Systematic selection of representative samples of batch products
By distinguishing between system and operating parameters and using dissimilarity metrics to select representative parametric simulation models, the problem of characterizing the uncertainty of ODD models in steer-by-wire and highly automated driving steering systems is solved, enabling efficient product sample selection and regulator design.
Patent Information
- Application Number
- CN202610664579.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2025-05-14
- Filing Date
- 2026-05-14
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies struggle to systematically select representative product samples to characterize model uncertainties in the Operational Design Domain (ODD) of mass-produced products, particularly in online steering and highly automated steering systems. This leads to discrepancies between simulation models and real-world behavior, and regulator designs lack effective model representativeness.
Dissimilarity measures such as gap metric, ν gap metric, and L2 metric are used to distinguish between system parameters and operating parameters. A representative parameterized simulation model is selected through computer implementation methods to ensure that the sample represents the entire ODD, reduce mathematical complexity and cost, and reduce over-engineering.
It enables the systematic selection of representative product samples, quantification of residual uncertainty, characterization of model uncertainty in ODD, support for regulator design and simulation of release process, and improves the design efficiency and accuracy of product regulation devices.
Smart Images

Figure CN123113702A_ABST
Abstract
Description
Technical Field
[0001] Mass-produced products (especially high-volume products), such as steering systems, suffer from variations in product parameters, such as friction, elasticity, and / or inertia, due to manufacturing tolerances and inaccuracies. Furthermore, mass-produced products suffer additional parameter variations due to aging, for example, caused by wear and / or environmental influences. The range and combinations of values for these parameter variations in mass-produced products that actually occur constitute their so-called "operational design domain" (ODD), which in German is, for example, "Betriebsbereich für die Auslegung" (operational design domain). Background Technology
[0002] Typically, only basic or relevant characteristics of the product are incorporated into a simulation model (hereinafter referred to as the "model") of the product, thus resulting in a simplification-induced deviation between real behavior and modeled behavior. Due to the dispersion of parameters in the ODD, additional deviations arise for each individual batch of product between its real behavior and the associated simulation model, which often focuses on representing the product with nominal parameters. This simulation model, including model biases or uncertainties caused by simplification and dispersion throughout the ODD, often forms the basis for product conditioning development and / or simulation-based product release. Since a complete characterization of model uncertainties for batch products is currently often too costly, the simulation model is compared with a small number of real product samples (hereinafter referred to as "musters").
[0003] Typically, product samples are selected as follows: - Samples containing, for example, the minimum, maximum, and / or nominal product parameters that are most important according to expert opinion.
[0004] - Samples with product parameters defined by experts.
[0005] Therefore, it is not guaranteed that the selected product samples fully represent the entire ODD, and thus it is not possible to characterize the model uncertainties in the ODD caused by simplification and dispersion.
[0006] One problem that this disclosure is based on can be viewed, for example, as providing a method for selecting sufficiently representative product samples for a batch of products. Another problem can be viewed as enabling the characterization of model uncertainties in the ODD due to simplification and dispersion.
[0007] Compared to conventional steering systems, steer-by-wire (SbW) systems and / or steering systems used in highly automated driving (HAD) are subject to more stringent specifications regarding product release. To mitigate the significantly increased costs of real-world testing and trials resulting from these stricter release requirements for SbW and HAD steering systems manufactured in (high) production volumes compared to conventional steering systems, a focus on simulation-based release processes has emerged in the industry. For such simulation-based releases, empirically validated and / or verified simulation models of the steering system with known model uncertainties are necessary.
[0008] In principle, numerous different methods are known for validating and / or verifying various aspects of a simulation model or the entire model. However, currently, no method exists that allows for a systematic or complete characterization of the model uncertainty in the ODD (Organizational Determination) caused by simplification and dispersion in a (large) batch of products. In particular, no method exists for systematically or optimally selecting, for example, a small sample of products that represents the entire ODD with quantifiable residual uncertainty, thereby achieving a systematic characterization of the model uncertainty in the ODD.
[0009] Furthermore, steering systems suffer from non-negligible parameter dispersion, and most of it operates within a closed-loop control circuit. Therefore, robust regulators for steering systems are often designed using methods established in control techniques. However, these design methods presuppose models with known uncertainties for the product. Currently, there is no method for establishing one or more models with known uncertainties for regulator design, models that also have a known range of representativeness within the ODD and collectively represent the entire ODD.
[0010] Therefore, another problem that this disclosure is based on can also be seen as: designing one or more models with known uncertainties for the regulator, which in turn have a known range of representativeness in the ODD and collectively represent the entire ODD.
[0011] In systems theory, different metrics are known that quantify the dissimilarity of two systems (where systems and products can be equated in the following text) and thus allow for comparison. The gap metric, the ν-gap metric, and the L2 metric are described below.
[0012] The gap metric uses a scalar in the real interval [0,1] to quantify the dissimilarity in stability and performance characteristics of two systems G1 and G2 (more precisely, their simulated models G1 and G2) in their unregulated (open-loop) input / output behavior during regulated operation (closed-loop). A metric close to 0 indicates that the two systems are very similar, and that a regulator stabilizing G1 also stabilizes system G2 with similar regulated performance. A metric result of 0 indicates that the systems under consideration, G1 and G2, behave exactly the same. On the other hand, a metric result close to or at 1 indicates that systems G1 and G2 are very dissimilar. Furthermore, using the gap metric, statements about the robust stability of closed-loop regulated loops with model uncertainties are possible. Explicit regulator design is required to evaluate the gap metric. Details regarding the definition and characteristics of gap measurement are described in Chapter 17 of the book "Essentials of Robust Control", Kemin Zhou and John C. Doyle, 1st edition, Pearson, 1997, ISBN: 9780135258332.
