Simulation-based parameter presetting for dynamic driving functions

Abstract

The invention relates to a method for establishing a set of functional parameters of a vehicle brake control device. In order to reduce the number of test drives, different values are selected for at least a subset of the functional parameters and, by means of a simulation platform, driving-dynamic measurement results for predetermined driving manoeuvres and environmental settings are simulated, which are evaluated by means of an objectification algorithm with at least one key characteristic number. Further, using the respective values with associated key characteristic numbers, a metamodel is trained, which therefore indicates a relationship between the respective subset of functional parameters and the key characteristic numbers. This metamodel is used in the parameter optimisation proper by means of an optimisation algorithm in order to find optimum parameter values.

Description

[0001] Simulation-based parameter presetting for dynamic driving functions

[0002] The invention relates to a method for controlling a vehicle brake by means of a vehicle brake control unit based on a set of functional parameters, as well as an associated method for determining the set of functional parameters of the vehicle brake control unit, a corresponding computer program product, and a manufacturing method.

[0003] Modern motor vehicles are equipped with driving dynamics controllers that primarily contribute to a safe condition in dynamic driving situations. A driving dynamics controller contains various individual functions, each responsible for specific driving situations (e.g., the anti-lock braking system (ABS) during braking or the electronic stability control (ESC) during dynamic cornering) and specifically influences driving behavior in terms of driving safety, driving comfort, or performance. These functions have variable parameters ("function parameters") that are individually adjusted for each vehicle model so that the driving dynamics controller works as optimally as possible and the driving behavior requirements are met. To date, it has only been possible to determine these parameters using simulation and computer support.These parameters are determined using objective criteria based on simulated or real-world driving behavior (Bechtloff, Jakob Philipp: “Estimation of the sideslip angle and driving dynamic parameters for improving model-based driving dynamics control,” 2017, Technical University of Darmstadt, Darmstadt). An example of these objectively determined function parameters are the parameters of the single-track model within the ESC function. In addition, there are function parameters that can be applied based on geometric and technical data of the vehicle and the braking system without real-world driving tests or the use of simulation (so-called “desktop calibration”). These parameters represent, for example, the vehicle mass or the coordinates of the vehicle's center of gravity.The majority of functional parameters for driving dynamics functions are currently largely applied in road tests, which involve comparatively high effort and involve real vehicles on various test tracks and under different environmental conditions. When applying these parameters, in addition to objective parameters, additional subjective evaluation criteria of the driver are also used. The driver can use their experience to evaluate the overall driving behavior and the overall impression of the vehicle and its surroundings and apply the parameters to the overall driving behavior, consisting of objective and subjective criteria. In doing so, the driver must consider and individually adjust physically determined compromises between various criteria (e.g., safety vs. performance).

[0004] There are several disadvantages to applying functional parameters based on a real vehicle: On the one hand, the traditional application of driving dynamics parameters is very time-consuming and costly, as it initially requires expensive prototype vehicles that must be laboriously manufactured in small quantities and equipped with measurement technology. These vehicles are then transported and used on various test tracks that meet different environmental conditions as part of the parameter application. Different environmental conditions, for example, can represent different friction coefficients of the road surface, for which specific functional parameters must be applied. The availability of the vehicles and the dependence on the environmental conditions require precise project planning and increase the effort (time and costs) within a project.Added to this are the time and organizational efforts required by the test drivers. On the other hand, real prototype vehicles are essential for traditional calibration, as the calibration of the driving-relevant parameters takes place using the real vehicle. Real vehicles are necessary because the complex overall driving behavior, the interaction between the vehicle and the vehicle dynamics controller installed within it, is evaluated holistically, objectively, and subjectively by the calibration engineer. At the same time, real vehicles enable the integration of the experienced driver (calibration engineer), who can subjectively evaluate the influence of the parameters on the resulting overall dynamic driving behavior.

