Method and system for controlling a production system

Abstract

A multitude of test data sets is read in, each comprising a first design data set (DR1) of a construction variant of a specified product (P) and a first target value (V1) quantifying a target variable of the construction variant. A first ranking (R1) of the first design data set (DR1) is determined with respect to the first target value (V1). Furthermore, a plurality of design evaluation modules (EV1,..., EVN) is provided, each for predicting a target value based on a design data set. Target values (VP1,..., VPN) of the first design data sets (DR1) are predicted by the design evaluation modules (EV1,..., EVN), respectively. Then, for the respective design evaluation module (EV1,..., EVN), a respective second ranking (R2) of the first design data sets (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. Then, a design evaluation module (EVS) is selected depending on the determined deviations (D). Furthermore, a multitude of second design data sets (DR2) is generated, for each of which a second target value (V2) is predicted by the selected design evaluation module (EVS). Then, from the second design data sets, a target value-optimized design data set (ODR) is derived depending on the second target values (V2) and output for manufacturing the product (P).

Description

Background Technology

[0001] Computer-aided design systems are being used on an increasingly large scale to manufacture complex technological products such as robots, engines, turbines, turbine blades, internal combustion engines, tools, motor vehicles, or components thereof. These design systems typically generate detailed design data specifying the product to be manufactured, which in turn allows control of production systems used to manufacture that specific product.

[0002] To optimize desired product characteristics, the design data of the product is typically optimized automatically. Product characteristics or target characteristics may refer to the product's performance, power, output, speed, weight, operation, accuracy, error rate, resource consumption, efficiency, effectiveness, pollutant emissions, stability, wear, service life, physical properties, mechanical properties, chemical properties, electrical properties, magnetic properties, secondary conditions to be met, or other target variables.

[0003] To optimize the target variables of a product, multidimensional optimization methods, also known as MDO methods, are often used. These MDO methods typically simulate a large number of construct variants of the product specified by design data, and specifically select construct variants of the simulated target variables for optimization.

[0004] However, such simulations typically require a significant amount of computation, especially since a vast number of construction variants must usually be evaluated. To reduce the required computation, so-called surrogate models are often used. These surrogate models are trained, in particular, through machine learning methods to predict relevant simulation outcomes or target variables without requiring detailed simulations. However, these surrogate models often have low or fluctuating accuracy. Summary of the Invention

[0005] The objective of this invention is to describe methods and systems for controlling production systems, which allow for more efficient design optimization.

[0006] This task is accomplished by a method having the features of the present invention, a system having the features of the present invention, a computer program product having the features of the present invention, and a computer-readable storage medium having the features of the present invention.

[0007] To control the production system to manufacture a product optimized for multiple target variables, a large number of test datasets are read in. Each test dataset includes a first design dataset specifying a variant of the product and a first target value quantifying the target variable of that variant. Here and below, optimization is also understood as approaching the optimal value. A first ranking of the first design dataset is determined with respect to the first target value. Furthermore, multiple design evaluation modules are provided, each used to predict the target value based on the design dataset. The target values ​​of the first design dataset are predicted by the design evaluation modules respectively. Then, for the corresponding design evaluation module, a corresponding second ranking of the first design dataset with respect to the predicted target value is determined, and a corresponding deviation between the corresponding second ranking and the first ranking is determined. A design evaluation module is then selected based on the determined deviation. Furthermore, a large number of second design datasets are generated, and a second target value is predicted for each second design dataset by the selected design evaluation module. Then, based on the second target value, a target value-optimized design dataset is derived from the second design dataset and output for manufacturing the product.

[0008] To perform 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.

[0009] The method and system according to the invention can be executed or implemented, for example, by means of one or more computers, processors, application-specific integrated circuits (ASICs), digital signal processors (DSPs), and / or so-called "field-programmable gate arrays" (FPGAs).

