Methods and systems for controlling production systems

By jointly training machine learning modules across product versions and those specific to product versions, the problems of high computational workload and low accuracy in existing technologies are solved, achieving efficient product performance optimization.

CN117795519BActive Publication Date: 2026-11-17SIEMENS AG
View PDF 1 Cites 0 Cited by

Patent Information

Application Number
CN202280055186.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Priority Date
2021-08-06
Filing Date
2022-08-02
Publication Date
2026-11-17
Estimated Expiration
2042-08-02

AI Technical Summary

Technical Problem

Existing technologies involve a large computational workload and low accuracy of proxy models in product design optimization, making it difficult to efficiently optimize product performance.

Method used

A joint training method is adopted, which combines cross-product version machine learning modules and product version-specific machine learning modules. By utilizing transfer learning, linear regression models, neural networks, and other technologies, product performance can be predicted and optimized.

Benefits of technology

It significantly improves training efficiency and optimization speed, increases performance prediction accuracy, and reduces the need for training data for new product versions.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN117795519B_ABST
    Figure CN117795519B_ABST
Patent Text Reader

Abstract

To control the production system (PS) used to manufacture the first product version (PV1), multiple product version-specific training datasets are read in for each product version. These training datasets include: design datasets (DR1, DR2) specifying design variants for the corresponding product version; and performance values ​​(P1, P2) for that design variant. The design datasets (DR1, DR2) are fed into a cross-product version machine learning module (LG_3). The output signal (OS) of this cross-product version machine learning module (LG_3) is fed not only into a first product version-specific machine learning module (L1_3) but also into a second product version-specific machine learning module (L2_3). These machine learning modules (LG_3, L1_3, L2_3) are jointly trained such that the output data (O1) of the first machine learning module (L1_3) reproduces the performance value (P1) of the first product version, and the output data (O2) of the second machine learning module (L2_3) reproduces the performance value (P2) of the second product version. Then, multiple integrated design datasets are generated and fed into a trained cross-product version machine learning module (LG_3). The output signal (OS) obtained by the trained cross-product version machine learning module (LG_3) is fed into a trained first machine learning module (L1_3). Based on the output data (O1) obtained by the trained first machine learning module, a performance-optimized design dataset (ODR) is derived from these integrated design datasets, and the performance-optimized design dataset is output to manufacture the first product version (PV1).
Need to check novelty before this filing date? Find Prior Art

Description

Background Technology

[0001] Computer-aided design systems are becoming increasingly prevalent in the manufacture of complex technological products, such as robots, electric motors, turbines, turbine blades, internal combustion engines, tools, motor vehicles, or components thereof. These systems typically generate design data that specifies in detail the product to be manufactured, and this design data allows control over the production systems used to manufacture the specified product.

[0002] To improve desired product characteristics, efforts are often made to automatically optimize product design data, which are typically expressed as performance parameters. These performance parameters may specifically relate to a product's power, output, speed, weight, runtime, accuracy, failure rate, resource consumption, efficiency, effectiveness, pollutant emissions, stability, wear, service life, physical properties, mechanical properties, electrical properties, constraints to be met, or other target parameters.

[0003] To optimize the performance of a product to be manufactured, multidimensional optimization methods, also known as MDO (Multi Dimensional Optimization) methods, are typically used. These MDO methods usually simulate multiple design variations of the product specified by the design data, and specifically select those design variations that optimize the simulated performance.

[0004] However, such simulations typically require a considerable amount of computation, especially since a large number of design variations often need to be evaluated. To reduce the required computational effort, so-called surrogate models are often used. These surrogate models are trained, in particular, using machine learning methods to predict relevant simulation results without requiring detailed simulations. However, such surrogate models often have low or fluctuating accuracy. Summary of the Invention

[0005] The objective of this invention is to describe a method and system for controlling a production system, which allows for more efficient design optimization.

[0006] This task is accomplished by a method according to the invention, a system according to the invention, a computer program product according to the invention, and a computer-readable storage medium according to the invention.