[0005] The objective of this invention is to develop a method that allows the application effort of the functional parameters to be significantly reduced.

[0006] The problem is solved by a method, in particular a computer-implemented method, for defining a set of functional parameters of a vehicle brake control unit. Different values ​​are selected for at least a subset of the functional parameters, and driving dynamics variables are simulated using a simulation platform for predetermined driving maneuvers and environmental settings, which correspond to measurement results during real test drives. In other words, a vehicle with a vehicle brake control unit is simulated in the simulation platform, the functional parameters of which are set, i.e., adjusted, to these values. Furthermore, the settings for the environmental conditions, i.e., the environmental settings of the simulation platform, are set, and simulations for various driving maneuvers are carried out. The simulation then provides the driving dynamics variables, such as, for example, lateral accelerations, longitudinal accelerations, and many others.These are sent to an objectification algorithm, which evaluates these variables and outputs at least one key figure as a rating. In combination with the simulation platform and the objectification algorithm, a rating in the form of a key figure is assigned to specific values ​​of the function parameters. A metamodel, in particular an artificial neural network, is then trained to take on precisely this task by being taught training data. The respective values ​​of the function parameters with the corresponding key figures are used as training data. The trained metamodel thus specifies a relationship between the respective subset of function parameters and the key figures. The metamodel can now be used for the actual parameter optimization, i.e. for finding the best values ​​for the function parameters using an optimization algorithm.

[0007] The method according to the invention, hereinafter referred to as "SIPA" (Simulation-Based Pre-Adjustment), can thus eliminate the significant disadvantages of conventional parameter calibration, in particular the complexity and availability of prototype vehicles for vehicle dynamics controllers. The simulation-based approach inherent in the method eliminates dependencies on both the environmental conditions and the prototype vehicles. The environmental conditions and the prototype vehicles are digitally modeled in the simulation according to the vehicle's requirements and technical data, allowing these virtual models to be simulated in conjunction with a virtual vehicle dynamics controller, and the functional parameters to be calibrated using the simulated driving behavior in a virtual environment.

[0008] In a preferred embodiment of the invention, the objectification algorithm carries out the evaluation based on several types of key performance indicators, and a separate metamodel is trained for each type of key performance indicator, which relates the respective functional parameters to the corresponding type of key performance indicator.

[0009] In a preferred embodiment of the invention, the selection of the values ​​for at least the subset of the functional parameters is carried out by means of a design of experiments unit, in particular according to the Latin hypercube sample method.

[0010] In a preferred embodiment of the invention, the backpropagation approach is used to train the metamodel. Backpropagation, also known as backpropagation of error, is a preferred method for training artificial neural networks. It belongs to the group of supervised learning methods and is applied to multilayer networks as a generalization of the delta rule. In a preferred embodiment of the invention, a genetic algorithm is used as the optimization algorithm.

[0011] In a preferred embodiment of the invention, the optimization algorithm takes into account the different metamodels for the different types of key performance indicators, in particular weighted ones.

[0012] In a preferred embodiment of the invention, at least one additional metamodel is trained for a further subset of functional parameters for other driving maneuvers and / or environmental settings. This is then also used by the optimization algorithm. The driving maneuvers, with their different environmental conditions, can be weighted to favor certain properties (e.g., stability) under specific environmental conditions (e.g., on snow / low road friction).

[0013] In a preferred embodiment of the invention, the subset of functional parameters comprises general parameters and environment- and / or maneuver-dependent parameters, wherein the further subset of functional parameters comprises the same general parameters and other environment- and / or maneuver-dependent parameters.

[0014] In a preferred embodiment of the invention, the total set of parameters of the set of functional parameters of the vehicle brake control unit is divided into subsets, with optimal parameter values ​​for the subsets being found sequentially. For each of these functional parameters of the subset, individual driving maneuvers and environmental conditions are assigned in which the functional parameters have an influence, in particular an above-average influence, on the driving dynamics measurement results. The object is further achieved by a method for controlling a vehicle brake by means of a vehicle brake control unit based on a set of functional parameters that are defined by an above-mentioned method and stored in the vehicle brake control unit. Control commands for the vehicle brake are generated based on the stored functional parameters.

