Method for optimising a functional parameter in order to set an operating variable of a technical system

Abstract

The disclosure comprises a method (100) for optimising at least one functional parameter in order to set an operating variable of a technical system (96), said method comprising the following steps: - defining (102) at least one first initial functional parameter (4) for a first iteration step; - determining (104) at least one actual operating variable (93) by supplying the at least first functional parameter (4) to the technical system (96), - generating (106) a second functional parameter (6), which corresponds to the first actual operating variable (93), by supplying the determined first actual operating parameter (93) as an input variable to an ML model (20) in order to pre-train the ML model (20); - generating (108) a third functional parameter (8), which corresponds to a specified target operating variable (94), by supplying the specified target operating variable (94) to the pre-trained ML model (20), the ML model (20) being intended to be trained to said target operating variable; - determining (110) an auxiliary parameter (32) for an optimiser (30) as an input variable for the optimiser (30), wherein the auxiliary parameter (32) is determined from the first actual operating variable (93) and the specified target operating variable (94), - determining (112) a fourth functional parameter (10) by supplying the determined auxiliary variable (32) to the optimiser (30), - supplying (114) the third functional parameter (8), output by the ML model (20), and the fourth functional parameter (10), output by the optimiser (30), as new input variables for the technical system (96) for at least one next iteration step in order to find the optimised functional parameter for the operating variable, to be controlled, of the technical system (96).

Description

[0001] Description

[0002] Title:

[0003] Method for optimizing a functional parameter to adjust an operating variable of a technical system

[0004] The present invention relates to a method for optimizing a functional parameter in order to adjust an operating variable of a technical system.

[0005] State of the art

[0006] Various methods are known in the prior art for determining functional parameters for a control unit in order to adjust or control an operating variable for a technical system, such as a plant or a vehicle. DE 10 2009 054 905 A1 discloses an example of such a method.

[0007] In general, these known methods are usually based on the use of different optimizers whose goal is to optimize, i.e., either maximize or minimize, one or more target variables or target operating variables. Examples related to a vehicle might be a minimum braking distance or an optimally adjusted wheel speed.

[0008] For high-dimensional target variables, such as the desired temporal progression of an operating variable, common methods always require the creation of one or more auxiliary variables. For this purpose, for example, the mean square error between the desired measured or operating variable over time and the measured variable resulting from each iteration over time is calculated. The optimizer used then aims to iteratively determine the optimal or optimized function parameters by minimizing this square error.

[0009] It is therefore the object of the present invention to provide a solution by means of which one or more functional parameters for setting an operating variable can be determined or optimized in an efficient and reliable manner.

[0010] Disclosure of the invention

[0011] This object is achieved by a method for optimizing at least one functional parameter in order to adjust an operating variable of a technical system, having the features of the independent claim.

[0012] According to a first aspect, the disclosure relates to a method for optimizing at least one functional parameter in order to adjust an operating variable of a technical system, comprising the following steps:

[0013] In a first step, at least one first initial function parameter is defined for a first iteration step.

[0014] In a second step, at least one actual operating variable is determined by feeding at least the first functional parameter to the technical system.

[0015] In a third step, a second function parameter is generated which corresponds to the first actual operating variable by feeding the determined first actual operating variable to a machine learning ML model as an input variable in order to pre-train the ML model.

[0016] In a fourth step, a third function parameter is generated, which corresponds to a specified target operating variable, by feeding the pre-trained ML model the specified target operating variable on which the ML model is to be trained. In a fifth step, an auxiliary parameter is determined for an optimizer as an input variable for the optimizer. The auxiliary parameter is determined from the first actual operating variable and the specified target operating variable.

[0017] In a sixth step, a fourth function parameter is determined by feeding the determined auxiliary variable to the optimizer.

[0018] In a seventh step, the third function parameter output by the ML model and the fourth function parameter output by the optimizer are fed as a new input variable for the technical system for at least one next iteration step in order to find the optimized function parameter for the operating variable of the technical system to be controlled.