[0007] To control the production system to manufacture the first product version based on design datasets for the first and second product versions, multiple product version-specific training datasets are read in. These training datasets include: design datasets specifying design variants for the corresponding product versions; and performance values ​​quantifying the performance of the design variants. These design datasets are fed into a cross-product version machine learning module. The output signal of the cross-product version machine learning module is fed not only into a first product version-specific machine learning module but also into a second product version-specific machine learning module. According to the invention, these machine learning modules are jointly trained such that the output data of the first machine learning module reproduces the performance values ​​of the first product version, and the output data of the second machine learning module reproduces the performance values ​​of the second product version. Then, multiple integrated design datasets are generated and fed into the trained cross-product version machine learning module. The output signal obtained by the trained cross-product version machine learning module is fed into the trained first machine learning module. Based on the output data obtained by the trained first machine learning module, a performance-optimized design dataset is derived from these integrated design datasets, and the performance-optimized design dataset is output to manufacture the first product version.

[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 methods and systems 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] This invention allows the use of data from optimizations and training performed on other, especially previous, product versions to predict the performance of a first product version to be manufactured, and to optimize that first product version. This application of previous optimization or training results to machine training for a new application is often referred to as transfer learning.

[0011] In many cases, this cross-product version machine learning module can be used to learn the common interrelationships between design and performance across different product versions very efficiently. Training data from previous product versions can often be used very advantageously. Conversely, these product version-specific machine learning modules can take into account the differences between these product versions.

[0012] This approach typically leads to a significant improvement in training efficiency, especially when product versions are similar. In many cases, less training data is required to manufacture a new, first-generation product version. Furthermore, it often improves optimization speed and / or prediction accuracy.

[0013] Advantageous embodiments and extensions of the present invention are described below.

[0014] According to an advantageous embodiment of the invention, for multiple design datasets, performance values ​​of design variants specified by the respective design datasets can be determined by simulation. The respective design datasets, together with the corresponding predicted performance values, can then be used as training datasets. Various efficient simulation methods are available for determining performance values ​​by simulation, particularly for mechanical, electrical, and / or thermal simulations of products or their components.

[0015] According to another advantageous embodiment of the invention, a first product specification to be satisfied by a first product version and a second product specification to be satisfied by a second product version can be read in. During joint training of these machine learning modules, the first product specification can be fed into the first machine learning module, and the second product specification can be fed into the second machine learning module. Then, when the resulting output signal is fed into the trained first machine learning module, the first product specification can also be fed into the trained first machine learning module accordingly. In this way, the correlation between product specifications, design datasets, and / or performance values ​​can be modeled or learned through training, and these correlations can be used to predict performance.

[0016] Furthermore, these product-version-specific machine learning modules can each implement a linear regression model, and the cross-product-version machine learning module can implement a Gaussian process. In this way, the linear correlation between performance values ​​across different product versions can be learned efficiently. Since linear relationships can typically be modeled with relatively few parameters, training often requires relatively little training data.

[0017] Alternatively or additionally, these machine learning modules may each comprise a neural network. Here, the cross-product version machine learning module can be implemented in particular using a multilayer perceptron. These product version-specific machine learning modules may each comprise one or more neural layers. Through these neural layers, the non-linear correlations between performance values ​​of different product versions can also be learned efficiently.

[0018] According to an advantageous extension of the invention, multiple machine learning module groups can be provided, each group having one cross-product version machine learning module and at least two product version-specific machine learning modules. Furthermore, the number or set of available training datasets can be determined. Then, based on the determined number or set, one of these groups can be selected to perform training and derive a performance-optimized design dataset. In this way, machine learning module groups particularly suited to the number or set of available training data can be selectively chosen. This is especially beneficial for cases with relatively limited training data, such as one or more linear regression models.

[0019] It can also check whether product specifications exist for the first product version and the second product version, respectively.

