CONTROLS OF A PRODUCTION SYSTEM

Description

[0001] Computer-aided design systems are increasingly used in the production of complex technical products, such as robots, motors, turbines, turbine blades, internal combustion engines, tools, motor vehicles, or their components. These design systems typically generate design data that specifies the product to be manufactured in detail and can be used to control production systems for manufacturing that product.

[0002] To optimize desired product characteristics, the design data for the product is often optimized automatically. These product characteristics or target characteristics can relate to performance, power, yield, speed, weight, runtime, precision, error rate, resource consumption, efficiency, emissions, stability, wear, lifespan, physical property, mechanical property, chemical property, electrical property, magnetic property, constraints, or other target product parameters.

[0003] To optimize target parameters for a product, multidimensional optimization methods, so-called MDO methods (MDO: Multi-Dimensional Optimization), are frequently used. These MDO methods typically simulate a large number of design variants of the product specified by design data and specifically select those that optimize the simulated target parameters.

[0004] However, such simulations often require considerable computational effort, especially since a large number of design variants frequently need to be evaluated. To reduce the required computational effort, so-called surrogate models are often used. These models are trained, particularly through machine learning methods, to predict relevant simulation results or target variables without detailed simulation. However, such surrogate models often exhibit low or fluctuating accuracy.

[0005] EP3722977A1 ​​relates to a computer-implemented method and apparatus for generating a design for a technical system or product, wherein a design for the technical system or product is generated depending on a set of first parameters (P1) specifying physical properties and second parameters (P2) specifying perceptible properties of the technical system or product.

[0006] WO2020053991A1 discloses a device for supporting the design of a manufacturing system in order to create a design plan for a manufacturing system that meets the requirements of manufacturing.

[0007] WESTERMANN PAUL ET AL: "Using Bayesian deep learning approaches for uncertainty-aware building energy surrogate models" (2021-03-01), reveals the training of two types of Bayesian models, neural dropout networks and stochastic variational Gaussian process models, to emulate a complex high-dimensional building energy simulation problem.

[0008] The object of the present invention is to provide a method and a system for controlling a production system that allow for more efficient design optimization.

[0009] This problem is solved by a method with the features of claim 1, by a system with the features of claim 9, by a computer program product with the features of claim 10, and by a computer-readable storage medium with the features of claim 11.

[0010] To control a production system for manufacturing a product optimized with respect to several target variables, a large number of test data sets are imported. Each test data set comprises an initial design data set specifying a design variant of the product, as well as initial target values ​​quantifying the target variables of that design variant. Here, and in the following, optimization is understood to mean an approximation of an optimum. Regarding the initial target values, an initial ranking of the first design data sets is determined. Furthermore, several design evaluation modules are provided, each for predicting target values ​​based on design data sets. These design evaluation modules predict target values ​​for the initial design data sets.For each design evaluation module, a second ranking of the first design data sets is determined with regard to the predicted target values, along with the deviation of each second ranking from the first ranking. Based on these deviations, a design evaluation module is selected. A large number of second design data sets are then generated, for each of which the selected design evaluation module predicts second target values. Based on these second target values, a target-optimized design data set is derived from the second design data sets and transmitted to the production system for product manufacturing. The product is then manufactured according to this target-optimized design data set.

[0011] To determine the first and / or second ranking for the first design data sets, a Pareto optimization is performed using the target variables as Pareto target criteria, thereby determining a Pareto front. For each first design data set, a respective distance is calculated. eg A Euclidean distance to the Pareto front is determined. This establishes the first and / or second ranking of the initial design data sets according to their distance to the Pareto front. Specifically, a smaller distance can be assigned a higher rank than a larger distance. In this way, multiple independent target variables can be naturally considered when determining a ranking of design data sets. Furthermore, the above procedure scales largely without problems with the number of target variables in many cases.

[0012] To carry out the method according to the invention, a system for controlling a production system, a computer program product and a computer-readable, preferably non-volatile storage medium are provided.

[0013] The method and system according to the invention can be carried out or implemented, for example, using one or more computers, processors, application-specific integrated circuits (ASICs), digital signal processors (DSPs) and / or so-called "Field Programmable Gate Arrays" (FPGAs).

[0014] The invention allows, in particular, the selection of design evaluation modules for design optimization that reproduce as accurately as possible a ranking of design data sets derived from predefined data. Such design evaluation modules can thus be advantageously used to reliably identify a target-value-optimized design data set as such, or to distinguish it from less optimal design data sets. In particular, even a design evaluation module that systematically misjudges the predicted target values ​​can still be very advantageously applicable, provided that it reproduces the actual ranking as accurately as possible.