[0019] A fundamental idea of ​​the present invention is that, in order to optimize a functional parameter for setting a desired operating variable, such as a wheel speed of a vehicle, with a control unit, a combination of a so-called classical optimizer and a machine learning ML model, which is ideally implemented as a neural network, is used.

[0020] The method according to the invention makes it possible to use the advantages of the classic optimizer and the ML model to find an optimal functional parameter for setting a target operating variable of a technical system, such as a vehicle or a plant, by means of a control unit.

[0021] For this purpose, the ML model is used at least twice in the following way, as explained in simplified form in the inventive principle below:

[0022] When the ML model is used for the first time, it is trained to establish a functional relationship between an input and an output. The input can be an initial function parameter, and the output can be a corresponding and simulated or otherwise practically determined actual behavior of an operating variable, such as a vehicle's wheel speed.

[0023] In the next step, an ideal or target profile of the desired operating variable is fed into the ML model. The ML model provides corresponding and improved functional parameters as output. These functional parameters generated by the ML model are then used as input data for a new simulation of the operating variable in order to find improved functional parameters in a further iteration step.

[0024] At the same time, however, functional parameters of an optimizer are also used as additional input data for the re-simulation of the operating variable. The optimizer is fed with an auxiliary variable that represents the difference between the actual and the target curve of the desired operating variable. The optimizer attempts to minimize the difference between the actual and the target or target curve of the desired operating variable, i.e., to make it as small as possible.

[0025] The output data of the ML model and the optimizer can be entered alternately into the simulator to determine the actual course of the company size.

[0026] One embodiment of the method provides that the step of determining the actual operating variable is carried out using at least one of the following test methods: simulation, measurement, road test, or test bench. In this way, the actual course of the operating variable can be efficiently determined.

[0027] One embodiment of the method provides for the auxiliary parameter to be calculated by applying a mean square error function between the actual operating variable and the target operating variable. This allows the auxiliary parameter to be determined simply and efficiently. One embodiment of the method provides for the optimizer to use at least one of the following algorithms: genetic algorithm, hyperband, Bayesian Optimization and Hyperband (BOHB), Sobol Sequence Guided Search, or Gaussian Process-based Bayesian Optimization. This allows the functional parameter to be determined flexibly and optimized depending on the application scenario.

[0028] One embodiment of the method provides that, in order to determine the optimal functional parameter for setting the operating size of the technical system, the aforementioned steps are repeated after the first initial functional parameter has been determined.

[0029] According to a second aspect, the disclosure relates to a control unit having a computing unit, wherein the control unit is designed to determine at least one optimized functional parameter for controlling an operating variable of a technical system using methods according to the invention according to the first aspect.

[0030] According to a second aspect, the disclosure relates to a detection system for classifying an object to be detected, comprising at least one ultrasonic sensor which is designed to carry out the method according to the invention, wherein the at least one ultrasonic sensor can be used in a vehicle.

[0031] According to a third aspect, the disclosure relates to a computer program containing machine-readable instructions which, when executed on one or more computers and / or computer instances, cause the computer or computer instances to carry out the method according to the invention.

[0032] According to a fourth aspect, the disclosure relates to a machine-readable data carrier and / or downloadable product with the computer program. According to a fifth aspect, the disclosure relates to one or more computers and / or computer instances with the computer program, and / or with the machine-readable data carrier and / or the downloadable product.

[0033] Further measures improving the invention are described in more detail below together with the description of the preferred embodiments of the invention with reference to figures

[0034] Examples of implementation

[0035] It shows:

[0036] Figure 1 Schematic flow diagram of the method 100 for optimizing at least one functional parameter in order to adjust an operating variable of a technical system 96; and

[0037] Figure 2 Schematic representation of an architecture for carrying out the method 100 according to an embodiment of the present invention.

[0038] Figure 1 shows a schematic flow diagram of the method 100 for optimizing at least one functional parameter in order to adjust an operating variable of a technical system 96.

[0039] In step 102, at least one first initial function parameter 4 or a first function parameter set 4 is defined for a first iteration step.

[0040] In step 104, at least one actual operating variable 93 is determined by supplying the first functional parameter 4 to the technical system 96.