[0013] The system theory statement and implications of the ν gap metric are very similar to those of the gap metric; however, the two metrics are defined fundamentally differently. Regulator design is not required to evaluate the ν gap metric, but a winding study of the systems G1 and G2 to be compared is necessary. Details regarding the definition and characteristics of the ν gap metric are described in Chapter 17 of the book "Essentials of Robust Control," Kemin Zhou and John C. Doyle, 1st edition, Pearson, 1997, ISBN: 9780135258332, or in the publication "Frequency domain uncertainty and the graph topology," Glenn Vinnicombe, IEEE Transactions on Automatic Control, vol. 38, no. 9, pp. 1371-1383, Sept. 1993, DOI: 10.1109 / 9.237648.
[0014] The definition of the L2 metric corresponds to the ν gap metric in the case of no winding number study. For this reason, the principled statements and implications of the two metrics are similar, but the L2 metric has less theoretical persuasiveness. Details regarding the definition and properties of the L2 metric are described in Chapter 17 of the book "Essentials of Robust Control", Kemin Zhou and John C. Doyle, Istition, Pearson, 1997, ISBN: 9780135258332.
[0015] All three dissimilarity measures introduced are known to have the following properties: - The metric quantifies the dissimilarity of two systems in terms of stability and performance characteristics during regulated operation based on unregulated input / output behavior.
[0016] -Based on the measurement results, it can be determined whether two systems G1 and G2 are similar enough that G1 can be considered representative of G2.
[0017] - The measurement result can also be interpreted as the (system theory) distance between the systems being compared.
[0018] - For all systems to be compared, the following applies: L2 result ≤ ν gap result ≤ gap result.
[0019] For each simulation model of (batch) products, the model parameters can be divided into at least the following two groups: - System parameters, which may represent, for example, the internal characteristics of the product (e.g., gearbox efficiency, bearing friction within the product).
[0020] - Operating parameters, which may represent external effects on the product (e.g., supply voltage, resilient connection to the environment).
[0021] In the systematic selection and / or tolerance determination of product prototypes for uncertainty characterization, there is currently no distinction between system parameters on one hand and operating parameters on the other. Therefore, it is potential to realize many representative prototypes, even though the operating parameters can often be easily and / or cost-effectively achieved in (bench) experiments, and particularly can be externally set, with these representative prototypes associated with specific operating parameters and corresponding external effects.
[0022] Therefore, another problem that this disclosure is also based on can be seen as solving the problems mentioned above more efficiently and / or better. Summary of the Invention
[0023] A first general aspect of this disclosure relates to a computer-implemented method for determining representative parametric simulation models of parametric simulation models for products, particularly steer-by-wire systems and / or steering systems for highly automated driving. The parametric simulation model includes one or more system parameters and one or more operational parameters. The method includes calculating a dissimilarity metric on one pair of multiple pairs of parametric simulation models, wherein a distance is obtained for each pair, and thus multiple distances are obtained. This dissimilarity metric may, for example, be based on a gap metric, a ν-gap metric, and / or an L2 metric.
[0024] The method further includes selecting a predetermined number of parameterized simulation models based on the plurality of distances, wherein a plurality of representative parameterized simulation models are obtained, wherein changes in one or more operating parameters of one or both of the corresponding pairs of parameterized simulation models keep the associated distances constant.
[0025] A second general aspect of this disclosure relates to a computer system designed to perform a computer implementation of the method for determining a representative parametric simulation model for a product, as described in the first general aspect (or its embodiments).
[0026] A third general aspect of this disclosure relates to a computer program or signal designed to perform a computer implementation of the method for determining a representative parameterized simulation model for a parameterizable simulation model of a product, as described in the first general aspect (or its embodiments).
[0027] The fourth general aspect of this disclosure relates to a computer-readable medium that stores and / or contains a computer program according to the third general aspect (or its embodiments).
[0028] The method presented herein, according to the first general aspect (or its implementation), enables the systematic selection of a small number of product samples based on model-based criteria, such that the product samples represent the entire ODD of a batch of products with quantifiable residual uncertainty.
[0029] In particular, the method presented herein according to the first general aspect (or its embodiments) can achieve the following advantages over the prior art: - A systematic operating procedure for selecting product samples that represent the entire ODD; - Quantifiable residual uncertainty of the product sample relative to the entire ODD; -The systematic characterizability of model uncertainties caused by simplification and dispersion throughout the ODD; - The calculable representativeness range of the corresponding product samples in the entire ODD; - The systematic and derivable requirements for the manufacturing precision of product samples; - The requirement for a systematically derivable design for robust product control devices; and / or - Optional search for product parameters or combinations that significantly alter product behavior.
[0030] The methods described herein, as per the first general aspect (or its implementation), can be used during the design phase and / or during system development (i.e., after the design phase) when developing a real product—for example, in the case of a steering system by way of wire (SbW)—in accordance with the first general aspect (or its implementation).
[0031] In the design phase, for example, this method can be used to systematically select (e.g., a small number) product samples that represent all possible ranges and combinations of parameter values for the entire batch of products (i.e., for the entire ODD). Furthermore, the intermediate results of this method can be used to identify individual product parameters or combinations that have a particularly strong impact on product behavior.
[0032] Based on representative product samples, further system development can systematically characterize the uncertainty between the actual product behavior and its modeled behavior throughout the ODD, the uncertainty arising from simplification and parameter dispersion during modeling. Subsequently, simulation models, including the characterized model uncertainties, can be used to develop product conditioning devices and / or for simulation-based product release.
[0033] Additionally, for example in system development, the methods presented herein can be used to systematically establish any number of regulator conception models (i.e., one or more representative models for a regulator conception), including associated model uncertainties, each having a known range of representativeness in the ODD and collectively representing the entire ODD. For example, these regulator conception models can be used to systematically develop gain-scheduling regulation devices for products that take into account known model uncertainties and ranges of representativeness.
[0034] The methods proposed herein, as described in the first general aspect (or its implementation), can be used, particularly in the context of processes for simulation-based release of SbW and HAD steering systems. In this case, it is necessary to select product samples representative of the entire ODD for verification and / or validation of the simulation model used, for example, the steering system. Furthermore, the methods proposed herein can also be used, in principle, for model verification and / or validation and / or in simulation-based release processes in the case of other (high-volume) products.