[0015] Advantageous embodiments and further developments of the invention are specified in the dependent claims.

[0016] According to an advantageous embodiment of the invention, machine learning modules can be provided as design evaluation modules. These modules have been trained using training datasets that do not match the test datasets to reproduce corresponding training target values ​​based on a training design dataset. Such trained machine learning modules typically require significantly fewer computing resources than detailed simulation models.

[0017] Nevertheless, simulation modules can also be provided as design evaluation modules, each of which uses a design data set specifying a design variant to predict the target values ​​of that variant through simulation. In this way, simulation modules can also be evaluated with regard to their accuracy in predicting rankings using the invention.

[0018] According to a particularly advantageous embodiment of the invention, a design evaluation module can output a statistical distribution of target values ​​for a given first design dataset to predict these values. For this purpose, a design evaluation module can implement a Bayesian neural network and / or a Gaussian process. Such a statistical distribution can be specified, in particular, by a mean, median, variance, standard deviation, uncertainty, reliability, probability distribution, distribution type, and / or trend of the target values. Based on the respective statistical distribution, a sample of target values ​​can then be selected, particularly randomly. This allows the determination of the second ranking with respect to the selected target value samples.Furthermore, the selected target value samples, the determined second-order rankings, and / or the determined deviations can be aggregated over several iterations of process step e) of claim 1 to select a design evaluation module. This aggregation can include averaging, integration, and / or the calculation of a minimum, maximum, quantile, and / or percentile. By considering the statistical distributions of the target values, their respective uncertainties can be estimated. In this way, more robust or reliable design variants can be selected.

[0019] According to a further advantageous embodiment of the invention, the deviation of the respective second ranking from the first ranking can be determined using a Kendall tau metric.

[0020] Advantageously, the Kendall tau metric can be weighted for this purpose. Specifically, when applying the Kendall tau metric, an initial design data set with a smaller distance to the Pareto front can be weighted more heavily than an initial design data set with a larger distance to the Pareto front. In this way, better designs can be weighted more heavily than worse designs when comparing the rankings. This is advantageous because ultimately, the ranking of the best or near-best designs is decisive for selecting the target-optimized design data set.

[0021] An embodiment of the invention is explained in more detail below with reference to the drawing. The drawings illustrate, in schematic form: Figure 1: a design system controlling a production system for manufacturing a product, Figure 2: a design system according to the invention in a configuration phase, Figure 3: predicted distributions of target values, Figure 4: a determination of a Pareto front for design data sets, and Figure 5: the design system according to the invention in an application phase.

[0022] In the figures, identical or corresponding reference symbols denote identical or corresponding entities that may be described, implemented, or realized in connection with the figure in question.

[0023] Figure 1Figure 1 schematically depicts a design system DS coupled to a production system PS. The design system DS is used for the computer-aided design of a product P optimized with respect to several target parameters and for controlling the production system PS to manufacture product P. The production system PS can comprise a manufacturing plant, a robot, a machine tool, and / or other equipment for manufacturing or processing product P or a component thereof based on design data. The manufactured product P can be an engine, a robot, a turbine blade, a wind turbine, a gas turbine, a motor vehicle, or another technical structure or component of these products.

[0024] Each design or design variant of the product P to be manufactured is specified by design data in the form of a design data set. Such a design data set can, in particular, specify a geometry, a structure, a property, a production step, a material, and / or a component of the product P.

[0025] According to the invention, the design system DS is to be enabled to generate a realistic design data set (ODR) optimized for product P with respect to several predefined target parameters, largely automatically. These target parameters can relate in particular to performance, power, yield, speed, quality, weight, runtime, precision, defect rate, resource consumption, efficiency, vibration tendency, stiffness, thermal conductivity, aerodynamic efficiency, material fatigue, emissions, stability, wear, service life, physical properties, mechanical properties, chemical properties, electrical properties, magnetic properties, and / or other design criteria or constraints to be met by product P.

[0026] The design system DS generates such a target-value-optimized design data set ODR and transmits it to the production system PS. Based on the target-value-optimized design data set ODR, the production system PS is instructed to manufacture a target-value-optimized product P according to the target-value-optimized design data set ODR.

[0027] Figure 2 Figure 1 illustrates a design system DS according to the invention in a configuration phase. The design system DS has one or more processors PROC for executing the method according to the invention and one or more memory MEMs for storing data to be processed.

[0028] The design system DS according to the invention is to be configured to evaluate design data sets as accurately as possible with regard to the target parameters to be optimized for the product P. In particular, a predicted ranking of the design data sets, oriented towards the target parameters, should reproduce a predetermined, actual ranking as accurately as possible. For this purpose, the design system DS evaluates a plurality of provided design evaluation modules EV1, ..., EVN with regard to their respective accuracy in reproducing these rankings.