[0041] Optionally, in step 104, the actual operating variable 93 is determined using at least one of the following test methods: simulation, measurement, simulated and / or real driving test, test bench. In step 106, a second function parameter 6, which corresponds to the first actual operating variable 93, is generated by feeding the determined first actual operating variable 93 to an ML model 20 as an input variable in order to pre-train the ML model 20. In this way, the ML model 20 learns a (functional) relationship between the input variable, the initial function parameter 4, and the determined actual operating variable or the actual curve of the desired operating variable 93, e.g., a wheel speed of a vehicle, deceleration of a vehicle during braking.

[0042] In step 108, a third function parameter is generated, which corresponds to a predetermined target operating variable 94, by supplying the pre-trained ML model 20 with the predetermined target operating variable 94 on which the ML model 20 is trained.

[0043] In step 110, an auxiliary parameter 32 for an optimizer 30 is determined as an input variable for the optimizer 30. The auxiliary parameter 32 is determined from the first actual operating variable 93 and the specified target operating variable 94.

[0044] Optionally, the auxiliary parameter 32 is formed by applying a mean square error function between the actual operating variable 93 and the target operating variable 94.

[0045] The optimizer 30 can use at least one of the following algorithms: genetic algorithm, hyperband, Bayesian Optimization and Hyperband (BOHB), Sobol Sequence guided search, Gaussian Process based Bayesian Optimization.

[0046] In a step 112, a fourth function parameter 10 is determined by feeding the determined auxiliary variable 32 to the optimizer 30.

[0047] In a step 114, the third functional parameter 8 output by the ML model 20 and the fourth functional parameter 10 output by the optimizer 30 are supplied as a new input variable for the technical system 96 for at least one subsequent iteration step in order to find the optimized functional parameter for the operating variable of the technical system 96 to be controlled. Steps 104 to 114 can be repeated as often as desired, as described above, in order to determine the optimal functional parameter for adjusting the operating variable of the technical system 96.

[0048] In the context of the present invention, it should be mentioned here that a first, second, third or fourth functional parameter can be understood as meaning not only one functional parameter but also a number or a plurality of functional parameters.

[0049] The method described above for finding the optimized functional parameter for the operating variable to be adjusted can be carried out by a control unit 98 with a computing unit. The control unit 98 can, in particular, be a component of a technical system 96, i.e., the technical system 96 can be controlled via the control unit 98. The technical system 96 can be, for example, a system or a vehicle or a device to be automated.

[0050] Figure 2 shows a schematic representation of an architecture for carrying out the method 100 according to an embodiment of the present invention.

[0051] As already explained, the ML model 20 is first trained to establish a functional relationship between an input variable and an output variable. The input variable can be an initial functional parameter 4, and the output variable can be a corresponding and simulated or other practical, such as measured, (real / simulated) driving tests, test bench use, or the actual course of an operating variable 93, such as a vehicle's wheel speed.

[0052] The determined actual trend of the operating variable 93 is then fed into the ML model 20 as input data to obtain a corresponding new function parameter 6. In this way, the ML model learns the relationship between the actual trend of the operating variable 93 and the (initial) function parameter 4.

[0053] In a next step, an ideal or target / target curve of the desired operating variable 94 is fed into the ML model 20. The ML model 20 supplies corresponding and improved (third) function parameters 8 as output or initial data. This function parameter 8 generated by the ML model 20 is then used again as input or first input data for a new simulation of the actual operating variable 93 in order to find improved function parameters in a new iteration step.

[0054] At the same time, however, function parameters 10 of an optimizer 30 are also used as additional or second input data for the renewed simulation of the actual operating variable 93. The optimizer 30 is fed with an auxiliary variable 32, which represents a difference between the actual curve 93 and the target curve 94 of the desired operating variable. The optimizer 30 attempts to minimize the difference between the actual curve 93 and the target or target curve 94 of the desired operating variable, i.e., to make it as small as possible.

[0055] The output data of the ML model 20 and the optimizer 30 can be entered alternately into the simulator to determine the actual course 93 of the operating variable.