[0035] The method presented herein according to the first general aspect (or its implementation) can be implemented entirely or partially in a numerical manner. This is advantageous because it does not depend on the existence of a parameterizable simulation model in analytical form.
[0036] Currently, prototype selection and / or tolerance determination for characterizing model uncertainty are invariably based on the results of gaps, ν-gap, L2, or other known dissimilarity measures used for surrogate model pairs, representing the input-output behavior of the considered (product) simulation model. Thus, the measurement result (i.e., distance) indirectly depends on system parameters and operating parameters, without distinguishing the influence of a single parameter or two groups of parameters. For this reason, in prototype selection and / or tolerance determination directly based on measurement results, no distinction is made between system parameters on one hand and operating parameters on the other. If any distinction is made, a representative prototype is currently realized, associated with specific operating parameters and corresponding externalities. However, operating parameters can often be easily and / or cost-effectively implemented (e.g., set) in (bench) experiments, while implementing system parameters is significantly more complex and / or more expensive when constructing a representative prototype. Furthermore, treating system parameters and operating parameters equally increases the mathematical dimension and complexity of prototype selection and / or tolerance determination. This may also negatively impact prototype selection and potentially increase the number of prototypes required, thereby incurring higher costs across the range of uncertainty characterization. For example, it might happen that multiple representative prototypes are selected, initially distinguished by their operating parameters. To address this situation and achieve prototyping and / or cost reduction, currently (if any) costly (manual) post-processing of prototype selection would be required. Furthermore, the increased mathematical dimensionality and / or complexity resulting from treating two parameter sets equally also leads to stricter tolerances in prototype construction, resulting in overengineering.
[0037] The method presented herein according to the first general aspect (or its implementation) is therefore particularly advantageous in that it specifies a distinction between system parameters on one hand and operating parameters on the other. This is achieved by using a dissimilarity metric previously unknown and disclosed herein, which in particular enables the systematic selection of representative prototypes and / or tolerance determination for prototype construction with respect to system parameters, while considering operating parameters simultaneously but treating them differently. Therefore, due to such a dissimilarity metric, a separation occurs between system parameters on one hand and operating parameters on the other, rather than equal treatment. In particular, at least the following advantages are derived: - Regarding system parameters, systematically select representative prototypes to reduce mathematical dimensionality / complexity, reduce the number of prototypes, and / or reduce costs.
[0038] - Regarding system parameters, tolerance determination is performed systematically for prototype construction with reduced mathematical dimensionality / complexity and / or relaxed tolerances. The latter can reduce so-called over-engineering.
[0039] Furthermore, it should be emphasized that the systematic selection of representative prototypes in the proposed method is therefore better and / or more efficient, not only because the method does not select representative prototypes that differ (essentially) only from external factors, i.e., from the operating parameters that can be set in the (test bench) experiment—which would also be possible through the aforementioned post-processing—but because the representativeness of the selected prototypes, particularly the predetermined number of prototypes, is based on such a measure of dissimilarity that distinguishes between system parameters and operating parameters. In other words: the mentioned post-processing may in any case result in a reduced number of representative prototypes, but not the representativeness of the predetermined and desired number of prototypes. Attached Figure Description
[0040] Figure 1 The illustrative map solution is an exemplary implementation of a computer-based method for determining a representative parametric simulation model of a parametric simulation model for a product, particularly a steer-by-wire system and / or a steering system for highly automated driving.
[0041] Figure 2 An exemplary implementation of a computer-based method based on an adjacency matrix is illustrated.
[0042] Figure 3a The illustration illustrates an exemplary ODD, which has multiple parametric simulation models, from which three are selected as representative parametric simulation models.
[0043] Figure 3b The illustration is from Figure 3a An exemplary ODD has three selected representative parameterized simulation models. Detailed Implementation
[0044] The method 100 proposed in this disclosure can be designed to select a sufficiently representative sample of products for a batch production. Alternatively or additionally, the method 100 can also be designed to characterize model uncertainties in the ODD caused by simplification and dispersion. Alternatively or additionally, the method 100 can also be designed to establish one or more models with known uncertainties for regulator design, said models having a known range of representativeness in the ODD and collectively representing the entire ODD. In all these cases, the method 100 distinguishes between system parameters on one hand and operating parameters on the other.
[0045] The systematic and model-based selection of representative product samples is described below.
[0046] Method 100 may first be designed to determine a representative parametric simulation model of a parametric simulation model. Alternatively or additionally, method 100 may also be designed to determine a representative sample of parameters in the ODD. Alternatively or additionally, method 100 may also be designed to select a representative sample of products.
[0047] The selection of representative product samples can be based on the representative parameterized simulation model determined by the parameterized simulation model and / or on the representative parameter samples in the ODD.
[0048] This method is 100% or partially numerical, and therefore can be applied even if the parameterizable simulation model does not exist in analytical form.
[0049] First, a computer-implemented method 100 is disclosed for determining a representative parameterized simulation model of a parameterizable simulation model for a product. The product can be, in particular, a batch product, i.e., produced in batches. This method 100 is particularly meaningful if a large number of similar products should be manufactured, but the products (e.g., due to production and / or material) can be different—even though this method 100 can also be applied to non-batch or small-batch products. A large number of similar products can, for example, include >1e5 products per year, >5e5 products per year, or >1e6 products per year.
[0050] The product could be, for example, a steer-by-wire system. Alternatively or additionally, the product could be a steering system for autonomous driving, particularly highly automated driving.
[0051] A parameterizable simulation model includes one or more system parameters and one or more operational parameters. Therefore, a parameterizable simulation model can be parameterized not only by one or more system parameters but also by one or more operational parameters. A parameterizable simulation model can, but does not have to, be analytical.
[0052] For example in Figure 1 The method 100, which provides a schematic map solution, includes calculating a dissimilarity metric 130 on each of a plurality of pairs (typically different) parameterized simulation models, wherein a distance is obtained for each pair, and thus a plurality of distances are obtained, wherein a change in one or more operating parameters of one or both of the two parameterized simulation models of the corresponding pair makes the associated distance invariant.