[0020] Here, it is also possible to check whether these product specifications involve product parameters common to both product versions. If the check is positive, then the following machine learning module group can be selected, and the product specifications should be fed into the product version-specific machine learning module of that machine learning module group. Attached Figure Description

[0021] The embodiments of the present invention will then be described in more detail with reference to the accompanying drawings. Here, schematic diagrams are provided respectively:

[0022] Figure 1 It clarifies the design system for controlling production systems used to manufacture products;

[0023] Figure 2 The design system according to the present invention during the training phase is illustrated;

[0024] Figure 3 It clarifies the different groups of machine learning modules; and

[0025] Figure 4 The design system according to the present invention, which is in the application stage, is explained.

[0026] If the same or corresponding reference numerals are used in these figures, those reference numerals denote the same or corresponding entities, which in particular can be realized or designed as described in conjunction with the relevant figures. Detailed Implementation

[0027] Figure 1A schematic diagram illustrates the design system DS coupled to the production system PS. The design system DS is used for computer-aided design of a new version of a product PV1 with optimized performance, and for controlling the production system PS to manufacture product version PV1. In this case, the production system PS may include production facilities, robots, machine tools, and / or other equipment for manufacturing or processing the product PV1 or its components based on the design data. The product PV1 to be manufactured may, in particular, be an electric motor, robot, machine tool, turbine blade, wind turbine, gas turbine, motor vehicle or other machinery, other technical structures, or components of these products.

[0028] The corresponding design or design variant of product version PV1 to be manufactured is specified through design data in the form of a corresponding design dataset. Such design datasets can specify, in particular, the geometry, structure, characteristics, manufacturing steps, materials and / or components of product version PV1.

[0029] According to the present invention, the design system DS is designed to be able to automatically generate, to a large extent, a practical design dataset ODR that optimizes the performance of product version PV1. In this application, "optimization" should also be understood as an approximation of the optimal value. Here, the performance may specifically relate to the power, output, speed, weight, runtime, accuracy, failure rate, resource consumption, vibration tendency, temperature resistance, stiffness, material fatigue, efficiency, effectiveness, pollutant emissions, stability, wear, service life, physical properties, mechanical properties, electrical properties, constraints to be met, or other target parameters of product PV1.

[0030] For the purposes described above, the design system DS is trained using machine learning methods—as explained in more detail below—based on the available design dataset of the first product version PV1 to be manufactured and additionally based on the available design dataset of at least one other product version.

[0031] As a result of the optimization, the design system DS generates a performance-optimized design dataset ODR and transmits it to the production system PS. Based on the performance-optimized design dataset ODR, the production system PS manufactures a performance-optimized new product version PV1 according to the performance-optimized design dataset ODR.

[0032] Figure 2 The design system DS according to the invention is illustrated during the training 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 the data to be processed.

[0033] To train the design system DS, multiple training datasets are read in or generated for the first product version PV1 to be manufactured, and additionally, multiple training datasets are read in or generated for at least one other product version. This at least one other product version may, in particular, be a previous version of the product to be manufactured, for which performance data or other empirical values ​​already exist. For clarity, only the data and processing steps for two product versions are explicitly presented and described in the accompanying figures.

[0034] The training datasets specific to the first product version PV1 include: a design dataset DR1 specifying a design variant of the first product version PV1; and a performance value P1 quantifying the performance of the design variant. Correspondingly, the training datasets specific to the second product version include: a design dataset DR2 specifying a design variant of the second product version; and a performance value P2 quantifying the performance of the design variant.

[0035] In addition, for the first product version PV1, the first product specification E1 that PV1 should satisfy and the second product specification E2 that the second product version should satisfy are also read in. Product specifications E1 and E2, which are specific to the product version, are fixed and pre-defined, and unlike the design dataset, cannot be optimized. Product specifications E1 and E2 can be used to specify, for example, product type, product geometry, product dimensions, product environment, operating environment, operating specifications, boundary conditions, constraints, installation location, ambient temperature range, and / or other environmental data.

[0036] For the current embodiment, it should be assumed that: multiple design datasets DR1 and product specification E1 of the first product version PV1, and multiple design datasets DR2 with relevant performance values ​​P2 and product specification E2, are read from the database DB by the design system DS.