[0029] In this embodiment, the design evaluation modules EV1, ..., EVN are implemented using machine learning modules, for example, artificial neural networks. The machine learning modules EV1, ..., EVN are trained using known and evaluated training datasets to reproduce corresponding training target values, particularly in the form of a statistical distribution of target values, based on a training design dataset. For training, training design datasets, for example, vectors of design parameters for the product P, are fed into each respective machine learning module EV1, ..., or EVN as input data. The resulting output data of each machine learning module EV1, ..., or EVN is compared with the corresponding training target values, for example, a vector of target values, and any deviation is minimized through the training of the respective machine learning module.

[0030] In this context, training is generally understood as the optimization of a mapping of input data to output data in a machine learning module. This mapping is optimized according to predefined criteria during a training phase. For example, in prediction models, criteria could include a prediction error; in classification models, a classification error; or in control models, the success of a control action. Through training, the network structures of neurons in a neural network and / or the weights of connections between neurons can be adjusted or optimized to best meet the predefined criteria. Training can thus be viewed as an optimization problem. A multitude of efficient optimization methods are available for such optimization problems in the field of machine learning.In particular, gradient descent methods, particle swarm optimizations and / or genetic optimization methods can be used.

[0031] In the present embodiment, the design evaluation modules EV1, ..., EVN are preferably implemented as so-called Bayesian neural networks. Such Bayesian neural networks can be understood, among other things, as statistical estimators. As such, a Bayesian neural network predicts a statistical distribution (VPD) of target values ​​instead of a point prediction. In this way, in addition to an estimate of the target values, information about their uncertainty is also obtained. Such statistical distributions can be characterized, in particular, by means and variances.

[0032] Figure 3Such predicted statistical distributions (VPD) are illustrated for different target variables T1 and T2, with respect to which the product P is to be optimized. The target variables T1 and T2 could, for example, be the weight and vibration tendency of a turbine blade. Figure 3 Each predicted statistical distribution (VPD) is illustrated as a contour line diagram of a probability density for a given target value pair.

[0033] Efficient training methods for Bayesian neural networks can be found, for example, in the publication "Pattern recognition and machine learning" by Christopher M. Bishop, Springer 2011.

[0034] After the design evaluation modules EV1, ..., EVN have been trained as described above, - as Figure 2Further illustrated – known and evaluated test datasets for testing the design evaluation modules EV1, ..., EVN are read from a database DB linked to the design system DS. To avoid overfitting effects, the test datasets are not taken from the set of training datasets used for training. Analogous to the training datasets, the test datasets each comprise a first design dataset DR1 specifying a design variant of the product P, as well as first target values ​​V1 quantifying the target variables of this design variant.

[0035] The first design data sets DR1, along with their respective initial target values ​​V1, are fed into a Pareto optimizer OPTP of the design system DS. The Pareto optimizer OPTP is used to perform a Pareto optimization.

[0036] Pareto optimization is a multi-criteria optimization in which several different target criteria, known as Pareto target criteria, are considered independently. In this example, the target variables to be optimized constitute the Pareto target criteria. The result of the Pareto optimization is a so-called Pareto front.

[0037] Figure 4 This illustrates the determination of a Pareto front (PF) for the first design data sets DR1. Figure 4 The first design data sets DR1 are each plotted as circles in a coordinate system with the target values ​​T1 and T2 as the coordinate axes, according to their assigned target values. For clarity, only a few of the first design data sets DR1 are marked with a reference symbol as examples.

[0038] The Pareto front (PF) is formed by those solutions to a multi-criteria optimization problem where one objective criterion cannot be improved without worsening another. A Pareto front thus represents, in a sense, a set of optimal compromises. In particular, solutions not included in the Pareto front (PF) can still be improved with respect to at least one objective criterion and can therefore be considered suboptimal.

[0039] In Figure 4 The first design data records DR1 of the Pareto front PF are illustrated by filled circles, and the first design data records DR1 not belonging to the Pareto front PF are illustrated by unfilled circles. As shown from Figure 4As can be seen, the Pareto optimization is performed towards larger target values ​​for the target variables T1 and T2. A variety of efficient standard routines, especially those from machine learning, are available for such Pareto optimizations and for determining a Pareto front.