[0010] With the aid of this invention, a design evaluation module can be selected for design optimization, which reproduces the ordering of the design dataset derived from pre-given data as accurately as possible. Therefore, such a design evaluation module can be advantageously used to relatively reliably identify a target-value-optimized design dataset as a target-value-optimized design dataset or to distinguish a target-value-optimized design dataset from a less optimized design dataset. Here, it is particularly advantageous to apply design evaluation modules that systematically misestimate the predicted target value, provided that the design evaluation module only reproduces the actual ordering as well as possible.

[0011] Advantageous embodiments and extensions of the present invention are described below.

[0012] According to an advantageous embodiment of the invention, a machine learning module can be provided as a design evaluation module. This machine learning module has been trained using a training dataset inconsistent with the test dataset to reproduce the associated training target values ​​based on the training design dataset. Such a trained machine learning module typically requires significantly fewer computational resources in its application than a detailed simulation model.

[0013] Nevertheless, a simulation module can also be provided as a design evaluation module, each simulation module predicting the target value of a specified construct variant based on a design dataset. In this way, the accuracy of the simulation module in predicting ranking can also be evaluated using this invention.

[0014] According to a particularly advantageous embodiment of the invention, the corresponding design evaluation module can output the statistical distribution of the target values ​​of the corresponding first design dataset to predict the target values. For this purpose, the corresponding design evaluation module can implement a Bayesian neural network and / or a Gaussian process. This statistical distribution can be specified, in particular, by the mean, median, variance, standard deviation, uncertainty specification, reliability specification, probability distribution, distribution type, and / or variation process specification of the target values. Based on the corresponding statistical distribution, a corresponding sample of target values ​​can then be selected, particularly randomly. A corresponding second ranking can then be determined for the selected sample of target values. Furthermore, the selected sample of target values, the determined second ranking, and / or the determined deviation can be aggregated during multiple runs of the method of the invention for use in selecting the design evaluation module. In this case, the aggregation can include averaging, integrating, and / or forming minimum values, maximum values, fractions, and / or percentages. By considering the statistical distribution of the target values, the uncertainty of each target value can be estimated. In this way, more robust or reliable construction variants can be selected, in particular.

[0015] According to another advantageous embodiment of the invention, in order to determine the first ranking and / or the corresponding second ranking, Pareto optimization can be performed on the first design dataset using the target variable as the Pareto objective criterion. Here, the Pareto front can be determined. Then, the corresponding distance to the Pareto front, such as the Euclidean distance, can be determined for the corresponding first design dataset. Thus, the first ranking and / or second ranking of the first design dataset can be determined based on the distance between the first design dataset and the Pareto front. In particular, smaller distances can be assigned higher ranks in the ranking than larger distances. In this way, multiple independent target variables can be naturally considered when determining the ranking of the design dataset. Furthermore, in many cases, the above process can scale largely without problems with the number of target variables.

[0016] According to another advantageous embodiment of the invention, the deviation between the corresponding second sort and the first sort can be determined by means of the Kendall-tau metric.

[0017] Therefore, it is advantageous to weight the Kendall-tau metric. Specifically, when applying the Kendall-tau metric, higher weights can be assigned to the first design dataset that is closer to the Pareto front than to the first design dataset that is farther from the Pareto front. In this way, when comparing rankings, better designs can have higher weights than worse designs. This is advantageous because the ranking of the best or near-best designs is ultimately decisive for the selection of the design dataset for objective value optimization. Attached Figure Description

[0018] Embodiments of the invention are explained in more detail below with reference to the accompanying drawings. These are illustrated schematically as follows: Figure 1 : A design system for controlling production systems used to manufacture products. Figure 2 The system according to the invention is in the configuration phase. Figure 3 : Predicted distribution of target value Figure 4 : Determining the Pareto front of the design dataset, and Figure 5 The design system according to the present invention is in the application stage.