[0037] The database DB contains multiple training datasets for evaluating this performance, particularly for previous second-generation product versions. Here, the performance value P2 can come from measurements of previous product versions or simulations of those previous versions.

[0038] It should also be assumed that no performance indication exists for the design variant of the first product version PV1 specified by the design dataset DR1. For this reason, the design dataset DR1 and product specification E1 are fed to the simulator SIM of the design system DS. The simulator SIM is used to simulate the design variant specified by the fed design dataset (here, DR1), and to determine the corresponding performance (here, P1) of the corresponding design variant of the first product version PV1 in a simulated manner. Various known simulation methods or simulators for mechanical, electrical, or thermal simulations are available to perform such simulations.

[0039] Design dataset DR1, along with the associated performance value P1 determined through simulation and product specification E1, is transmitted from simulator SIM to selection module SEL of design system DS. Design dataset DR2, with associated performance value P2 and product specification E2, is also fed into selection module SEL.

[0040] Multiple machine learning module groups SL1, SL2, and SL3 are coupled to, or at least addressable by, the selection module SEL. The purpose of the selection module SEL is to select one of these groups, SL1, SL2, or SL3, based on the existence of product specifications (here, E1 and E2) and the set or number of available training datasets. For this purpose, the selection module SEL determines the number or set of training datasets. In particular, the number or set of training datasets with high-performance values ​​can be determined, provided that such training datasets are especially important for evaluating high-performance design variants. Furthermore, the selection module SEL checks whether the product specifications (here, E1 and E2) are available for both product versions and involves product parameters common to both product versions.

[0041] exist Figure 3 The diagram schematically illustrates different machine learning module groups SL1, SL2, and SL3. In this case, group SL1 includes a cross-product version machine learning module LG_1; and product version-specific machine learning modules L1_1 and L2_1. Correspondingly, group SL2 includes a cross-product version machine learning module LG_2; and product version-specific machine learning modules L1_2 and L2_2. Furthermore, group SL3 includes a cross-product version machine learning module LG_3; and product version-specific machine learning modules L1_3 and L2_3. Here, product version-specific machine learning modules L1_1, L1_2, and L1_3 are assigned to the first product version PV1, while product version-specific machine learning modules L2_1, L2_2, and L2_3 are assigned to the second product version.

[0042] The corresponding groups SL1, SL2, or SL3, when selected, should be trained to predict the performance of the first product version PV1 and the second product version, as specified herein, based on the input design dataset.

[0043] For this purpose, design datasets DR1 and DR2 are fed as input data to the relevant groups SL1, SL2, or SL3, respectively. Output data O1 and O2 are derived and output from the input data DR1 and DR2 through the relevant groups SL1, SL2, or SL3. In the current embodiment, product specifications E1 and E2 are additionally fed as input data to group SL3.

[0044] These output data, O1 and O2, are intended to reproduce, as accurately as possible, the actual performance values ​​of the design variants specified by the design datasets DR1 and DR2.

[0045] Training should generally be understood as optimizing the mapping from input data (here, DR1, DR2, and E1 and E2 if necessary) to output data (here, O1 and O2). This mapping is optimized during the training phase according to pre-given criteria. In this case, for example, prediction error or reproduction error can be used as the criterion in the case of a prediction model, classification error can be used as the criterion in the case of a classification model, and / or deviation from the optimized control action can be used as the criterion in the case of a control model. Through this training, for example, model parameters, process parameters, the network structure of neurons in the neural network, and / or the weights of the connections between these neurons can be adjusted or optimized to satisfy these criteria as well as possible. Therefore, this training can be understood as an optimization problem. For such optimization problems, there are a variety of efficient optimization methods available. In particular, genetic optimization, gradient descent, backpropagation, and / or particle swarm optimization can be used.

