Method and apparatus for performing a tolerance-robust design optimization

Abstract

The invention relates to a computer-implemented method for performing tolerance simulations of design components to find suitable design components, comprising the following steps: - Providing (S1) a variety of design components, each defined by component properties; - Determine (S2) one or more key parameters of the design components using a simulation method; - Providing a probabilistic machine learning model trained to map the component properties of a design component to one or more key variables and to determine an evaluation uncertainty, - For each design component to be evaluated, perform a tolerance robustness evaluation, using the probabilistic machine learning model (S7) if the evaluation of the component properties of the design component in question with the machine learning model results in an evaluation uncertainty below a specified threshold, and using a robustness simulation (S8) if the evaluation of the component properties of the design component in question with the machine learning model results in an evaluation uncertainty above the specified threshold. - Determine (S9, S10) the relevant design component as a suitable design component depending on the result of the tolerance robustness evaluation.

Description

Technical field

[0001] The invention relates to methods for simulating component designs while taking component tolerances into account. The invention further relates to the suitable selection of design components based on an aggregated key parameter and its tolerance sensitivity to deviations in component properties. Technical background

[0002] In tolerance-robust design optimization, the tolerance sensitivity of a large number of design components is investigated with respect to component properties varied within the tolerance range. This involves, for example, performing simulations of the design components and determining an aggregated key metric that quantitatively assesses the performance of the design component. For each design component, a series of tolerance variations is performed, in which one or more component properties are varied within a predefined tolerance range. The influence of each variation on the key metric is evaluated through a separate simulation. The distribution of the key metric as component properties are varied allows for an assessment of the robustness of the design component.

[0003] Several algorithms are known for design optimization, some of which are tailored to Pareto optimization. This involves simultaneously determining and optimizing multiple key parameters.

[0004] To accelerate the evaluation of design components through simulation, the simulation can be replaced by machine learning models that specify a relationship between the design component's design parameters and the aggregated key metric.

[0005] The effort required to evaluate the robustness of a design component is very high due to the large number of simulations needed, since for each design of a design component the key parameters must first be determined and then at least several dozen tolerance simulations must be carried out for the robustness evaluation with regard to the component tolerances in order to determine the sensitivity of the key parameters to deviations in component properties.

[0006] Tolerance simulations are computationally intensive, so design optimization with robustness evaluation involves considerable effort when evaluating a larger number of design components, up to 100,000.

[0007] It is an object of the present invention to accelerate the determination of optimized design components by significantly reducing the number of necessary simulations. Disclosure of the invention

[0008] This problem is solved by the method for performing tolerance simulations of design components to find an optimized design component according to claim 1 and a corresponding device according to the dependent claim.

[0009] Further details are specified in the dependent claims.

[0010] According to a first aspect, a computer-implemented method for performing tolerance simulations of design components to find a suitable design component for a technical system is provided, comprising the following steps: - Providing a wide variety of design components, each defined by component properties; - Determining one or more key parameters of the design components using a simulation method; - Providing a probabilistic machine learning model that is trained to map the component properties of a design component to one or more key parameters and to determine an evaluation uncertainty, - For each design component to be evaluated, perform a tolerance robustness evaluation, using the probabilistic machine learning model for the tolerance robustness evaluation if the evaluation of the component properties of the design component in question with the machine learning model results in an evaluation uncertainty below a specified threshold, and using a robustness simulation for the tolerance robustness evaluation if the evaluation of the component properties of the design component in question with the machine learning model results in an evaluation uncertainty above the specified threshold, and determine the design component in question as a suitable design component depending on the result of the tolerance robustness evaluation.

[0011] A design component corresponds to the design of a specific component to be optimized and is defined by component properties for at least one part of the design component. These component properties can include a dimension and / or arrangement, a material parameter, a mechanical property, or an electrical property.

[0012] The appropriate design component can be used, in particular, for the manufacture or development of a technical system. The design component is considered or incorporated into the design of the technical system, taking into account its corresponding component properties.