[0015] The object is also achieved by a computer program product comprising instructions which, when the program is executed by a computer, cause the computer to carry out one of the above methods.

[0016] The object is also achieved by a manufacturing method for a motor vehicle brake system comprising a vehicle brake control unit which comprises program code for executing vehicle functions, which contains function parameters, wherein values ​​for the function parameters which were determined by means of one of the above methods are stored in the control unit.

[0017] SIPA enables the simulation-based calibration of selected functional parameters of the vehicle dynamics controller in a virtual environment. As shown in Fig. 1, the method is divided into a "preprocessing process" and a "parameter application process" and builds on an existing approach (Rot, Ivan: "Method for model-based calibration of the shift sequence control of transmission control units in a virtual environment," 2017, Technical University of Darmstadt, Darmstadt).

[0018] Within the "preprocessing process," a simulation platform is used that contains the vehicle, the vehicle environment (road, route), the virtual driver, and the vehicle dynamics controller on which the functional software to be applied runs. The environmental conditions and vehicle characteristics can be adapted within the simulation platform, ensuring application coverage taking into account different environmental conditions. The simulation platform must represent all relevant vehicle parameters and driving conditions. Due to its complexity, the simulation environment requires a large amount of computing resources. This negatively impacts its applicability in the subsequent "parameter application process," as this requires a high number of simulation runs (iterations) as part of the parameter optimization (application).Therefore, the goal of the "preprocessing process" is to first derive a simplified metamodel (hereinafter referred to as "training") using the complex simulation environment. This metamodel can represent the relevant properties for parameter application and requires fewer computing resources during execution. This less complex and comprehensive metamodel can then be used in the "parameter application process" in a resource-efficient manner.

[0019] To train the metamodel, the driving behavior simulated using the simulation environment is evaluated, and specific information is extracted from it that comprehensively describes the driving behavior. First, a test plan is created using a Design of Experiments (DoE) approach. This test plan adjusts the functional parameters to be optimized for each test, for example, using the Latin Hypercube Sample method, so that the system generates the greatest possible information gain with the least possible effort (tests). The tests are carried out one after the other using the simulation platform for a specified driving maneuver, and the generated measurement data is then evaluated using an objectification algorithm, for example, as described in DE 10 2019 217 431 A1.Based on the measured driving behavior measurements, the objectification algorithm calculates objective rating models that correspond to a subjective assessment of a real driver for the driving maneuver under consideration. In addition to calculating the subjective ratings, the objectification algorithm also calculates maneuver-specific objective criteria, which are used together with the subjective rating models to evaluate the driving maneuvers. The subjective rating models and the objective criteria are summarized and referred to below as KPIs (Key Performance Indicators). Based on the functional parameters known from the DoE and the KPIs determined from the simulated driving tests, the metamodel is trained in the final step of the "preprocessing process." Various approaches can be used to train the metamodel, such as the backpropagation algorithm from the field of machine learning methods.The goal of the trained metamodel is to efficiently calculate a mathematical relationship between the functional parameters (input of the metamodel P) and the KPIs (output of the metamodel y_KNN). A separate metamodel with individual model parameters is implemented to calculate each KPI. The metamodel can represent any mathematical model that can calculate the functional behavior of an output depending on the model inputs. An example is an artificial neural network (KNN) in the form of a multilayer perceptron (MLP) with one hidden layer, which fulfills the above-mentioned properties: with

[0020] Input: P: [p , ...,p e

[0021] Offset: b

[0022] Activation function: f KNN

[0023] Weight: g

[0024] Number of input parameters: n

[0025] After the metamodels have been trained, they are applied within the "parameter application process" for calculating the KPIs depending on the function parameters. The function parameters P are specified by the optimization algorithm. In SIPA, an optimization algorithm is used that is able to find a global optimum taking into account several function parameters to be optimized (applied) for a nonlinear system behavior. For example, the genetic algorithm (GA) fulfills these properties and can be used for the application (optimization) of the function parameters [6]. The optimization algorithm has the task of determining the objective function / ZF (P) to minimize: min ZF (P), P e D with and number of KPIs: m. Depending on the application, other, even local, optimization algorithms can also be applied.