[0046] In the current context, the following fact should be leveraged for this training: similar product versions exhibit similar interrelationships or correlations between their design datasets and performance values. Common interrelationships or correlations can typically be learned very efficiently by training a shared machine learning module on different but similar product versions. Furthermore, different interrelationships or correlations can be efficiently mapped across different product versions using product version-specific machine learning modules. In this way, efficient, so-called transfer learning can be achieved, where training results or data from previous product versions are advantageously used for the current product version to be manufactured, in this case, PV1.

[0047] The different machine learning module groups SL1, SL2, and SL3 allow you to select the group SL1, SL2, or SL3 that is particularly suitable for the corresponding training conditions, based on the available training data or product specifications.

[0048] A more detailed description follows. Figure 3 The groups SL1, SL2, and SL3 are shown as examples.

[0049] In group SL1, the cross-product version machine learning module LG_1, shared by both product versions, preferably implements a Gaussian process and / or a Bayesian model, while the product version-specific machine learning modules L1_1 and L2_1 implement linear regression models, respectively. Data sets DR1 and DR2 are designed as input data and fed into the shared machine learning module LG_1. The output signal OS derived from the input data DR1 and DR2 by the shared machine learning module LG_1 is fed into both machine learning modules L1_1 and L2_1. From this output signal OS, output data O1 is derived and output through machine learning module L1_1, and output data O2 is derived and output through machine learning module L2_1. This group of machine learning modules is often referred to as a common-region model.

[0050] Linear regression models L1_1 and L2_1 can be used to learn the linear relationship between performance values ​​of different product versions very efficiently. Such linear regression models can also be modeled with relatively few parameters; in many cases, a relatively small amount of training data is sufficient to optimize such linear regression models. Therefore, in the current embodiment, if the amount of available training data is relatively small and, for example, below a pre-given threshold, the selection module SEL selects group SL1.

[0051] In the current embodiment, group SL2 includes an artificial neural network, particularly a multilayer perceptron, as a machine learning module LG_2 common to these product versions. Product version-specific machine learning modules L1_2 and L2_2 are each designed as an artificial neural network consisting of one or more layers. In this case, these layers can function linearly or non-linearly. Design datasets DR1 and DR2 are fed as input data into the common machine learning module LG_2. The common machine learning module LG_2 derives an output signal OS, which is fed into both machine learning modules L1_2 and L2_2. From the output signal OS, output data O1 is derived and output through machine learning module L1_2, and output data O2 is derived and output through machine learning module L2_2.

[0052] Unlike group SL1, product version-specific machine learning modules L1_2 and L2_2 can also effectively model or reproduce the non-linear relationships between the performance of different product versions. However, the greater flexibility achieved in this way typically requires more modeling parameters than in the linear case. Therefore, if a larger amount of training data is available or exceeds a pre-defined threshold, the selection module SEL preferably chooses group SL2.

[0053] In the current embodiment, group SL3 includes: a machine learning module LG_3 common to both product versions; and product version-specific machine learning modules L1_3 and L2_3. Machine learning modules LG_3, L1_3, or L2_3 can be implemented using artificial neural networks and / or Gaussian processes, respectively. Unlike groups SL1 and SL2, group SL3 can also explicitly handle product version-specific specifications (here, E1 and E2) and model these product version-specific specifications regarding their correlation with design datasets and / or performance values.

[0054] Design datasets DR1 and DR2 are fed as input data into a shared machine learning module LG_3, which derives an output signal OS and feeds it into two machine learning modules L1_3 and L2_3. Additionally, product specification E1 and product specification E2 are fed as additional input data into machine learning module L1_3 and machine learning module L2_3, respectively. Output data O1 is derived and output from the output signal OS and product specification E1 via machine learning module L1_3. Correspondingly, output data O2 is derived and output from the output signal OS and product specification E2 via machine learning module L2_3. In many cases, because product version-specific machine learning modules L1_3 and L2_3 additionally consider product version-specific product specifications E1 and E2, the correlation between product specifications, design datasets, and performance values ​​can be identified and learned. However, because product specifications E1 and E2 offer additional distinguishing possibilities, more training data is typically required for efficient training compared to training other groups SL1 and SL2. Correspondingly, if product versions-specific specifications (here, E1 and E2) exist and the amount of training data exceeds a pre-defined threshold, the selection module SEL preferably selects group SL3.