[0013] In general, the above procedure can be applied to all technical systems where component tolerances must be considered during design / development. This is particularly relevant when component tolerances strongly influence key parameters (target values / performance parameters), especially if these tolerances are not constant within the design space, making meaningful design possible only by taking them into account.

[0014] The technical system can correspond to an electric machine, where the design component can be a component of the electric machine, such as a rotor, a stator, or a stator / rotor combination. The design component is defined by its component properties, which specify, in particular, the geometry of a stator and / or a rotor, one or more material properties of the stator and / or the rotor, and an electrical property, in particular a winding resistance or a number of turns. The one or more key parameters in this context can include a specification of the magnetic field distribution in the rotor and / or stator or in an air gap between the rotor and stator, electromagnetic radiation, a temperature distribution, a maximum temperature, or a torque.

[0015] When designing components for electric machines, tolerances in the component geometry, as well as variations in the magnetic strength of individual magnets, play a significant role. This affects the expected torque of the electric machine, and especially the torque ripple during rotation. This torque ripple is a key parameter for many electric machines and should generally be very low. Studies have shown that the torque ripple for the nominal machine differs from the average value across machines when component tolerances are taken into account by factors between approximately 1 and 10. This factor is not constant and varies considerably within the design space.

[0016] If the time required to simulate the electric machine is so high that a sufficient number of machines with component tolerances cannot be simulated, then a design based on these tolerances is not possible. The method described above can significantly reduce the number of machine designs to be simulated. These designs are still fully evaluated, but unnecessary simulations are largely avoided. This then enables a system design. Tolerance-robust design optimization is very computationally intensive, as various tolerance variations must be checked for each design component through a separate simulation.For each design component, a predefined number of variations of component properties are determined. These are selected randomly or using suitable sampling methods, such as Latin Hypercube Sampling (which can provide faster and more accurate results), from the respective tolerance ranges. These tolerance design components are then analyzed using robustness simulation. For each design component, a statistical deviation from one or more standard values ​​of a key parameter is determined and evaluated against a predefined error limit. The standard procedure in EM / EDS2, for example, is to analyze 50 variations of tolerance design components to determine the statistical deviation for robustness evaluation with sufficient accuracy.

[0017] Due to the high computational time required for a large number of simulations for robustness evaluation, such design optimizations can currently only be carried out for design components that can be simulated with low computational time.

[0018] One approach is to use a machine learning model to perform robustness optimization in an improved manner. Instead of a costly simulation, a pre-trained machine learning model is evaluated to determine one or more aggregated key parameters for the combination of varied component properties. This allows the tolerance deviations for the component properties to be correlated with the statistical deviation of one or more key parameters. These key parameters can include performance parameters of the design component, such as weight, stability, electrical properties, and the like.

[0019] Thus, the existing robustness simulation serves as a basis, whereby a simulation is then extended by the machine learning model and a concrete process plan in such a way that design optimization under component tolerances is made possible.

[0020] Key parameters for tolerance analysis can include, for example, torque and torque ripple (i.e., the fluctuations in torque over one motor revolution). Here, the mean and standard deviation over the robustness variations can be considered, or combined values ​​such as mean + 3 x standard deviation or mean + 6 x standard deviation.

[0021] Since the evaluation of a corresponding machine learning model can be performed much faster than through simulation, allowing more tolerance design components to be checked for robustness within an acceptable timeframe, a robustness evaluation based on this method is fundamentally suitable for use in design optimization. The machine learning model can be implemented as a neural network or similar and trained on training data obtained from more complex, higher-precision simulations, thus enabling a better assessment of the tolerance robustness of the design component. However, the evaluation of tolerance design components is also time-consuming and, especially with a large number of components to be checked, involves a significant strain on computing resources.

[0022] The core of the above method is to reduce the number of design components subjected to robustness testing in order to minimize the duration of tolerance simulations. To achieve this, a machine learning model is identified and trained based on the results of a tolerance simulation of the design components. The machine learning model is preferably a probabilistic model and outputs not only the value but also a quantification of the uncertainty. Examples of such models include a Gaussian process model or a Bayesian neural network.