[0026] The objective function depends on the KPIs calculated by the metamodels (y KNN ). Within the objective function, weightings (w) are defined for each KPI, which can be specified by the user in order to specifically influence the characteristics of the driving behavior. For example, the weighting for a KPI relevant to driving dynamics stability can be set so that the function parameters assume certain values, thereby ensuring that the vehicle has the highest possible stability in the driving maneuver under consideration. The objective function is iteratively minimized by the optimization algorithm until one or more defined optimality criteria are met. An optimality criterion can, for example, be defined as the convergence of the objective function to an undetermined value as the iteration progresses.

[0027] Certain functional parameters depend on specific driving situations (e.g., driving speed) or environmental conditions (e.g., the coefficient of friction p of the road-tire contact). Other functional parameters, however, are general and not specific to any specific situation. Within a test drive, a driving and environmental scenario can always be defined and simulated or run. However, the general parameters must be applied to all possible driving and environmental scenarios. In reality, these parameters are adjusted iteratively and based on experience. In SIPA, this situation is addressed using the following approach:

[0028] To explain the approach, three environments (alternatively driving states) A, B and C are considered as examples. For the application, there are parameters W that are valid for each environment. In addition, there are parameters A, B and C that are only valid within the environment of the same name A, B and C. In order to find a cross-environment parameter set for W, A, B and C that meets the required driving dynamics requirements, the present approach first conducts driving tests (simulated) for each environment individually according to the "preprocessing process" and then trains metamodels with the associated parameters (see Figure 2). Thus, each metamodel contains information about a specific environment, which can calculate the KPIs depending on the environment-independent and the specific environment-dependent parameters.A universally valid (cross-environment) application of the considered parameters is achieved by applying the learned metamodels in parallel within the subsequent "parameter application process." In each iteration step, the optimization algorithm calculates the complete input parameter set (A, B, C, W), which is then divided into input variables for the respective metamodels (see Figure 2 "Calculation for a KPI"). The substitute KPI is then calculated. calculated for all considered environments: with number of environments: k KPl value within an environment: y KNN.Umg Weighting of an environment: w Umg .

[0029] With the weighting w UmgThe KPIs of individual environments can be specifically weighted. This allows for optionally preferential driving behavior for a specific environment while considering the overall driving behavior. The objective function value is calculated taking all KPIs and KPI weightings (w) into account as follows:

[0030] Figure 2 illustrates the approach schematically. A function can contain several dozen or hundreds of function parameters that influence driving behavior and are to be applied accordingly to the individual vehicle. Within a function, these parameters determine the driving behavior for a total of several driving situations and environmental conditions. Attempting to consider all parameters in a single process step (driving maneuver) proves to be very complex from a mathematical, technical, and therefore practical perspective. Training a metamodel that has several hundred parameters as input variables means creating a large number of simulation variables that are used to train the metamodel. The need for training data for training a metamodel increases exponentially with the number of its input variables. This leads to very high computing resources that are currently largely unavailable in practice.A further problem arises from the fact that a driving maneuver would have to be performed very extensively and comprehensively to cover all driving situations influenced by the parameters to be calibrated. This is illustrated by an example in Figure 3 under "Calibration approach without calibration strategy." Accordingly, all relevant functional parameters in a complex driving maneuver would have to be adjusted in a single calibration step to best meet the requirements for all sub-events within this maneuver. Mathematically, each parameter represents a degree of freedom in which the optimal value must be found.