[0055] Instead of groups SL1, SL2, and SL3, or in addition to these groups, one or more other groups of machine learning modules operating in a similar manner may be provided. Furthermore, the corresponding groups or corresponding machine learning modules may also include: Bayesian neural networks, recurrent neural networks, convolutional neural networks, autoencoders, deep learning architectures, support vector machines, k-nearest neighbor classifiers, and / or decision trees.

[0056] For the current embodiment, it should be assumed that a large amount of training data is available, and therefore, the number of training datasets determined by the selection module SEL exceeds the threshold provided for group SL3. As long as the selection module SEL also finds that product specifications E1 and E2 exist for the two product versions and that these product specifications involve common product parameters, then—as Figure 2 As further clarified, group SL3 is selected via the selection module SEL. Correspondingly, design dataset DR1 with product specification E1 and design dataset DR2 with product specification E2 are selectively forwarded as input data to the selected group SL3 by the selection module SEL. As described above, group SL3 derives and outputs output data O1 and O2 from the input data DR1, E1, DR2, and E2.

[0057] Then, the output data O1 is compared with the performance value P1, and the output data O2 is compared with the performance value P2, and the deviation D is determined. Here, in particular, the Euclidean distance |O1-P1| between the output data O1 and the performance value P1, and the Euclidean distance |O2-P2| between the output data O2 and the performance value P2 can be calculated. Thus, the deviation D can be determined, for example, as the sum of these distances according to D = |O1–P1| + |O2–P2|, as a weighted sum of individual deviations, or as other combinations of individual deviations.

[0058] like Figure 2 As shown by the dashed arrow, the bias D is fed back to group SL3. Based on the fed-back bias D, the machine learning modules LG_3, L1_3, and L2_3 contained in group SL3 are jointly trained to minimize the bias D. Several efficient numerical methods are available to perform this minimization, such as gradient descent, backpropagation, particle swarm optimization, and / or genetic optimization.

[0059] By minimizing the bias D, the machine learning modules LG_3, L1_3, and L2_3 of group SL3 are trained to predict the relevant performance values ​​P1 or P2 of two product version design variants using their output data O1 or O2, given a pre-given design dataset DR1 or DR2 and product specification E1 or E2.

[0060] The cross-product version machine learning module LG_3 can efficiently learn the commonalities between different product versions. Conversely, the differences between these product versions can be mapped by the product version-specific machine learning modules L1_3 and L2_3.

[0061] If, instead of group SL3, the selection module SEL selects one of the other groups, SL1 or SL2, then obviously the training can be performed in a similar manner.

[0062] Figure 4 The design system DS, trained to generate performance-optimized design datasets ODR, is described. For this purpose, multiple synthetic design datasets DR1 are generated by the generator GEN of the design system DS, each specifying a design variant of the first product version PV1 to be manufactured. This generation is at least partially randomized.

[0063] These integrated design datasets DR1 are fed into the selection module SEL by the generator GEN. Additionally, the product specification E1 of the first product version PV1 is also fed into the selection module SEL. Next, the selection module SEL selectively feeds the integrated design datasets DR1 and product specification E1 into the selected and trained group SL3. Here, these integrated design datasets DR1 are fed as input data to the common machine learning module LG_3. Then, the output signal OS obtained from the common machine learning module LG_3, along with product specification E1, is fed as input data into the machine learning module L1_3 specific to the first product version PV1 to be manufactured. The output data O1 derived from the machine learning module L1_3 is output and fed into the optimization module OPT of the design system DS. Furthermore, these integrated design datasets DR1 are also fed into the optimization module OTP.

[0064] The optimization module OPT derives the performance-optimized design dataset ODR from these integrated design datasets DR1 based on the input output data O1, i.e., based on the predicted performance values. Thus, for example, the integrated design dataset that predicts the highest performance value O1 can be selected as the performance-optimized design dataset ODR. Alternatively, the performance-optimized design dataset ODR can be interpolated from multiple integrated design datasets that have the highest or particularly high predicted performance values ​​O1.