[0053] Dissimilarity measures can be based on gap measures, ν-gap measures, and / or L2 measures. Specifically, dissimilarity measures can be gap measures, ν-gap measures, or L2 measures. Alternatively, dissimilarity measures can be based on a combination of gap measures, ν-gap measures, and / or L2 measures. A dissimilarity measure outputs a quantitative measure of the dissimilarity of a pair of parametric simulation models, i.e., a quantitative measure of how similar or dissimilar the two parametric simulation models in that pair are. In this regard, a dissimilarity measure can also be called a similarity measure or a comparison measure. The quantitative measure output by a dissimilarity measure can be called a distance (Distanz). If the two parametric simulation models in a pair are similar, the distance can be small. In particular, if the two parametric simulation models in a pair are identical (i.e., maximally similar), the distance can be zero. Conversely, if the two parametric simulation models in a pair are dissimilar, the distance can be large.
[0054] For example in Figure 1 As illustrated in the schematic diagram, the method 100 further includes: selecting a predetermined number of parameterized simulation models based on multiple distances, thereby obtaining multiple representative parameterized simulation models. Figure 3a The -b section shows an example result of such a selection process.
[0055] System parameters can represent (or represent) the internal characteristics of a product (also known as a system) (e.g., gearbox efficiency, bearing friction, etc.). One or more system parameters can be summarized in a vector, for example... ,in And must be an integer. Operating parameters can represent (or) external effects on the product (e.g., supply voltage, resilient connectivity to the product / system environment). One or more operating parameters can be summarized in a vector, for example... ,in And they are integers. Unlike system parameters, runtime parameters can be set, for example, outside the system. For this reason, it is meaningful to distinguish between system parameters and runtime parameters.
[0056] Dissimilarity measurement can be a (numerical) measure of the dissimilarity (or similarity) between parametric simulation models. Here, if a first parametric simulation model is replaced by another first parametric simulation model that differs in one or more operating parameters but not in one or more system parameters, the dissimilarity measurement between the first and second parametric simulation models remains unchanged. Similarly, if a second parametric simulation model is replaced by another second parametric simulation model that differs in one or more operating parameters but not in one or more system parameters, the dissimilarity measurement between the first and second parametric simulation models remains unchanged.
[0057] Here, the dissimilarity measure (and therefore) need not be a measure in the sense of the metric space (X,d). Because, for example, the necessary condition is not met here. On the other hand, dissimilarity measures can be defined on equivalence classes, in which parameterized simulation models are distinguished only by one or more running parameters.
[0058] Selecting a predetermined number of 140 parameterized simulation models may, but is not required to, include specifying one or more operating parameters for these parameterized simulation models.
[0059] The dissimilarity measure between two corresponding parametric simulation models can be based on other measures. For example, this other measure can be based on a gap measure, a ν-gap measure, and / or an L2 measure. In particular, this other measure can be a gap measure, a ν-gap measure, or an L2 measure.
[0060] For example, the dissimilarity measure between two corresponding parameterized simulation models can be defined by a computational criterion that includes this additional measure. This additional measure can include values in real numbers, particularly in the interval [0,1]. The additional measure (denoted by δ in subsequent examples) can be selected according to the technical application and particularly influences the system theory of the dissimilarity measure and its implications (denoted by Δ in subsequent examples).
[0061] The dissimilarity measure between two corresponding parametric simulation models can be based on multiple values, where each value is based on other measures between the other two parametric simulation models respectively.
[0062] For example, the first parameterized simulation model With the second parameterized simulation model The dissimilarity measure between (where) It is a system parameter and These are runtime parameters. (and is an integer) can be defined by a calculation criterion, which includes multiple values (i.e., the value).
[0063] Furthermore, the value among multiple values can be defined by a calculation criterion that incorporates two other parameterized simulation models (e.g. Between ) (i.e., other first-parameterized simulation models) Other second-parameterized simulation models Other metrics between) (e.g. Therefore, the criteria for calculating the value can depend on other first-parameterized simulation models. The (multidimensional) system parameters, if necessary. These system parameters can be compared with the first parametric simulation model. The (multidimensional) system parameters are the same. The criteria for calculating this value can also depend on other second-parameterized simulation models. The (multidimensional) system parameters, if necessary. These system parameters can be compared with the second parameterized simulation model. The (multidimensional) system parameters are the same. Furthermore, the criteria for calculating this value can also depend on other first-parameterized simulation models. The (multidimensional) operating parameters, if necessary. Furthermore, the criteria for calculating this value can also depend on other second-parameterized simulation models. (Multidimensional) operating parameters, if necessary.
[0064] Each of the multiple values can be defined according to the same computational criterion. Given the aforementioned correlation of the computational criteria, it can be concluded that all values among the multiple values are assigned to the first parametric simulation model. (Multidimensional) system parameters, and the second parameterized simulation model, if necessary. The system parameters (multidimensional if necessary). Furthermore, values from multiple parameters can be individually assigned to the corresponding other first-parameterized simulation models. (Multidimensional) operating parameters, and corresponding other second-parameterized simulation models, if necessary. The (multidimensional) running parameters. Here, multiple values can be viewed as a set of values mapping from the product space (Produktraum) of the running parameters to real numbers. Therefore, values can vary with the running parameters (especially with tuples of the running parameters).
[0065] The dissimilarity measure between two corresponding parametric simulation models can be based on the maximum value or (other) statistical parameters of the distribution of the values. For example, the dissimilarity measure between two corresponding parametric simulation models can be defined by a calculation criterion that includes the maximum value and / or (other) statistical parameters of the distribution of the values.