[0031] This complex problem is solved within the scope of this invention by implementing and applying an application strategy, which represents a particularly preferred variant of the method according to the invention. The application strategy is defined and implemented individually for the application of each function. The essential approach of the application strategy is the division of a complex problem into smaller, less complex sub-problems and sub-steps. This creates a practical and computationally resource-efficient process for the virtual application of a function's function parameters.

[0032] The calibration strategy defines a process individually for each function, which provides a sequence in which certain parameters (collectively referred to as a parameter group) are applied using a driving maneuver specifically defined for the application of this parameter group. This driving maneuver can be divided into one or more driving maneuvers with different environments (environmental conditions).

[0033] After the function parameters have been applied within the first step of the defined sequence, the next parameter group with the next associated driving maneuver is applied based on these then applied parameters. These steps are performed sequentially until all planned parameter groups have been applied within the calibration strategy. This approach takes into account any potential interdependencies between the parameters. One goal of the calibration strategy is to reduce the number of parameters in a parameter group as much as possible within a calibration process in order to reduce the problem as much as possible from the perspective of computing resources and to find a suitable parameter set (calibration = parameter optimization) (in other words, to divide the complex problem into as many small sub-problems as possible).

[0034] Figure 3 shows the essential features of the application approach with application strategy schematically

[0035] The special feature of the SIPA method is that, within the virtual calibration process, the driving characteristics are individually considered through targeted weighting of the KP Is, and the functional parameters can be automatically applied according to the weighting. Furthermore, different driving and environmental conditions can be considered simultaneously within the calibration process, ensuring that the applied functional parameters are optimally designed for all considered driving and environmental conditions.

[0036] The application strategy allows the complex application of many function parameters to be reduced to several small sub-problems in order to reduce the overall computing resources and the effort required to find suitable parameter application processes.

[0037] The method represents a generic approach and can generally be applied to the calibration of the functional parameters of all quasi-static and dynamic driving functions. For example, the method can be applied to other (dynamic or quasi-static) driving functions with appropriate adjustment of the KPIs and calibration strategy. One example is the calibration of the parameters of the ABS (anti-lock braking system) or ACC (adaptive cruise control) functions.

[0038] In the following example, SIPA is applied to a virtual application for four function parameters of the yaw control function (Active Yaw Control Function). Thus, an exemplary parameter group with an associated driving maneuver within the application strategy is considered. A sine dwell driving maneuver is considered, in which the vehicle is tested in a total of three independent environments with a road-tire friction coefficient of = 0,3, |i ß = 0,65 und |i c= 1.0 under otherwise identical conditions, the yaw control function deliberately causes the vehicle to oversteer, and the yaw control function returns the vehicle to a stable driving state as a result of the oversteer. To exclude the influence of other potential control functions, all other driving dynamics functions are deactivated so that only the control function to be applied for vehicle oversteer is active and influenced by the parameters to be applied. In the first step, a plan of experiment (DoE) is created in the "preprocessing process," and the maneuver is individually and virtually performed for each environment within the simulation platform using different parameter combinations.The simulation variables, which are also referred to as measurement signals because they correspond to real-world measurement variables, are then evaluated using the objectification algorithm, and the KPIs relevant for describing the driving dynamics quality - "performance," "stability," "handling," and "comfort" - are determined. In the last step of the "preprocessing process," the metamodels are trained for each environment and each KPI (a total of 12 metamodels in this example). The quality of the metamodels in this example is shown in Figures 4, 5, and 6. To demonstrate quality, the KPIs calculated using the trained metamodels are compared with the KPIs determined directly based on the measurement signals (coming from the simulation platform) using the objectification algorithm. The closer the points (one point corresponds to the evaluation of a driving maneuver) are to the diagonal line, the higher the prediction accuracy of the metamodel.The quality of the metamodels considered can be described as good overall.

