulate different loads, it is necessary, for example, to predetermine a plurality of speed profiles for a predetermined path and on this basis determine the corresponding loads using known load models.

[0004] For the design of vehicle components, it is desirable to determine the loads of the components in typical driving profiles. However, in reality, even if training drives are performed by professional drivers on the same training route, the training routes based on which the load models are created differ slightly from one another and lead to different load profiles and, accordingly, to different loads. In this case, the loads and load profiles are subject to a probability distribution which describes the probability of a certain load state depending on the training vehicle and training route profile used. In the methods known from the prior art, such effects are not taken into account when simulating the loads and load profiles.

[0005] The load is the total load acting on a component during driving, such as a load spectrum, while the load curve is the time series of individual load values ​​acting on a component during driving. The load can be determined from the load curve.

[0006] In the context of the following description of the invention, data used to teach a load model are marked with the prefix “training.” Data generated using a load model have the prefix “prediction.” Summary of the invention

[0007] The object of the present invention is to further develop the method known from the prior art for determining a load curve or a load.

[0008] According to the invention, the object is achieved by a method for determining a load forecast of a component of a motor vehicle, wherein, in a model creation process, a load model is created using an identification method based on a training driving data set, a training vehicle data set, and a training load set, wherein the training driving data set includes a large number of training driving data groups, the training driving data groups each including specific route curve data and / or driving accompanying data for a training drive of a training motor vehicle on a training route, wherein the training vehicle data set includes vehicle data of the training motor vehicle used in the corresponding training drive, wherein the training load data set has training load data, wherein the training load data each have the load of the component acquired by measurement technology or the load of the component calculated based on the measured quantity corresponding to the training driving data group, wherein the load model approximates the probability distribution of the load occurring under a predetermined training route and / or under predetermined driving accompanying data and / or based on predetermined vehicle data, so that the load model is a model of a probability distribution, and wherein, in a model evaluation process, a load forecast is determined for a predetermined forecast route and / or for predetermined forecast accompanying data and / or based on predetermined forecast vehicle data of the forecast vehicle with the help of the load model. Advantageously, during the model evaluation process, the load model for the predicted vehicle is simulated several times, so that the dispersed usage conditions can be better covered, so that ultimately a better understanding of the possible minimum and maximum loads can be obtained.

[0009] The modeling of probability distributions represents the generation of non-deterministic or statistical models based on training data that are extrapolable for new parameter combinations or groups of parameter combinations or vehicle configurations or for new driving routes, wherein in both cases described there are no comparative measurements. Here, the term "probability distribution of output variables" plays an important role, since the trained model maps the probability of occurrence of a specific load (e.g. the probability that the actual load has a specific value at a specific point in time due to the current state). Specifically, when simulating such a load model, each new simulation of the trained model leads to different results. For example, if the task is to simulate a time series, each simulation based on the same parameters will provide different time series that differ from each other but remain within the realistic framework due to the probability distribution (e.g. each model evaluation process, i.e. each simulation run of the trained model, generates different results or separate time series, even if the parameter inputs remain the same). A real example of this behavior in a real system is the performance of a full vehicle test under driving discipline: here, the measurement runs differ in principle from each other, since very many influencing variables cannot have an influence, even if the test driver tries to drive in a reproducible manner as much as possible. This leads to a common and unavoidable dispersion. It is this problem that is solved by the method according to the invention, in which this dispersion is modeled quite specifically and then simulated. This is essentially different from the common modeling methods that generate static or deterministic models.

[0010] Once such a non-deterministic or statistical model is generated, a large number of simulations (e.g., 10,000 simulations) can be performed with the same parameters to represent the full dispersion at a fine granularity, where each simulation leads to a different result and where all the results together depict the modeled probability distribution. In this way, the bandwidth of possible system responses can be covered and more realistic results or simulations of the acoustic field can be obtained.

[0011] The load model according to the invention therefore does not represent a model that maps reality deterministically "as well as possible" (like usual modeling methods), but rather represents an artificial intelligence-based data-driven model that learns the dispersion of loads due to influencing variables that cannot be captured and maps reality with these uncertainties that cannot be captured.

[0012] This method is not yet comparable to a control method, which can take into account uncertainties but can only work with feedback from the real system. In the method according to the invention, only a pure simulation model is used that is learned from the real system in order to be able to map a system that does not yet exist. The control system is not capable of such predictions.