[0019] In the accompanying drawings, the same or corresponding reference numerals denote the same or corresponding entities that can be implemented or realized as described in conjunction with the accompanying drawings. Detailed Implementation

[0020] Figure 1 A design system DS coupled to a production system PS is illustrated in the diagram. The design system DS is used to construct, with computer assistance, an optimized product P in terms of multiple objective variables, and to manipulate the production system PS to manufacture product P. In this case, the production system PS may include manufacturing facilities, robots, machine tools, and / or other devices for manufacturing or processing product P or its components based on design data. The manufactured product P may be an engine, a robot, turbine blades, a wind turbine, a gas turbine, a motor vehicle, or other technical structures or components of these products.

[0021] The corresponding design or a corresponding design variant of the product P to be manufactured is specified through design data in the form of a corresponding design dataset. This design dataset can specifically specify the geometry, structure, characteristics, manufacturing steps, materials and / or components of product P.

[0022] According to the present invention, the design system DS should be able to generate, to a large extent automatically, a realistic and optimized design dataset ODR for multiple pre-given target variables of product P. In this case, the target variables may be, in particular, performance, power, output, speed, mass, weight, runtime, accuracy, error rate, resource consumption, efficiency, oscillation tendency, stiffness, thermal conductivity, aerodynamic efficiency, material fatigue, pollutant emissions, stability, wear, service life, physical properties, mechanical properties, chemical properties, electrical properties, magnetic properties, and / or other construction criteria or secondary conditions that product P must satisfy.

[0023] The design system DS generates an optimized design dataset ODR and transmits it to the production system PS. Based on the optimized design dataset ODR, the production system PS manufactures the optimized product P according to the optimized design dataset ODR.

[0024] Figure 2 A design system DS according to the invention is shown in the configuration phase. The design system DS has one or more processors PROC for executing the method according to the invention and one or more memories MEM for storing data to be processed.

[0025] The design system DS according to the present invention should be able to evaluate the design dataset as accurately as possible with respect to the target variable to be optimized for product P through configuration. In particular, the predicted ordering of the design dataset for the target variable should reproduce the pre-given actual ordering as accurately as possible. To this end, the reproduction accuracy of these orderings is evaluated by the numerous design evaluation modules EV1, ..., EVN provided by the design system DS.

[0026] In this embodiment, the design evaluation modules EV1, ..., EVN are implemented using machine learning modules (e.g., artificial neural networks). The machine learning modules EV1, ..., EVN are trained to reproduce the associated training target values, particularly in the form of a statistical distribution of the target values, based on a known and evaluated training dataset. For training, a training design dataset, for example, in the form of a vector of design parameters for product P, is fed as input data to the corresponding machine learning module EV1, ..., or EVN. The output data thus obtained by the corresponding machine learning module EV1, ..., or EVN is compared with the associated training target values ​​(e.g., a vector of target values), and the bias is minimized through training of the involved machine learning modules.

[0027] In this context, training should generally be understood as the optimization of the mapping from input data to output data of a machine learning module. This mapping is optimized during the training phase according to pre-given criteria. These criteria could be, for example, prediction error for a prediction model, classification error for a classification model, or the success of control actions for a control model. Through training, specifically by adjusting or optimizing the network structure of neurons in the neural network and / or the weights of connections between neurons, the pre-given criteria are satisfied as much as possible. Therefore, this training can be viewed as an optimization problem. For such optimization problems in machine learning, a large number of efficient optimization methods are available. In particular, gradient descent, particle swarm optimization, and / or genetic optimization methods can be used.

[0028] In this embodiment, the evaluation modules EV1, ..., EVN are preferably implemented as a so-called Bayesian neural network. This Bayesian neural network can be particularly viewed as a statistical estimator. Therefore, the Bayesian neural network predicts the statistical distribution VPD of the target value rather than point prediction. In this way, in addition to the estimation of the target value, a description of the uncertainty of that estimation is obtained. This statistical distribution can be characterized, in particular, by the mean and variance.

[0029] Figure 3 The diagram illustrates the predictive statistical distribution (VPD) of various target variables T1 and T2, for which product P should be optimized. Target variables T1 and T2 could be, for example, the weight and vibration trend of turbine blades. Figure 3 The corresponding predicted statistical distribution (VPD) is plotted as a contour map of the probability density of the corresponding target value pairs.