[0023] Tolerance robustness evaluation can be carried out by varying the component properties of the design component in question within its specified tolerance ranges and determining the distribution of the resulting one or more key parameters, where the distribution of the one or more key parameters corresponds to the result of the tolerance robustness evaluation.

[0024] The tolerance robustness evaluation results in an acceptable or unacceptable variation of at least one key parameter, where the specification defines, for example, mean values, standard deviations, or combinations of these metrics for a specific key parameter with a minimum or maximum value. The result is unacceptable if the limit values ​​specified in the specification are exceeded or fallen below.

[0025] It may be stipulated that the tolerance robustness evaluation is always performed using robustness simulation as long as the machine learning model has not been trained with a sufficient number of training data sets.

[0026] The machine learning model is trained in an initial optimization phase for a number of design components, for each of which a tolerance simulation is also performed. As described above, the tolerance simulation involves simulating many tolerance design components with values ​​varied within the tolerance ranges of the component properties. In the further course of the design optimization, the machine learning model is successively trained further when one or more key parameters are available for a design component through simulation.

[0027] The machine learning model is trained using training datasets that map the component properties (i.e., the values ​​of the quantified component properties, such as dimensions, electrical parameters, material parameters, and the like) of the design component and the already specifically simulated tolerance design components to one or more aggregated key variables. The machine learning model is trained using these training datasets under supervision. Due to the preferably probabilistic design of the machine learning model, an evaluation uncertainty can also be specified during its inference, indicating how accurately the machine learning model can predict the key variable.

[0028] In a subsequent optimization phase, a robustness evaluation is then performed using the machine learning model. A new design component to be evaluated can then be analyzed using the machine learning model to determine the relevant one or more key parameters.

[0029] Instead of simulation, the evaluation of the machine learning model will be used to determine the variation of one or more key parameters depending on the variation of the tolerance-related component properties / design parameters. Beforehand, it will be checked whether the evaluation uncertainty for the relevant component properties of the design component under consideration is sufficiently low to allow for reliance on the machine learning model evaluation instead of simulation. In this way, as the machine learning model's training progresses, robustness evaluation via simulation can be dispensed with for an increasing number of design components, thereby significantly accelerating the process of finding an optimized design component.

[0030] It may be provided that one or more key dimensions for each design component are checked according to at least one key dimension criterion, whereby the design component in question is rejected and no tolerance robustness evaluation is carried out if it does not meet the at least one key dimension criterion.

[0031] This is made possible by the result of a simulation that evaluates the non-tolerance component properties (standard values ​​of the design parameters) of a specific design component with respect to the resulting key parameter, for example, by checking the key parameter against a required design property (specification). For example, the key parameter could correspond to a maximum operating temperature of the design component, which must not exceed a certain predefined maximum temperature.

[0032] Alternatively or additionally, it may be provided that the multiple key parameters for each design component are checked according to a Pareto optimum of at least two of the key parameters, whereby the design component in question is rejected and no tolerance robustness evaluation is carried out if the at least two associated key parameters are outside a tolerance range around the Pareto optimum.

[0033] Thus, the process and the suitable design components can be embedded in an optimization algorithm, and the choice of the design of the suitable design component of the technical system can be determined by the Pareto optimum of the design optimization.

[0034] This allows verification of whether the design component to be evaluated is close to a Pareto front, where two or more key variables are in a Pareto relationship. A tolerance distance to the Pareto front can be defined, within which a design component is permitted for simulation of a robustness evaluation; otherwise, it is rejected. This ensures that the design component to be simulated exhibits sufficient performance before a robustness evaluation is performed. This saves additional computational effort that would otherwise be incurred either by the simulation itself or by the evaluation using the machine learning model.

[0035] The results of the robustness simulation can be used for further training / retraining of the machine learning model. Brief description of the drawings

[0036] The embodiments are explained in more detail below with reference to the accompanying drawings. These show: Fig. 1. A schematic representation of a computer system for performing robustness-optimized design optimization of a design component; and Fig. 2. A flowchart illustrating a procedure for performing a design optimization of a design component with a robustness evaluation. Description of embodiments

[0037] Fig. Figure 1 shows a schematic representation of a computer system 1 for performing a simulation to determine a key parameter for a design component. The computer system 1 comprises a processor 2 and a data and program memory 3 and enables the display of the simulation results via an output device 4.