[0066] The following discloses an example of a dissimilarity measure that utilizes the distinction between system parameters and operational parameters to achieve systematic prototype selection and / or tolerance determination. However, mathematical notation must first be introduced: For any prototype, all operational parameters (such as supply voltage, resilient environmental connections, etc.) can be summarized in a vector. ( And all system parameters (such as gearbox efficiency, bearing friction, etc.) can be summarized in a vector. ( And it is an integer). Furthermore, linear time-invariant (LTI) systems... —That is, a parametric simulation model—can use parameters σ and ε to represent the dynamic behavior of the prototype. The gap, ν gap, L2, or other known dissimilarity measures used for two LTI systems G1 and G2 can be annotated as (Other metrics). Furthermore, It can be a polytope (i.e., a generalized polygon of arbitrary dimensions). Furthermore, a distribution (function) can be defined for the randomly distributed operating parameter ε. (That is, probability distribution). Distribution (function) This can be considered, for example, in prototype selection and / or tolerance determination, see below. A confidence interval (typically) for any random variable with probability P, such as symmetric or as small as possible (multidimensional), can be noted as... The specific interval selection for the subsequently modified metric may depend on the application, as detailed below.
[0067] In the first example (“deterministic metric”), the first parameterized simulation model With the second parameterized simulation model The dissimilarity measure Δ between them—where again It is a system parameter and These are runtime parameters. And it must be an integer—this can be defined using the following formula: Among multiple values Through other first-parameter simulation models Other second-parameterized simulation models Other metrics δ between are defined, where It is also a running parameter, and For example, it is a polyhedron.
[0068] Therefore, value Based on two other parameterized simulation models , Other metrics δ between them, and their calculation criteria depend on one or more operating parameters of the corresponding two other parameterized simulation models. Therefore, the first parameterized simulation model With the second parameterized simulation model The dissimilarity measure Δ between them is also based on other measures δ.
[0069] Furthermore, here, the first parametric simulation model With the second parameterized simulation model The dissimilarity measure Δ between the two models depends on the system parameters σ1 and σ2, but not on the operating parameters ε1 and ε2. Therefore, the two parameterized simulation models (i.e., the first parameterized simulation model) and / or a second parameterized simulation model A change in one or more operating parameters of one or two of the parameterized simulation models causes the associated distance to be affected. The difference remains unchanged. Other metrics δ can be based on gap metrics, ν gap metrics, and / or L2 metrics, specifically, for example, gap metrics, ν gap metrics, or L2 metrics. In this example, the dissimilarity metric Δ between the two parameterized simulation models is expressed through the value... The maximum value is obtained. Therefore, here, the dissimilarity measure Δ between the two corresponding parametric simulation models is based on the maximum value. Thus, the dissimilarity measure Δ can be understood as considering, for example, a polyhedron. The worst-case measure of dissimilarity or distance δ between two prototypes with system parameters σ1 and σ2, under all operating parameters.
[0070] In the second example (“random metric”), the first parameterized simulation model Second parameterized simulation model The dissimilarity measure Δ between them—where again It is a system parameter and These are runtime parameters. And it must be an integer—this can be defined using the following formula: Among multiple values Based on the distribution (function) Randomly distributed operating parameters , Through other first-parameter simulation models Other second-parameterized simulation models Other metrics δ are used to define the dissimilarity. From this, the distribution of values is derived again, from which a statistical parameter can be derived, through which the dissimilarity metric Δ can be defined. In this example, this statistical parameter can be defined as... Therefore, it is defined as the upper limit of the confidence interval for a predetermined probability P. For example, the median (i.e., the p-quantile with probability P = 0.5, see below) can be determined first, and then a 5% to 95% confidence interval for that median can be determined. The upper limit of the confidence interval is thus derived here from the median up to the 95% confidence interval. Here, the metric Δ can be understood as the upper limit of the confidence interval for the dissimilarity or distance δ between two prototypes with system parameters σ1 and σ2, taking into account the probability distribution of the operating parameters.
[0071] Alternatively, in the second example, the statistical parameter can (and only) be, for example, the maximum value, so the formula for the dissimilarity measure Δ is: Alternatively, in the second example, the statistical parameter could be, for example, the P-quantile for a predetermined probability P (e.g., P is at least 0.9), and the formula for the dissimilarity measure Δ would be: In such an exemplary case, multiple values can be determined individually using the Monte Carlo method. This is often simply referred to as rolling dice. An empirical distribution can then be determined from these values, from which, for example, the P-quantile can be determined (e.g., from the inverse of the empirical distribution).
[0072] Here, the first parametric simulation model With the second parameterized simulation model The dissimilarity measure Δ between them does not actually depend on the running parameters ε1 and ε2. Therefore, the two parameterized simulation models (i.e., the first parameterized simulation model) and / or a second parameterized simulation model A change in one or more operating parameters of one or two of the parameterized simulation models causes the associated distance to be affected. The same applies. Here, other metrics δ can also be based on gap metrics, ν gap metrics, and / or L2 metrics, specifically, for example, gap metrics, ν gap metrics, or L2 metrics. In this example, the dissimilarity metric Δ between two parameterized simulation models is defined by one of the exemplary statistical parameters of the distribution of values or by the maximum value (which can also be regarded as a statistical parameter of the distribution of values).
[0073] Secondly, two parameterized simulation models (i.e., the first parameterized simulation model) and / or a second parameterized simulation model A change in one or more operating parameters of one or two of the parameterized simulation models causes the associated distance to be affected. constant.
[0074] The dissimilarity measure between two corresponding parameterized simulation models can be based on a constraint on the set of values of statistical parameter pairs based on the distribution of the one or more operating parameters, and on the maximum value in the constrained set of those values.