[0013] Furthermore, the accuracy of the model must not be confused with the probability distribution of the model output. The accuracy of the model refers to the deviation between the model and one or more measurements, and the model provides the same results in each simulation with the same parameters and input quantities. However, the evaluation of the probability distribution is only possible when a model, such as the load model according to the present invention, provides different results with the same parameters and input quantities. Since real measurements are never 100% repeatable and always show deviations, the described method provides a more comprehensive and more realistic mapping of reality. Therefore, the method according to the present invention can improve the accuracy of load forecasting by covering possible uncertainties of modeling based on probability distributions.

[0014] According to the invention, it is advantageously provided that the load forecast is a load curve forecast, wherein the training load set has training load time series, wherein the training load time series each have a load curve of a component corresponding to the training driving data set, acquired by measurement technology or calculated based on measured quantities, wherein the load model approximates the probability distribution of the occurring load curves, and wherein the load curve forecast is determined in a model evaluation process. In the model evaluation process, a plurality of forecast load time series are advantageously determined, wherein the individual load states respectively calculated in the forecast load time series again follow the probability distribution. In this way, for example, slightly different load curves can be simulated in the case of a training drive on a specific training route with a predetermined training vehicle. The determined load curves are each a time series. Since a model of the probability distribution of the load curves, i.e. a load model, is used for the simulation, the simulated load curve forecasts differ from each other, although the same input quantities, i.e. a forecast route and / or forecast driving accompanying data and / or forecast vehicle data of a forecast vehicle, are used for the simulation.

[0015] Advantageously, the training drive is carried out as a planned test drive. However, it can also be provided and provided according to the invention that the training drive is a normal road drive of any motor vehicle, wherein the data required for the modeling are collected in the motor vehicle and transmitted, for example wirelessly, to a central server.

[0016] Advantageously, according to the invention, it is provided that the training drive data set is a training drive time series. According to the invention, different training drive time series can represent time series parts of a single training drive.

[0017] The training driving data set can be predetermined or collected by measurement technology. For example, the route curve can be determined and predetermined according to a map of an existing training route, or the route curve can be determined during the training drive, for example, by GPS.

[0018] According to the invention, the load model can be determined based on a probabilistic model structure. Different model structures are known from the prior art, which allow the probability distribution to be approximated.

[0019] Advantageously, according to the invention, it is provided that the load model is determined based on the generated model structure. In this way, the amount of calculations for creating the load model can be significantly reduced. Advantageously, fully visible Bayesian networks or directional models, such as variational autoencoders, can be considered as model structures. In a particularly advantageous embodiment of the method according to the invention, it is provided that a recurrent conditional generative adverserial network is used as the model structure of the load model. The use of these model structures for probabilistic models is known to those skilled in the art and is described, for example, in the article by Ian Goodfellow, Yoshua Bengio, Aaron Courville: Deep Learning, mitp Verlag Frechen, ISBN: 9783958457003.

[0020] In order to check the model accuracy of the determined load model, it is provided according to the invention that the model evaluation process is first performed multiple times, i.e., for example, 10,000 times, with a predetermined forecast path and / or predetermined forecast driving data and / or based on predetermined forecast vehicle data of the forecast vehicle. Subsequently, a classical deterministic model is determined based on the simulated load forecast, for example, using a neural network method. Then, as is known from classical deterministic modeling, the accuracy of the deterministic model can be verified based on real driving and load data. For this purpose, the average minimum error between the simulation and the real measurement data is usually used. As long as a suitable accuracy can be achieved, it can be assumed that the overall probability distribution of the simulated load forecast corresponds to the real probability distribution of the load and that the model quality of the load model is sufficient.

[0021] The method according to the invention can be used to determine the actual load and load profile under predetermined conditions, ie for the route curve and the vehicle type. In this way, for example, a probability distribution of the driving behavior attributable to one or more training drivers can be mapped.

[0022] However, it can also be provided, and according to the invention, that different training drive data sets contain route curve data and travel companion data that were performed under mutually different training conditions. Thus, in an advantageous manner, training drives can be performed on training routes with different training vehicles of the same vehicle type in order to determine the probability distribution of load curves that are affected by manufacturing dispersion or other influences.

[0023] According to the invention, it is provided that the different training conditions include performing training runs with at least two different training vehicles and / or performing training runs on at least two different training routes and / or performing training runs under at least two different environmental conditions. It can also be provided, and it is provided according to the invention, that different test parameters are varied in order to determine a generally valid model of the probability distribution under the different conditions.

[0024] Advantageously, the vehicle data include the vehicle mass, the air resistance coefficient and the tire diameter of the training vehicle. The vehicle data are in principle predetermined parameters of the training vehicle. In contrast, the vehicle data are not parameters that change during operation of the vehicle, such as the vehicle speed.