[0038] Fig.Figure 2 schematically illustrates, using a flowchart, the procedure for carrying out a tolerance-robust design optimization of a design component for a technical system to be developed or manufactured.

[0039] In step S1, a large number of design components are initially provided. These components are defined based on their properties, which specify the dimensions and / or arrangement of individual components, material parameters, mechanical properties (such as surface roughness, modulus of elasticity, etc.), and electrical properties (such as electrical resistance, etc.). In particular, the component properties and design parameters are determined in such a way that they define the design components and their structure.

[0040] In step S2, a nominal design evaluation is performed. This is typically done using a simulation to determine one or more key parameters, which can represent aggregated quantities. The key parameter corresponds to a performance parameter that indicates the performance of the simulated design component with respect to its intended application. The one or more key parameters can be determined, for example, from a finite element simulation that simulates the field distribution of a physical quantity within an area or volume of the design component, so that the key parameter results from an aggregation of the field distribution of the physical quantity. For example, the key parameter could correspond to a maximum temperature within the simulated area.

[0041] In step S3, it is checked for each selected design component whether a machine learning model can be used for evaluation. The evaluation uncertainty of the machine learning model can be used here. If this uncertainty is significantly above a predefined threshold, the tolerance variations cannot be replaced. The specific requirements typically depend on the respective products and design specifications. If the machine learning model cannot be used for evaluation (alternative: No), the process continues with step S8. Otherwise, the process continues with step S4.

[0042] In step S4, the one or more key parameters are first checked against a key parameter criterion to determine whether they meet the requirements for the design component to be selected. For example, if a key parameter corresponds to a maximum temperature within the design component, it can be checked whether this temperature is exceeded. Since exceeding a predefined maximum temperature limit is not permitted, the design component may be rejected, and no subsequent tolerance robustness evaluation is required. If the one or more key parameters meet the key parameter criteria (alternative: Yes), the procedure continues with step S5; otherwise (alternative: No), the procedure continues without a tolerance robustness evaluation with step S3 to analyze the next design component, and the design component in question is rejected.

[0043] In step S5, it is checked whether the one or more key parameters within a range lie on a Pareto front. This front is defined by several key parameters or by one or more key parameters and a cost parameter, which represents the cost of the design component defined by the component properties. The cost can correspond to the estimated production or manufacturing costs of the design component. If it is determined that the design component lies sufficiently close to the Pareto front (alternative: Yes), the procedure continues with step S6. Otherwise, the procedure continues with step S3 without tolerance robustness evaluation to analyze the next design component, and the design component in question is discarded.

[0044] The machine learning model can be designed as a probabilistic machine learning model, such as a Gaussian process model or a Bayesian neural network. For a specific design component, the component properties are used in step S6 to check whether the evaluation uncertainty of the respective component properties of the design component under review is sufficiently low. This can be done using a threshold comparison.

[0045] If the evaluation uncertainty is sufficiently low (alternative: Yes), the machine learning model can be used in step S7 instead of the tolerance robustness simulation to determine the robustness to tolerance variations for the selected design component. For this purpose, the machine learning model is configured to assign the component properties to one or more key parameters. By varying the component properties within their predefined tolerance ranges (specified by the specification), it can be determined within which range of values ​​the one or more key parameters vary.

[0046] Otherwise (alternative: No), a corresponding robustness evaluation is performed for the design component in step S8 using simulations of tolerance variations. Similarly, the component properties are assigned to one or more key parameters. By varying the component properties within their predefined tolerance ranges (specified by the specification), it can be determined within which range of values ​​the one or more key parameters vary. The result of the robustness evaluation simulations can be used to further refine / retrain the machine learning model, thus reducing the number of tolerance simulations required throughout the entire process, as the robustness model becomes sufficiently accurate.