[0040] As in Figure 2As indicated, the Pareto optimizer OPTP determines a Pareto front PF for the first design data records DR1 with respect to the first target values ​​V1. Based on the determined Pareto front PF, the distance to the Pareto front PF is then calculated for each of the first design data records DR1. This distance can be, for example, a Euclidean distance. The first design data records DR1 are then sorted according to their distance to the Pareto front PF. This sorting creates an initial ranking R1 of the first design data records DR1. That is, the initial ranking R1 orders the first design data records DR1 according to their distance to the Pareto front PF, where a smaller distance corresponds to a higher rank than a larger distance. The determined initial ranking R1 serves as a benchmark for the subsequently determined rankings.

[0041] According to the invention, a second ranking R2(1), ..., R2(N) is determined for each of the design evaluation modules EV1, ..., EVN. For this purpose, the first design data records DR1 are entered into a respective trained design evaluation module EV1, ... .... or EVN is fed in. The respective design evaluation module EV1, ... or EVN consequently outputs a predicted statistical distribution of the target values ​​for each initial design data record DR1. From each output statistical distribution, a corresponding target value sample VP1, ... or VPN is then drawn according to this respective statistical distribution. VP1 here denotes the target value sample drawn from the output of the design evaluation module EV1, and VPN, correspondingly, the target value sample drawn from the output of the design evaluation module EVN.

[0042] The target value samples VP1, ..., VPN are fed into the Pareto optimizer OPTP. The latter determines, for each of the target value samples VP1, ..., VPN, a design evaluation module-specific Pareto front PF1, ... or PFN of the first design data records DR1, as described above. Furthermore, for each design evaluation module EV1, ..., EVN and for each first design data record DR1, its distance to the respective Pareto front PF1, ... or PFN is determined, also as described above. Finally, for each design evaluation module EV1, ... or EVN, a second ranking R2(1), ... or R2(N) is determined according to the calculated distances, as described above.

[0043] The second rankings R2(1), ..., R2(N) are then compared with the first ranking R1. In each case, a deviation D(1), ... or D(N) of the respective second ranking R2(1), ... or R2(N) from the first ranking R1 is determined. The respective deviation D(i), i = 1, ..., N, is preferably determined using a modified Kendall tau metric KT, according to D(i) = KT(R2(i), R1), i = 1, ..., N. The Kendall tau metric KT is modified in that higher ranks in the compared rankings are weighted more heavily, for example, with a factor of 1 / (R+1), where R denotes a respective rank. A smaller value of the rank R corresponds to a higher rank in the respective ranking.

[0044] The deviations D(1), ..., D(N) are fed into a selection module SEL of the design system DS. The selection module SEL serves to select the design evaluation module(s) EV1, ..., EVN that best reproduces the ranking of the initial design data sets DR1. For this purpose, the deviations D(1), ..., D(N) are evaluated by the selection module SEL. To account for the stochastic properties of the target values, the above-mentioned drawing of the target value samples VP1, ..., VPN and their subsequent processing are repeated several times. The resulting deviations D(1), ..., D(N) are then averaged over these repetitions. In addition to a mean or median, a variance or confidence interval of the deviations D(1), ..., D(N) is preferably also calculated.

[0045] Depending on the input deviations D(1), ..., D(N), the selection module SEL determines an index IMIN of the design assessment module EV1, ..., EVN that, on average, exhibits the lowest or at least a lower deviation D(IMIN) than the other design assessment modules EV1, ..., EVN. The design assessment module selected from the design assessment modules EV1, ..., EVN by the index IMIN is subsequently referred to as EVS.

[0046] Apparently, the selected design evaluation module EVS can reproduce a target-oriented ranking of design data sets better than the other design evaluation modules EV1, ..., EVN. In practice, it turns out that a design evaluation module selected in this way can generally perform very robust evaluations of design data sets.

[0047] A corresponding application of the DS design system with the selected EVS design evaluation module is achieved by Figure 5This is illustrated. The DS design system has a generator GEN for generating synthetic, secondary design data sets DR2, each specifying a design variant of the product P to be optimized. The generation of these secondary design data sets DR2 can also be randomly induced, in order to explore previously unknown areas of possible design variants. The secondary design data sets DR2 are fed by the generator GEN into the selected design evaluation module EVS and into an optimization module OPT of the DS design system.

[0048] As described above, the selected design evaluation module EVS predicts second target values ​​V2 for each second design dataset DR2, each in the form of a statistical distribution. A given statistical distribution can be defined, in particular, by a mean and its uncertainty. The second target values ​​V2 are fed by the selected design evaluation module EVS into the optimization module OPT. Based on the input data DR2 and V2, the optimization module OPT selects one or more of the second design datasets DR2 with the highest, or at least higher, second target values ​​V2 and / or with a lower uncertainty than other second design datasets DR2.