[0075] In the third example (“hybrid metric”), the first parameterized simulation model With the second parameterized simulation model The dissimilarity measure Δ between them—where again It is a system parameter and These are runtime parameters. And it must be an integer—this can be defined using the following formula: Among multiple values Based on confidence intervals with probability P (Multidimensional) operating parameters, if necessary , Through other first-parameter simulation models Other second-parameterized simulation models Other metrics δ are used to define the confidence interval. Operating parameters based on random distribution , Distribution (function) And it can (but does not have to) be multidimensional. For example, confidence intervals. It can be based on a distribution (function). Randomly distributed operating parameters , The 5% to 95% confidence interval of the median. In other words: here we can therefore first define the set of operating parameters of the parameterizable simulation model—that is, the confidence intervals. Running parameters , It can be located within the set. This is typically a polyhedron. The true limitation, namely For such operating parameters For each combination, values can then be defined individually. And the maximum value (or other statistical parameters of the distribution of values) can be defined from all these values. Therefore, the value is also derived. The limitation is based on a statistical parameter of the distribution of the one or more operating parameters, i.e., based on a confidence interval. Therefore, the dissimilarity measure between two corresponding parameterized simulation models is based on a constraint on the set of values of statistical parameter pairs based on the distribution of the one or more operating parameters. Since values are further determined here... The maximum value of Δ is thus used to measure the dissimilarity between two corresponding parameterized simulation models, which is also based on the maximum value over a restricted set. Here, the dissimilarity measure Δ can be understood as considering... The worst-case measure of dissimilarity or distance δ between two prototypes with system parameters σ1 and σ2, under all operating parameters within a (multidimensional) confidence interval.
[0076] Secondly, two parameterized simulation models (i.e., the first parameterized simulation model) and / or a second parameterized simulation model A change in one or more operating parameters of one or two of the parameterized simulation models causes the associated distance to be affected. constant.
[0077] Multiple representative parametric simulation models can be defined such that each parametric simulation model (in the ODD) minimizes the distance to the corresponding representative parametric simulation model (in the same ODD) that is closest to that parametric simulation model (given the dissimilarity metric). For example, these distances can be minimized under a predetermined finite-dimensional norm. It is also conceivable, for example, to modify, in particular to weight, the distances before applying the finite-dimensional norm, based on existing knowledge about the product and / or its ODD at the time of execution of method 100.
[0078] In this case, the minimum distance, particularly the minimum distance under the finite-dimensional norm, can be determined. Minimization can be performed mathematically rigorously or approximatingly. For example, it might be sufficient to find a local minimum rather than a global minimum. Therefore, a predetermined number of representative parametric simulation models (up to 140) can be selected precisely or approximatingly based on multiple distances.
[0079] The predetermined finite-dimensional norm can be, for example, the p-norm of any p in the interval [1, +Inf], where +Inf denotes positive infinity. In particular, the p-norm can be, for example, the summation norm (i.e., p=1). Here, therefore the sum of all distances can be minimized. In another example, the p-norm can be the Euclidean norm (i.e., p=2). Here, therefore the Euclidean distance can be minimized. In yet another example, the p-norm can be the maximal norm (p=Inf (infinity)). In the case of the maximal norm, multiple representative parametric simulation models can be defined such that the maximum value of all distances between each parametric simulation model and its corresponding nearest representative parametric simulation model is minimized.
[0080] For example in Figure 1 As illustrated in the diagram and as an option, method 100 may further include: determining multiple parameter samples in the Operational Design Domain (ODD) of product 110. The parameter samples in the ODD may each include one or more parameters of a parameterizable simulation model. The parameter samples in the ODD may each include at least one or more system parameters of the parameterizable simulation model and optionally one or more operational parameters. The ODD may be defined by a numerical and / or analytical description. For example, multiple parameter samples in the 110 ODD may be determined such that the ODD is sufficiently uniformly covered with respect to one or more system parameters. Such sufficiently uniform coverage may be based, for example, on pseudo-random numbers. Pseudo-random numbers (e.g., via the Mersenne-Twister algorithm) are typically uniformly distributed. Alternatively or additionally, such sufficiently uniform coverage may be based on Latin hypercube sampling. Alternatively or additionally, such sufficiently uniform coverage may be based on a Sobol sequence. In particular, such a sufficiently uniform coverage can be based on a combination of pseudo-random numbers, Latin hypercube sampling, and / or Sobol sequences. By assuming a sufficiently uniform coverage, representative parametric simulation models for parameterizable simulation models can be selected better and more efficiently.140
[0081] exist Figure 2 In an exemplary implementation, step 110 is referred to as “sampling the ODD”.
[0082] Alternatively, the determination of multiple parameter samples 110 may involve only a portion of the ODD (e.g., in cases where a specific problem is raised or for further study of the ODD).
[0083] For example in Figure 1 As illustrated and as an option, method 100 may further include: constructing a parameterized simulation model 120 based on a parameterized simulation model and multiple parameter samples in the ODD. Here, for example, the parameterized simulation model may be evaluated at one of the parameter samples respectively. Alternatively or additionally, a surrogate model for the parameterized simulation model may be created for one of the parameter samples respectively.
[0084] A parameterizable simulation model can, but does not have to, be analytical. If the parameterizable simulation model is not analytical (e.g., in the case of a blackbox simulation model), it may not be possible to simply evaluate the parameterizable simulation model at each of the parameter samples. In such cases, a surrogate model for the parameterizable simulation model can be created, for example, for each of the parameter samples using numerical simulation. Therefore, such a surrogate model can also be considered a parameterized simulation model of the parameterizable simulation model. Thus, creating a surrogate model for a corresponding parameter sample in the parameter samples can represent parameterizing the parameterizable simulation model. In this respect, the (parameterizable) simulation model can be parameterizable even if it does not exist in analytical form.
[0085] exist Figure 2 In an exemplary implementation, step 120 is referred to as “Determine PUM for all ODD samples based on PAM”, where PAM represents a parameterizable simulation model and PUM represents multiple parameterizable simulation models.
[0086] For example in Figure 1 As illustrated in the diagram and as an option, method 100 may further include: determining a representative measure of the representative parameterized simulation model 141 based on a predetermined number and / or minimum distance, particularly the minimum distance under a finite-dimensional norm. A representative measure may, for example, include the quantifiable residual uncertainty of the product sample with respect to the ODD.
[0087] For example, in Figure 1 As illustrated in the diagram and as an option illustration, method 100 may further include: if the representative measure does not (sufficiently) meet a predetermined criterion, then adaptation 150, in particular, increasing the predetermined quantity. For example, when the representative measure is large enough, the predetermined criterion can be met. After adaptation 150, in particular, increasing the predetermined quantity, method 100 can be repeated, that is, for example, until the representative measure meets the predetermined criterion (i.e., until the representative measure is large enough).