[0030] Effective training methods for Bayesian neural networks can be found, for example, in Christoper M. Bishop's 2011 Springer publication, "Pattern Recognition and Machine Learning".

[0031] After the design evaluation modules EV1, ..., EVN have been trained as described above, as follows: Figure 2 As further illustrated, known and evaluated test datasets for testing the design evaluation modules EV1, ..., EVN are read from the database DB coupled to the design system DS. To avoid overfitting, these test datasets are not from the training dataset set used for training. Similar to the training datasets, the test datasets each include a first design dataset DR1 specifying a variant of product P and a first target value V1 quantifying the target variable of that variant.

[0032] The first design dataset DR1, along with the assigned first objective value V1, is fed into the Pareto optimizer OPTP of the design system DS. The Pareto optimizer OPTP is used to perform Pareto optimization.

[0033] Pareto optimization is a multi-criteria optimization method that independently considers multiple different objective criteria, known as Pareto objective criteria. In this embodiment, the objective variable to be optimized forms the Pareto objective criteria. The result of Pareto optimization is the determination of the so-called Pareto front.

[0034] Figure 4 The determination of the Pareto front (PF) for the first design dataset DR1 is shown. Figure 4 In the diagram, the first design dataset DR1 is plotted as circles on a coordinate system with target variables T1 and T2 as axes, based on the target values ​​assigned to it. For clarity, only some of the first design dataset DR1 are labeled by a single figure number, as an example.

[0035] A Pareto front (PF) is formed from solutions to a multi-criteria optimization problem where one objective criterion cannot be improved without worsening another. Therefore, in a sense, the Pareto front forms a set of optimal compromise solutions. Specifically, solutions not included in the Pareto front (PF) are still improved with respect to at least one objective criterion and can therefore be considered suboptimal.

[0036] exist Figure 4 In the diagram, the first design data set DR1 of the Pareto front (PF) is shown as a solid circle, while the first design data set DR1 not belonging to the Pareto front (PF) is shown as a hollow circle. From Figure 4 As can be seen, Pareto optimization proceeds towards larger objective values ​​for objective variables T1 and T2. Numerous effective standard routines, particularly those for machine learning, can be used for this Pareto optimization and, consequently, for determining the Pareto front.

[0037] like Figure 2 As shown, the Pareto front (PF) of the first design dataset DR1 with respect to the first objective value V1 is determined using the Pareto optimizer OPTP. Then, based on the determined Pareto front (PF), the distance between each first design dataset DR1 and the Pareto front (PF) is determined. The distance can be, for example, Euclidean distance. The first design datasets DR1 are then sorted according to their distances to the Pareto front (PF). This sorting forms the first ranking R1 of the first design datasets DR1. This means that the first ranking R1 arranges the first design datasets DR1 according to their distances to the Pareto front (PF), where smaller distances correspond to larger distances and higher rankings. The determined first ranking R1 is used as a benchmark for subsequent determined rankings.

[0038] According to the present invention, a corresponding second ranking R2(1), ..., or R2(N) is also determined for each design evaluation module EV1, ..., EVN. For this purpose, a first design dataset DR1 is fed into the corresponding trained design evaluation module EV1, ..., or EVN. As a result, the corresponding design evaluation module EV1, ..., or EVN outputs a predicted statistical distribution of the target value for each first design dataset DR1. Then, a corresponding target value sample VP1, ..., or VPN is extracted from each output statistical distribution according to the corresponding statistical distribution. In this case, VP1 represents the target value sample extracted from the output of the design evaluation module EV1, and correspondingly, VPN represents the target value sample extracted from the output of the design evaluation module EVN.

[0039] The target value samples VP1, ..., VPN are fed into the Pareto optimizer OPTP. For each target value sample VP1, ..., VPN, the Pareto optimizer determines a Pareto front PF1, ..., or PFN specific to the design evaluation module for the first design dataset DR1, as described above. Furthermore, for each design evaluation module EV1, ..., EVN, and for each first design dataset DR1, its distance to the corresponding Pareto front PF1, ..., or PFN is also determined as described above. Finally, for the corresponding design evaluation module EV1, ..., or EVN, the corresponding second ranking R2(1), ..., or R2(N) is determined based on the determined distances, as described above.