[0056] The essential aspects of the method according to the invention are explained below using a concrete example:

[0057] The operating variable to be adjusted is a wheel speed of a vehicle. The function parameters P, l and D of the controller are responsible for this, for example. The aim of the present invention is therefore to determine the parameters P, l and D such that the resulting wheel speed has the smallest possible deviation from the ideal or desired / target curve of the operating variable. For this purpose, initial parameters are estimated. On these basis, the classic optimizer 30, for example using genetic algorithms, hyperband, BOHB, Sobol sequence guided search, Gaussian process based Bayesian optimization, etc., is fed with an auxiliary target variable, such as the mean square error, in order to find the ideal parameters.

[0058] Additionally, after each iteration (which can be conducted as a simulation or road test), the ML model 20, which can be implemented as a neural network, receives the resulting wheel speed over time as input and the underlying functional parameters as output. The neural network is thus trained in such a way that it can infer the underlying functional parameters based on measurements. Ideally, retraining with the newly generated data takes place after each iteration.

[0059] After this training of the measured / simulated curves and the underlying function parameters, the network is fed with the ideal curve (or target curve) as input, and the most probable function parameters for this result are returned based on the training data. This information is provided to the classical optimizer for the next iteration. Accordingly, the classical optimizer and the neural network, which learns general relationships between the curve and the desired function parameters, are executed alternately.

Claims

Claims 1. Method (100) for optimizing at least one functional parameter in order to adjust an operating variable of a technical system (96), comprising the following steps: Setting (102) at least one first initial function parameter (4) for a first iteration step; Determining (104) at least one actual operating variable (93) by supplying the at least first functional parameter (4) to the technical system (96), generating (106) a second function parameter (6) which corresponds to the first actual operating variable (93) by supplying the determined first actual operating variable (93) to an ML model (20) as an input variable in order to pre-train the ML model (20); generating (108) a third function parameter (8) which corresponds to a predetermined target operating variable (94) by supplying the pre-trained ML model (20) with the predetermined target operating variable (94) on which the ML model (20) is to be trained; Determining (110) an auxiliary parameter (32) for an optimizer (30) as an input variable for the optimizer (30), wherein the auxiliary parameter (32) is determined from the first actual operating variable (93) and the predetermined target operating variable (94), Determining (112) a fourth function parameter (10) by feeding the determined auxiliary variable (32) to the optimizer (30), Supplying (114) the third functional parameter (8) output by the ML model (20) and the fourth functional parameter (10) output by the optimizer (30) as a new input variable for the technical system (96) for at least one next iteration step in order to find the optimized functional parameter for the operating variable of the technical system (96) to be controlled.

2. Method (100) according to claim 1, wherein in the step (104) of determining the actual operating variable (93) is carried out by at least one of the following test methods: simulation, measurement, driving test, test bench.

3. Method (100) according to one of the preceding claims, wherein the auxiliary parameter (32) is formed by applying a mean square error function between the actual operating variable (93) and the desired operating variable (94).

4. The method (100) according to any one of the preceding claims, wherein the optimizer (30) uses at least one of the following algorithms: genetic algorithm, hyperband, Bayesian optimization and hyperband (BOHB, Sobol sequence guided search, Gaussian process based Bayesian optimization.

5. Method (100) according to one of the preceding claims, wherein steps (104) to (114) are repeated to determine the optimal functional parameter for adjusting the operating variable of the technical system (96).

6. Control unit (98) with a computing unit, wherein the control unit (98) is designed to determine at least one optimized functional parameter for controlling an operating variable of a technical system (96) using a method according to one of claims 1 to 5.

7. A computer program containing machine-readable instructions which, when executed on one or more computers and / or compute instances, cause the computer or computers or compute instances to carry out the method according to any one of claims 1 to 5.

8. Machine-readable data carrier and / or download product with the computer program according to claim 7.

9. One or more computers and / or compute instances with the computer program according to claim 7, and / or with the machine-readable data carrier and / or download product according to claim 8.

Images