[0088] For example in Figure 1 As illustrated in the diagram and as an option, method 100 may further include constructing a 131 (e.g., symmetric) adjacency matrix based on multiple distances. In particular, the components of the adjacency matrix may include distances. Using an adjacency matrix, the implementation of method 100 can be more efficient.
[0089] The adjacency matrix can also be constructed based on the order of the parametric simulation model. This order can be unrelated. For example, the order could be the order in which multiple parameter samples in the ODD are determined.
[0090] exist Figure 2 In an exemplary implementation, step 131 is referred to as “compute a metric-based adjacency matrix for all PUM pairs”, where again PUM represents multiple parameterized simulation models, and thus each PUM pair is a pair of two PUMs. Here, the metric-based adjacency matrix refers to the adjacency matrix.
[0091] A predetermined number of parametric simulation models based on multiple distance selections of 140 (resulting in multiple representative parametric simulation models) may include a full factorial search in the adjacency matrix. Alternatively or additionally, a predetermined number of parametric simulation models based on multiple distance selections of 140 may include (e.g., direct or gradual) clustering of the adjacency matrix. Alternatively or additionally, a predetermined number of parametric simulation models based on multiple distance selections of 140 may include converting the adjacency matrix into isometric auxiliary points (given a dissimilarity metric) and clustering the auxiliary points. A predetermined number of parametric simulation models based on multiple distance selections of 140 may include a combination of a full factorial search in the adjacency matrix, clustering the adjacency matrix, and / or converting the adjacency matrix into isometric auxiliary points (given a dissimilarity metric) and clustering the auxiliary points. The adjacency matrix may be clustered such that the entire adjacency matrix is provided for clustering (also referred to herein as direct clustering). Alternatively or additionally, the adjacency matrix can be clustered so that the adjacency matrix is used progressively and partially (i.e., block-by-block) for clustering. The latter can be computationally and / or storage-efficient. Clustering of auxiliary points and / or auxiliary points (alone or in combination) can have numerical advantages.
[0092] exist Figure 2 In an exemplary implementation, step 140 based on the adjacency matrix is referred to as “searching representative PUM and ODD samples based on the adjacency-matrix”, where again PUM represents multiple parameterized simulation models and the adjacency-matrix represents the adjacency matrix.
[0093] For example in Figure 1 As illustrated in the diagram and as an option illustration, method 100 may further include: determining 160 a representative parameter sample of one or more system parameters assigned to a representative parameterized simulation model. Method 100 may also include: determining or specifying one or more operating parameters assigned to the representative parameterized simulation model from the corresponding parameter sample.
[0094] For example, in Figure 1As illustrated in the diagram and as an option illustration, method 100 may further include: determining 161 product samples assigned to a representative parametric simulation model, wherein multiple representative product samples are obtained. The determination of the 161 product samples assigned to the representative parametric simulation model can be based, in particular, on representative parameter samples of the parametric simulation model. For example, the multiple representative product samples can be determined by comparing them with a database of stored parameter samples of real product samples, particularly based on similarity relationships. In the case of unsatisfactory similarity, one or more instructions can be output: how to manufacture sufficiently satisfactory representative product samples from real product samples (e.g., through fitting, shimming, etc.).
[0095] exist Figure 2 In the exemplary implementation, steps 160 and 161 are referred to as “defining representative samples based on representative PUM and ODD points”. Here, representative PUM again represents multiple representative parametric simulation models, and ODD points represent ODD samples.
[0096] For example in Figure 1 As illustrated in the diagram and as an option, method 100 may further include: designing the product's regulating device 170 based on at least one representative parametric simulation model. The design of the product's regulating device 170 can be based, in particular, on multiple representative parametric simulation models. This allows for the realization of a robust regulating device concept and / or design for the product. In particular, this allows for the design of a particularly robust and reliable regulating device for the product. This, in particular, can improve the safety of the product.
[0097] For example in Figure 1 As illustrated in the diagram and as an option illustration, method 100 may further include: verifying one or more requirements of 171 on the product based on at least one representative parametric simulation model. Verification of one or more requirements of 171 on the product can, in particular, be based on multiple representative parametric simulation models. This allows for robust and reliable product design. It also particularly improves product safety.
[0098] For example in Figure 1 As illustrated in the diagram and as an option illustration, the method 100 may further include: identifying one or more of the 172 representative parameterized simulation models and / or representative parameter samples that have a significant impact on the product and / or its behavior (e.g., turning behavior). This can be done, for example, through sensitivity analysis. This can improve the understanding of the product and its behavior. This can be particularly useful in the case of high-dimensional ODDs, since the ODD can no longer be directly observed.
[0099] Figure 2An exemplary implementation of method 100 is shown. In this case, for example, the following steps may be performed sequentially: - "Sampling of ODDs" - "Determine PUM for all ODD samples based on PAM" - "Calculate the metric-based adjacency matrix for all PUM pairs" - "Search for representative PUM and ODD samples based on the adjacency matrix" - "Define representative samples based on representative PUM and ODD points".
[0100] Figure 3a This diagram illustrates an exemplary ODD with multiple parametric simulation models. That is, for each of these parametric simulation models, the associated parameter samples are represented as solid points in the ODD. In this example, the ODD is a two-dimensional product space, meaning it is expanded by two parameters (one in the x-direction and another in the y-direction), which can each take values from an interval. Generally, the ODD does not have to be a product space, but can be an (arbitrary) manifold. The dimension of the ODD can be arbitrary.
[0101] In this example, three representative parametric simulation models, 140 and 160, have been selected from multiple parametric simulation models or their parameter samples as representative parametric simulation models or their parameter samples in the ODD. The latter are represented by crosses in the ODD. In this example, three representative parametric simulation models, 140 and 160, were selected such that (according to the dissimilarity metric) the distance between each parametric simulation model and the corresponding representative parametric simulation model that is closest to it is minimized.