[0040] Then, the second rankings R2(1), ..., R2(N) are compared with the first ranking R1. Here, the deviations D(1), ..., or D(N) between the corresponding second rankings R2(1), ..., or R2(N) and the first ranking R1 are determined respectively. In this case, the corresponding deviations D(i), i=1, ..., N, are determined according to D(i)=KT(R2(i), R1), i=1, ..., N, preferably by means of a modified Kendall-tau metric KT. The Kendall-tau metric KT is modified here to assign a higher weight to the higher ranking in each compared ranking, for example, using a factor of 1 / (R+1), if R represents the corresponding ranking. The smaller ranking value R in this case corresponds to the higher ranking in the corresponding ranking.

[0041] The deviations D(1), ..., D(N) are fed into the selection module SEL of the design system DS. The selection module SEL is used to select one or more design evaluation modules EV1, ..., EVN that best or as well as possible reproduce the order of the first design dataset DR1. For this purpose, the deviations D(1), ..., D(N) are evaluated by the selection module SEL. Here, in order to take into account the characteristics of random target values, the above extraction and further processing of the target value samples VP1, ..., VPN are repeated multiple times or frequently. The obtained deviations D(1), ..., D(N) are then averaged with respect to these repetitions. In addition to the mean or median, the variance or confidence interval of the deviations D(1), ..., D(N) is preferably also calculated.

[0042] Based on the input deviations D(1), ..., D(N), the selection module SEL determines the index IMIN of the following design evaluation modules EV1, ..., EVN, which has an average deviation D(IMIN) that is the smallest or at least smaller than the other design evaluation modules EV1, ..., EVN. The design evaluation module selected from the design evaluation modules EV1, ..., EVN by index IMIN is referred to as EVS below.

[0043] Clearly, the selected design evaluation module EVS reproduces the target value ranking of the design dataset better than other design evaluation modules EV1, ..., EVN. Practice has shown that design evaluation modules selected in this way typically provide very robust evaluations of design datasets.

[0044] Figure 5 A corresponding application of the design system DS with the selected design evaluation module EVS is illustrated. For this purpose, the design system DS has a generator GEN for generating a synthetic second design dataset DR2, each second design dataset specifying a construction variant of the product P to be optimized. The generation of the second design dataset DR2 can also be performed randomly, thereby opening up previously unknown regions of possible construction variants. The second design dataset DR2 is fed from the generator GEN into the selected design evaluation module EVS and then into the optimization module OPT of the design system DS.

[0045] As described above, the selected design evaluation module EVS predicts a second objective value V2 for the corresponding second design dataset DR2, with each second objective value V2 presented in the form of a statistical distribution. In this case, the corresponding statistical distribution can be given, in particular, by the mean and its uncertainty. The second objective value V2 is fed into the optimization module OPT by the selected design evaluation module EVS. Based on the input data DR2 and V2, the optimization module OPT selects one or more second design datasets DR2 that have the highest second objective value V2 or at least have a higher second objective value V2 than other second design datasets DR2 and / or have lower uncertainty than other second design datasets DR2.

[0046] Specifically, the following second design dataset in the second design dataset DR2 can be used as the output of the design dataset ODR for objective value optimization, which has the maximum weighted combination of the objective value V2 and its uncertainty. Alternatively or additionally, the design dataset ODR for objective value optimization can be derived by interpolation of multiple second design datasets DR2 selected according to the above criteria.

[0047] The design dataset ODR for target value optimization is ultimately output by the optimization module OPT, and, if combined with... Figure 1 The described method can be used to manipulate the production system PS to thereby manufacture a product P with optimized design.