[0039] In this example, two different optimization variants are implemented in the "parameter application process." Optimization variant 1 focuses on stable driving behavior, which, in the case of an oversteer maneuver, is equivalent to a smaller permissible sideslip angle that the vehicle develops during the maneuver. The KPI weightings used in the optimization for both optimization variants are shown in Table 1. For simplicity, the weightings of the three environments are set equal: w Umg = w UmgB = w Umgc . Table 1

[0040] The optimization results achieved for all three environments are presented in normalized form in Figures 7, 8, and 9 below. The results demonstrate that the use case specifications were generally implemented as expected via the weightings. In the example for optimization 1, the focus was on stable driving behavior. The environment-dependent and environment-independent parameters were applied in such a way that very high stability was achieved in all three environments (a normalized grade of 1 means that the maximum possible KPI grade was achieved under the given conditions).

[0041] In the example already presented, calibration results for the yaw control function (AYC function) for two different dynamic vehicle characteristics were explained. As already mentioned in the description of the method, the method can be applied to any other driving dynamics functions. To apply the method to a different function, it is first necessary to define the function parameters to be applied and to implement the relevant dynamic driving maneuvers in the simulation. The method can then be applied to the new calibration problem in a similar way to the example shown. Another example is the calibration of parameters of the ABS function (anti-lock braking system), in which, for example, the parameters influencing braking distance and stability are applied during straight-line braking or during so-called mue-split braking (braking on parallel, different friction coefficients).

Claims

Patent claims 1 . Method for determining a set of functional parameters of a vehicle brake control unit, wherein different values ​​are selected for at least a subset of the functional parameters and driving dynamics variables are simulated by means of a simulation platform for predetermined driving maneuvers and environmental settings, which are evaluated by means of an objectification algorithm with at least one key characteristic figure, and wherein a metamodel is trained by means of these respective values ​​of the functional parameters with the associated key characteristic figures, which thus indicates a relationship between the respective subset of functional parameters and the key characteristic figures, wherein the metamodel is used in the actual parameter optimization by means of an optimization algorithm to find optimal parameter values.

2. Method according to claim 1, characterized in that the objectification algorithm carries out the evaluation on the basis of several types of key figures and for each type of key figure a separate metamodel is learned which relates the respective functional parameters to the corresponding type of key figure.

3. Method according to one of the preceding claims, characterized in that the selection of the values ​​for at least the subset of the functional parameters is carried out by means of a design of experiments unit, in particular according to the Latin hypercube sample method.

4. Method according to one of the preceding claims, characterized in that the backpropagation approach is used to train the metamodel 5. Method according to one of the preceding claims, characterized in that a genetic algorithm is used as the optimization algorithm 6. Method according to one of the preceding claims, characterized in that the optimization algorithm takes into account the different metamodels for the different types of key figures, in particular in a weighted manner.

7. Method according to one of the preceding claims, characterized in that at least one further metamodel is learned for a further subset of functional parameters for other driving maneuvers and / or environmental settings.

8. Method according to one of the preceding claims, characterized in that the subset of functional parameters comprises general parameters and environment- and / or maneuver-dependent parameters, wherein the further subset of functional parameters comprises the same general parameters and other environment- and / or maneuver-dependent parameters.

9. Method according to one of the preceding claims, characterized in that the total set of parameters of the set of functional parameters of the vehicle brake control unit is divided into subsets, wherein the optimal parameter values ​​for the subsets are found one after the other.

10. Method for controlling a vehicle brake by means of a vehicle brake control unit based on a set of Functional parameters which are determined by a method according to claims 1 to 9.

11. A computer program product comprising instructions which, when executed by a computer, cause the computer to perform one of the above methods.

12. Manufacturing method for a motor vehicle brake system comprising a vehicle brake control unit which comprises program code for executing vehicle functions, which contains functional parameters, characterized in that values ​​for the functional parameters are stored in the control unit which were determined by means of one of the methods according to claims 1 to 9.

Images