[0102] For this example, in 3D, specifically in the z-direction orthogonal to the Blattebene plane, the distance (according to the dissimilarity metric) between each parameter sample in the ODD (more precisely, its parametric simulation model) and the closest representative parametric simulation model can be shown. Instead, in 2D, some isolates (isolines) are shown in the z-direction for these parameters. For example, three solid lines constitute an isolate, i.e., those with the same z-value. Additionally, three other isolates (roughly extending around a cross) are shown. Besides the isolates, the boundaries of the influence regions of the corresponding representative parametric simulation models are plotted (dashed lines intersecting at a point).
[0103] Figure 3b The illustration is from Figure 3a An exemplary ODD has three selected representative parameterized simulation models (more precisely, their parameter samples).
[0104] Furthermore, a computer system is disclosed, the computer system being designed to perform a computer-implemented method 100 for determining a representative parameterized simulation model for a parameterizable simulation model of a product. The computer system may include a processor and / or working memory.
[0105] Furthermore, a computer program is disclosed, designed to perform a computer implementation of method 100 for determining a representative parameterized simulation model for a parameterizable simulation model of a product. For example, the computer program may exist in an interpretable form or in a compiled form. The computer program may (also in part) be loaded into the computer's RAM for execution, for example, as a sequence of bits or bytes.
[0106] Furthermore, a signal is disclosed, which is designed to encode (and ultimately execute) a computer-implemented method 100 for encoding (and ultimately executing) a representative parameterized simulation model of a parameterizable simulation model for a product. This signal may be a bit sequence or a byte sequence.
[0107] Furthermore, a computer-readable medium for storing and / or containing computer programs is disclosed. This medium may include, for example, one of RAM, ROM, EPROM, HDD, SSD, etc., on which signals are stored / contained.
Claims
1. A computer-implemented method (100) for determining a representative parametric simulation model of a parametric simulation model for a product, particularly a steer-by-wire system and / or a steering system for highly automated driving; wherein the parametric simulation model includes one or more system parameters and one or more operating parameters; the method (100) includes: - Compute (130) dissimilarity measures on one of the multiple pairs of parametric simulation models, respectively, where a distance is obtained for each pair, and thus multiple distances are obtained, wherein a change in one or more running parameters of one or both parametric simulation models of the corresponding pair makes the distance therein constant; -Based on the multiple distance selections (140), a predetermined number of parameterized simulation models are obtained, wherein multiple representative parameterized simulation models are obtained.
2. The method (100) of claim 1, wherein the dissimilarity measure between the two parameterized simulation models of the corresponding pair is based on other measures, particularly wherein the other measures are based on gap measures, v-gap measures and / or L2 measures.
3. The method (100) of claim 2, wherein the dissimilarity measure between two corresponding parameterized simulation models is based on a plurality of values, wherein the values are based on other measures between the other two parameterized simulation models respectively.
4. The method (100) of claim 3, wherein the dissimilarity measure between two parameterized simulation models of the corresponding pair is based on a statistical parameter of the maximum value or the distribution of the values.
5. The method (100) of claim 3, wherein the dissimilarity measure between the two parameterized simulation models of the corresponding pair is based on a constraint on the set of values of statistical parameter pairs based on the distribution of the one or more operating parameters and based on the maximum value on the constrained set of values.
6. The method (100) according to any one of the preceding claims, wherein the plurality of representative parametric simulation models are defined such that each parametric simulation model is minimized in distance from the corresponding representative parametric simulation model that is closest to it, particularly at a predetermined finite-dimensional norm.
7. The method (100) according to any one of the preceding claims, wherein the predetermined finite-dimensional norm is a p-norm for p in [1, +Inf], in particular a summation norm, a Euclidean norm, or a maximum norm.
8. The method (100) according to any one of claims 6 or 7, comprising: - Determine the representative measure of the representative parameterized simulation model (141) based on a predetermined number and / or minimum distance, particularly the minimum distance under the finite dimension norm; - Optionally, if the representative measure does not adequately meet the predetermined criteria, then an adaptation (150), in particular, is performed to increase the predetermined quantity.
9. The method (100) according to any one of the preceding claims, comprising: - Determine (110) multiple parameter samples in the operational design domain (ODD) for the product, wherein the parameter samples in the ODD each include at least one or more system parameters of the parameterizable simulation model and optionally one or more operational parameters; - The parameterized simulation model is constructed based on the parameterized simulation model and multiple parameter samples in the ODD, specifically wherein the parameterized simulation model is evaluated at one of the parameter samples respectively, and / or wherein a surrogate model for the parameterized simulation model is created for one of the parameter samples respectively.
10. The method (100) of claim 9, wherein (110) a plurality of parameter samples in the ODD are determined such that, particularly based on pseudo-random numbers, Latin hypercube sampling and / or Sobol sequences, the ODD is sufficiently uniformly covered in terms of one or more of its system parameters.
11. The method (100) according to claim 9 or 10, comprising: -The corresponding parameter samples are assigned to one or more system parameters of the representative parameterized simulation model and determined (160) as representative parameter samples of the parameterized simulation model.
12. The method (100) according to any one of the preceding claims, comprising: - Based on multiple distances, construct an (131) adjacency matrix.
13. The method (100) of claim 12, wherein a predetermined number of parameterized simulation models based on a plurality of distance selections (140) comprise: - Perform a full factorial search in the adjacency matrix; - Cluster the adjacency matrix; and / or - Convert the adjacency matrix into equidistant auxiliary points and cluster the auxiliary points.
14. The method (100) according to any one of the preceding claims, comprising: - Determine (160) the product samples assigned to the representative parameterized simulation model, wherein multiple representative product samples are obtained, particularly representative parameter samples based on the parameterized simulation model.
15. The method (100) according to any one of the preceding claims, comprising: - Design the regulating device of the product (170) based on at least one representative parametric simulation model, or optionally based on multiple representative parametric simulation models; and / or - Verify one or more requirements of the product (171) based on at least one representative parametric simulation model, and optionally based on multiple representative parametric simulation models; and / or - Identify (172) one or more of the representative parameterized simulation models and / or the representative parameter samples that have a significant impact on the product.