Claims

1. A computer-implemented method for controlling a production system (PS) to manufacture a product (P) optimized with respect to multiple objective variables (T1, T2), wherein... a) Read in a large number of test datasets, each test dataset including a first design dataset (DR1) specifying a variant of the product (P) and a first target value (V1) quantifying the target variable of the variant. b) Determine the first ranking (R1) of the first design dataset (DR1) with respect to the first target value (V1). c) Provides multiple design evaluation modules (EV1, ..., EVN), each used to predict target values ​​based on the design dataset. d) Predict the target values ​​(VP1, ..., VPN) of the first design dataset (DR1) using the design evaluation module (EV1, ..., EVN). e) For the corresponding design evaluation modules (EV1, ..., EVN). - Determine the corresponding second ranking (R2) of the first design dataset (DR1) with respect to the predicted target values ​​(VP1, ..., VPN), and - Determine the corresponding deviation (D) between the second sort (R2) and the first sort (R1). f) Select the design evaluation module (EVS) based on the determined deviation (D). g) Generate a large number of second design datasets (DR2), and predict a second objective value (V2) for each second design dataset using the selected design evaluation module (EVS). h) Based on the second target value (V2), derive the target value optimized design dataset (ODR) from the second design dataset (DR2) and output it for manufacturing the product (P). This is to determine the first sort and / or the corresponding second sort (R1, R2). - Pareto optimization is performed on the first design dataset (DR1) using the target variables (T1, T2) as Pareto objective criteria, where the Pareto front (PF) is determined. - Determine the corresponding distance to the Pareto front (PF) for the corresponding first design dataset (DR1), and - Determine the first sort and / or second sort (R1, R2) of the first design dataset (DR1) based on the distance between the first design dataset (DR1) and the Pareto front (PF).

2. The method according to claim 1, characterized in that, A machine learning module is provided as part of the design evaluation module (EV1, ..., EVN), which has been trained on a training dataset to reproduce the associated training target values ​​based on the training design dataset. The test dataset is different from the training dataset.

3. The method according to any one of the preceding claims, characterized in that, Simulation modules are provided as design evaluation modules (EV1, ..., EVN), each simulation module simulates the target value of a specified construct variant based on a design dataset of that construct variant.

4. The method according to any one of claims 1 to 2, characterized in that, The corresponding design evaluation modules (EV1, ..., EVN) output the statistical distribution (VPD) of the target values ​​for the corresponding first design dataset (DR1) to predict the target values. Select the corresponding target value samples (VP1, ..., VPN) based on the corresponding statistical distribution (VPD). For the selected target value samples (VP1, ..., VPN), determine the corresponding second ranking (R2), and During multiple runs of method step e), the selected target value samples (VP1, ..., VPN), the determined second order (R2), and / or the determined deviation (D) are aggregated for use in selecting the design evaluation module.

5. The method according to claim 4, characterized in that, The corresponding statistical distribution (VPD) is specified by the mean, median, variance, standard deviation, uncertainty description, reliability description, probability distribution, distribution type and / or description of the process of change.

6. The method according to any one of claims 1 to 2, characterized in that, The Kendall-tau metric is used to determine the deviation between the corresponding second ranking (R2) and the first ranking (R1).

7. The method according to claim 6, characterized in that, When applying the Kendall-tau metric, the first design dataset (DR1) that is closer to the Pareto front (PF) is given a higher weight than the first design dataset (DR1) that is farther from the Pareto front (PF).

8. The method according to any one of claims 1 to 2, characterized in that, The corresponding design evaluation modules (EV1, ..., EVN) include artificial neural networks, Bayesian neural networks, recurrent neural networks, convolutional neural networks, multilayer perceptrons, autoencoders, deep learning architectures, support vector machines, data-driven trainable regression models, k-nearest neighbor classifiers, physical models, and / or decision trees.

9. A system (DS) for controlling a production system (PS) to manufacture a product (P) optimized in terms of a plurality of objective variables (T1, T2), said system being configured to perform the method according to any one of the preceding claims.

10. A computer program product configured to perform the method according to any one of claims 1 to 8.

11. A computer-readable storage medium having a computer program product according to claim 10.

Images