[0047] The result of the machine learning model evaluation and the simulation using tolerance variations is a distribution of one or more key variables. In step S9, this distribution can be checked against distribution limits, maximum values, and minimum values ​​to determine whether the variation of the one or more key variables lies within the permissible specification. Typically, the mean and standard deviation of the distributions are determined or predicted here; from these, combined values ​​such as mean + 3 x standard deviation or mean + 6 x standard deviation can also be calculated. The specification refers to these specific variables and defines the requirements (i.e., the simulation must be able to reproduce the variables from the specification).If one or more key parameters are within the permissible specification (alternative: Yes), the procedure continues with step S10, the design component is determined as a possible design component, and the procedure continues with step S3. Otherwise (alternative: No), the design component is discarded, and the procedure continues with step S3 to analyze the next design component.

Claims

[1] Computer-implemented method for performing tolerance simulations of design components to find suitable design components for a technical system, comprising the following steps: - Providing (S1) a variety of design components, each defined by component properties; - Determine (S2) one or more key parameters of the design components using a simulation method; - Providing a probabilistic machine learning model that is trained to map the component properties of a design component to one or more key parameters and to determine an evaluation uncertainty, - For each design component to be evaluated, perform a tolerance robustness evaluation, using the probabilistic machine learning model (S7) if the evaluation of the component properties of the design component in question with the machine learning model results in an evaluation uncertainty below a specified threshold, and using a robustness simulation (S8) if the evaluation of the component properties of the design component in question with the machine learning model results in an evaluation uncertainty above the specified threshold. - Determine (S9, S10) the relevant design component as a suitable design component depending on the result of the tolerance robustness evaluation. [2] Method according to claim 1, wherein the one or more key sizes for each design component are checked according to at least one key size criterion, wherein the design component in question is rejected and no tolerance robustness evaluation is carried out if it does not meet the at least one key size criterion. [3] Method according to claim 1 or 2, wherein the multiple key parameters for each design component are checked according to a Pareto optimum of at least two of the key parameters, wherein the design component in question is rejected and no tolerance robustness evaluation is carried out if the at least two associated key parameters are outside a tolerance range around the Pareto optimum. [4] Method according to one of claims 1 to 3, wherein the tolerance robustness evaluation is carried out by varying the component properties within their specified tolerance ranges for the design component in question and determining the distribution of the resulting one or more key parameters, wherein the distribution of the one or more key parameters corresponds to the result of the tolerance robustness evaluation. [5] Method according to any one of claims 1 to 4, wherein results of the robustness simulation are used for further training / retraining of the machine learning model. [6] Method according to any one of claims 1 to 5, wherein the component properties for at least one component of the design component define a dimension and / or an arrangement, a material parameter, a mechanical property or an electrical property. [7] Method according to any one of claims 1 to 6, wherein the tolerance robustness evaluation is always carried out using robustness simulation as long as the machine learning model has not been trained with a sufficient number of training data sets. [8] Method according to any one of claims 1 to 7, wherein the suitable design component is used for the manufacture or development of a technical system. [9] Method according to any one of claims 1 to 8, wherein the technical system corresponds to an electric machine, wherein the design component is a component of the electric machine, in particular a rotor, a stator or a stator / rotor combination, wherein the design component is defined by the component properties, which in particular specify a geometry of a stator and / or a rotor, one or more material properties of the stator and / or the rotor, and an electrical property, in particular a winding resistance or a number of turns, wherein the one or more key parameters include a specification of a distribution of the magnetic field in the rotor and / or stator or in an air gap between rotor and stator, of electromagnetic radiation, of a temperature distribution, of a maximum temperature or of a torque. [10] Method according to any one of claims 1 to 8, wherein the method and the suitable design components are embedded in an optimization algorithm and the choice of the suitable design of the technical system is determined by a Pareto optimum of the design optimization. [11] Device for carrying out one of the methods according to any one of claims 1 to 9. [12] Computer program product comprising instructions which, when the program is executed by a computer, cause it to perform the steps of the method according to any one of claims 1 to 9. [13] Machine-readable storage medium comprising instructions which, when executed by a computer, cause it to perform the steps of the method according to any one of claims 1 to 9.

Images