[0025] Advantageously, according to the invention, training motor vehicles of different vehicle types are used as prediction vehicles. In this way, different load curves of different types of vehicles can be determined with the help of the load model, which load curves follow the real load distribution.

[0026] In order to take into account the load dynamics in the load model, it is further provided according to the invention that the dynamics of the component loads over the traveled or predetermined route profile are taken into account in the load model. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Further advantageous embodiments of the method according to the invention are explained in more detail with reference to the exemplary embodiments shown in the accompanying drawings. In the drawings:

[0028] Figure 1 shows a schematic flow chart of a method according to the present invention,

[0029] Figure 2 A schematic diagram showing a load set consisting of load time series, and

[0030] Figure 3 A schematic diagram of a set of loads plotted over accumulated load cycles is shown. DETAILED DESCRIPTION

[0031] exist Figure 1 In the method 1 schematically shown in FIG. 1 , firstly, in a test process 2, a single training driver or different training drivers perform a plurality of training runs with a training vehicle or with different training vehicles on a single training route or on different training routes, and the load time series occurring in the different training runs are determined. In addition, route curve data are determined for each training run.

[0032] Based on the load time series set formed from these time series data and the training travel data set, as well as the vehicle data set of the training vehicle including vehicle data of the training vehicle or the training motor vehicle, a load model that approximates the probability distribution of the load curve is determined in the model creation process 3. Based on this model, a plurality of loads can then be determined in the model evaluation process 4 and used to simulate the loads of specific components of a motor vehicle that is still under development, for example. In this case, the determined load curve again follows the probability distribution of the training data determined in the training travel that is approximated by the model.

[0033] exist Figure 2 A load set 5 is schematically shown in FIG. 5 . The load set 5 consists of a large number of load time series that follow a probability distribution. The most likely load curve 6 and the dispersion range limit 7 are schematically shown in the figure.

[0034] Figure 3 A load set 5 is schematically shown. In this illustration, the loads are plotted over accumulated load cycles. In this illustration, larger loads occur less frequently than smaller loads. Furthermore, the scattering behavior of larger loads, which is characterized by the dispersion range limit 7, is greater than that of smaller loads.

Claims

1. A method (1) for determining a load forecast for a component of a motor vehicle, wherein: In the model creation process (3), a load model is created using an identification method based on a training driving data set, a training vehicle data set and a training load set, wherein the training driving data set includes a large number of training driving data groups, and the training driving data groups respectively include specific path curve data and / or driving accompanying data for training driving of a training motor vehicle on a training path, wherein the training vehicle data set includes vehicle data of the training motor vehicle used in the corresponding training driving, wherein the training load set has training load data, wherein the training load data respectively have loads of components corresponding to the training driving data groups, collected by measurement technology, or calculated based on measurement quantities, wherein the load model approximates the probability distribution of loads occurring under a predetermined training path and / or under predetermined driving accompanying data and / or based on predetermined vehicle data, so that the load model is a model of probability distribution, and wherein, in the model evaluation process (4), a load prediction is determined for a predetermined predicted path and / or for predetermined predicted driving accompanying data and / or based on predetermined predicted vehicle data of a predicted vehicle with the help of the load model.

2. The method (1) according to claim 1, characterized in that: The load prediction is a load curve prediction, wherein the training load set has a training load time series, wherein the training load time series respectively has a load curve (6) of a component corresponding to the training driving data set, which is collected by measurement technology or calculated based on measurement quantities, wherein the load model approximates the probability distribution of the occurring load curve, and wherein the load curve prediction is determined in the model evaluation process (4).

3. The method (1) according to claim 1 or claim 2, characterized in that: The training driving data set is a training driving time series.

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

5. The method (1) according to claim 4, characterized in that: A recursive conditional generative adversarial network is used as the model structure of the load model.

6. The method (1) according to any one of the preceding claims, characterized in that Different training travel data sets contain route curve data and travel accompanying data performed under training conditions different from each other.

7. The method (1) according to claim 6, characterized in that The different training conditions include performing the training run with at least two different training vehicles and / or performing the training run on at least two different training routes and / or performing the training run under at least two different environmental conditions.

8. The method (1) according to any one of the preceding claims, characterized in that The vehicle data includes the vehicle mass, air resistance coefficient and tire diameter of the training motor vehicle.

9. The method (1) according to any one of the preceding claims, characterized in that The training vehicles of different vehicle types serve as the predicted vehicles.

Citation Information

Patent Citations

  • Method for determining a degree-of-damage uncertainty of a motor vehicle

    WO2018178196A1

  • Vehicle testing system

    US20150338313A1

  • System and method for load-based structural health monitoring of a dynamical system

    US20180275044A1