[0065] Then, output the performance-optimized design dataset ODR derived from these comprehensive design datasets DR1, in order to manipulate the production system PS, such as by combining... Figure 1 As described, to manufacture an optimized product version PV1.

Claims

1. A computer-implemented method for controlling a production system (PS) to manufacture the first product version (PV1) based on design datasets (DR1, DR2) for a first product version and a second product version, wherein... a) reading in a plurality of product version-specific training data sets, the plurality of product version-specific training data sets each comprising: Specify the design dataset (DR1, DR2) for the design variant of the corresponding product version. And the performance values ​​(P1, P2) that quantify the performance of this design variant. b) The design datasets (DR1, DR2) are fed into the cross-product version machine learning module (LG_3). c) The output signal (OS) of the cross-product version machine learning module (LG3) is fed not only into the product version-specific first machine learning module (L1_3) but also into the product version-specific second machine learning module (L2_3). d) These machine learning modules (LG_3, L1_3, L2_3) are jointly trained such that the output data (O1) of the first machine learning module (L1_3) reproduces the performance value (P1) of the first product version, and the output data (O2) of the second machine learning module (L2_3) reproduces the performance value (P2) of the second product version, wherein the second product version is a previous product version. e) Generate multiple integrated design datasets and feed these datasets into a trained cross-product version machine learning module (LG_3). f) The output signal (OS) obtained by the trained cross-product version machine learning module (LG_3) is fed into the trained first machine learning module (L1_3), and g) Based on the output data (O1) obtained by the first machine learning module (L1_3) after training, derive the performance-optimized design dataset (ODR) from the comprehensive design dataset, and output the performance-optimized design dataset to manufacture the first product version (PV1).

2. The method according to claim 1, characterized in that, For multiple design datasets (DR1), the performance value (P1) of the design variant specified by the corresponding design dataset (DR1) is determined by simulation. The corresponding design dataset (DR1) and the corresponding predicted performance values ​​(P1) are used together as the training dataset.

3. The method according to claim 1 or 2, characterized in that, Read in the first product specification (E1) to be satisfied by the first product version (PV1) and the second product specification (E2) to be satisfied by the second product version. When jointly training these machine learning modules (LG_3, L1_3, L2_3), the first product specification (E1) is fed into the first machine learning module (L1_3), and the second product specification (E2) is fed into the second machine learning module (L2_3), and... When the obtained output signal (OS) is fed into the trained first machine learning module (L1_3), the first product specification (E1) is also fed into the trained first machine learning module (L1_3).

4. The method according to claim 1 or 2, characterized in that, The cross-product version machine learning modules implement Gaussian processes, and the product version-specific machine learning modules each implement linear regression models.

5. The method according to claim 1 or 2, characterized in that, These machine learning modules each include a neural network.

6. The method according to claim 1 or 2, characterized in that... Multiple machine learning module groups (SL1, SL2, SL3) are provided. Each of these machine learning module groups has one cross-product version machine learning module (LG_1, LG_2, LG_3) and two product version-specific machine learning modules (L1_1, L2_1, L1_2, L2_2, L1_3, L2_3). Determine the number or set of available training datasets, and Based on the determined number or set, select one of the groups (SL1, SL2, SL3) to perform the training and derive the performance-optimized design dataset (ODR).

7. The method according to claim 6, characterized in that, Check whether product specifications (E1, E2) exist for the first product version and the second product version, respectively. If the check result is positive, select the group (SL1, SL2, SL3) that should feed the product specifications (E1, E2) into its product version-specific machine learning module.

8. A system (DS) for controlling a production system (PS) to manufacture a first product version (PV1) based on design datasets (DR1, DR2) of a first product version and a second product version, said system being configured to perform the method according to any one of the preceding claims.

9. A computer program product configured to perform the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program product according to claim 9.

Citation Information

Patent Citations

  • Agricultural product quality nondestructive testing method based on cross-domain spectral information and generalizable system

    CN112651173A