[0049] In particular, the second design data set (DR2) that exhibits the maximum weighted combination of the target values ​​(V2) and their uncertainties can be output as the target-optimized design data set (ODR). Alternatively or additionally, a target-optimized design data set (ODR) can be interpolated from several second design data sets (DR2) selected according to the above criteria.

[0050] The target-optimized design data set ODR is ultimately output by the optimization module OPT and can be used, as described in the context of Figure 1 described, used to control the production system PS in order to produce the design-optimized product P.

Claims

1. Computer-implemented method for controlling a production system (PS) for producing a product (P) optimized with respect to multiple target variables (T1, T2), wherein a) a plurality of test datasets are read in, each comprising a first design dataset (DR1) specifying a design variant of the product (P), and first target values (V1) quantifying the target variables of that design variant, b) a first ranking (R1) of the first design datasets (DR1) is determined with respect to the first target values (V1), c) multiple design evaluation modules (EV1,...,EVN) each for predicting target values on the basis of design datasets are provided, d) target values (VP1,...,VPN) for each of the first design datasets (DR1) are predicted by the design evaluation modules (EV1,...,EVN), e) for each design evaluation module (EV1,...,EVN) - a respective second ranking (R2) of the first design datasets (DR1) with respect to the predicted target values (VP1,...,VPN) is determined, and - a respective deviation (D) of the respective second ranking (R2) from the first ranking (R1) is determined, f) one design evaluation module (EVS) is selected depending on the determined deviations (D), g) a plurality of second design datasets (DR2) is generated, for each of which second target values (V2) are predicted by the selected design evaluation module (EVS), and, h) depending on the second target values (V2), a target-value-optimized design dataset (ODR) is derived from the second design datasets (DR2) and transmitted to the production system to produce the product (P) in order to produce the product (P) in accordance with the target-value-optimized design dataset (ODR), wherein, to determine the first and / or the respective second ranking (R1, R2), - Pareto optimization is performed for the first design datasets (DR1) using the target variables (T1, T2) as Pareto target criteria, wherein a Pareto front (PF) is determined, - a respective distance from the Pareto front (PF) is determined for each first design dataset (DR1), and - the first and / or second ranking (R1, R2) of the first design datasets (DR1) is / are determined according to their distance from the Pareto front (PF).

2. Method according to Claim 1, characterized in that machine learning modules, which have been trained by means of training datasets to reproduce corresponding training target values based on a training design dataset, are provided as design evaluation modules (EV1,...,EVN) and in that the test datasets are different from the training datasets.

3. Method according to either one of the preceding claims, characterized in that simulation modules, each of which simulatively predicts the target values of that design variant on the basis of a design dataset specifying one design variant, are provided as design evaluation modules (EV1,...,EVN).

4. Method according to one of the preceding claims, characterized in that each design evaluation module (EV1,...,EVN) for predicting target values for a respective first design dataset (DR1) outputs a statistical distribution (VPD) of those target values, in that a respective target-value sample (VP1,...,VPN) is selected on the basis of the respective statistical distribution (VPD), in that the respective second ranking (R2) is determined with respect to the selected target-value samples (VP1,...,VPN), and in that the selected target-value samples (VP1,...,VPN), the determined second rankings (R2) and / or the determined deviations (D) are aggregated over multiple iterations of method step e) for the selection of a design evaluation module.

5. Method according to Claim 4, characterized in that a respective statistical distribution (VPD) is specified by means of a mean, a median, a variance, a standard deviation, an uncertainty figure, a reliability figure, a probability distribution, a distribution type, and / or a curve specification.

6. Method according to one of the preceding claims, characterized in that the deviation of the respective second ranking (R2) from the first ranking (R1) is determined using a Kendall-tau metric.

7. Method according to Claim 6, characterized in that in the use of the Kendall-tau metric, a first design dataset (DR1) at a smaller distance from the Pareto front (PF) has a higher weighting than a first design dataset (DR1) at a greater distance from the Pareto front (PF).

8. Method according to one of the preceding claims, characterized in that each design evaluation module (EV1,...,EVN) comprises an artificial neural network, a Bayesian neural network, a recurrent neural network, a convolutional neural network, a multilayer perceptron, an autoencoder, a deep-learning architecture, a support vector machine, a data-driven trainable regression model, a k-nearest-neighbour classifier, a physical model and / or a decision tree.

9. System (DS) for controlling a production system (PS) for producing a product (P) optimized with respect to multiple target variables (T1, T2), configured for implementing a method according to one of the preceding claims.

10. Computer program product configured for implementing a method according to one of Claims 1 to 8.

11. Computer-readable storage medium having a computer program product according to Claim 10.