Methods and systems for controlling the manufacturing of products using production equipment

Abstract

A machine learning module (BNN) is provided, trained to generate a first performance signal (PF1) and a prediction uncertainty (UC) for the predicted performance of a design variant of a specified product (P) from a design data record (DR). Furthermore, various design data records (DRs) are generated, each specifying a design variant of the product (P). For a given design data record (DR), the following steps are performed: - generating the first performance signal (PF1) and the corresponding prediction uncertainty (UC); - running or skipping a simulation (SIM) to generate a second performance signal (PF2) based on the prediction uncertainty (UC), the second performance signal (PF2) quantifying the simulated performance of the corresponding design variant; and - if the simulation (SIM) is run, deriving a performance value (PV) from the second performance signal (PF2), otherwise deriving a performance value (PV) from the first performance signal (PF1). Based on the derived performance value (PV), a performance-optimized design data record (ODR) is determined and output to control the production equipment (PP).

Description

[0001] Today, the production of complex products (such as robots, electric motors, turbines, turbine blades, internal combustion engines, machining tools, vehicles, or their components) typically relies on sophisticated design systems. Such systems usually provide detailed design data specifying the product to be manufactured. This design data allows for the specialized control of modern production equipment to produce the specified product.

[0002] To optimize product performance for given objectives or constraints, it is often desirable to automatically optimize the product's design data. Such performance can be related to the product's power, output, speed, runtime, accuracy, error rate, resource consumption, efficiency, pollutant emissions, stability, wear, lifespan, and / or other target parameters. For the purpose of performance optimization, some design systems use so-called Multidimensional Optimization (MDO) tools. These tools typically simulate design variations of the product defined by various design data and seek those design data that optimize the simulated performance of the product.

[0003] However, such simulations typically consume significant computational resources to determine the performance of each design variant. To reduce computational demands, so-called surrogate models based on machine learning can be used to predict simulation results. However, these surrogate models generally have poor accuracy.

[0004] The purpose of this invention is to provide a method and system for controlling the manufacturing of products using production equipment, which allows for more efficient design optimization.

[0005] This objective is achieved by the method according to claim 1, the system according to claim 11, the computer program product according to claim 12, and the computer-readable storage medium according to claim 13.

[0006] According to the present invention, a machine learning module is provided, which is trained to generate a first performance signal and predict the prediction uncertainty of performance from design data records of a design variant of a specified product, wherein the first performance signal quantifies the predicted performance of the design variant. Here and hereinafter, components or sub-products of a composite product may also be considered as products. Furthermore, various design data records are generated, each specifying a design variant of the product. For a given design data record, the following steps are performed:

[0007] - The machine learning module generates a first performance signal and the corresponding prediction uncertainty.

[0008] - Based on the predicted uncertainty, run or skip simulations of the corresponding design variant, the simulations generating a second performance signal that quantifies the simulated performance of the design variant, and

[0009] If the simulation is run, performance values ​​are derived from the second performance signal; otherwise, performance values ​​are derived from the first performance signal. Based on the derived performance values, performance optimization design data records are determined from the various design data records. These performance optimization design data records are then output for controlling the production equipment.

[0010] In order to perform an inventive method, a system, a computer program product, and a non-transient computer-readable storage medium are provided.

[0011] Creative methods and / or creative systems may be implemented by one or more processors, computers, application-specific integrated circuits (ASICs), digital signal processors (DSPs), programmable logic controllers (PLCs), and / or field-programmable gate arrays (FPGAs).

[0012] This invention allows skipping particularly expensive simulations where the predictions of the machine learning module are likely to be accurate. In this way, the number of simulations and therefore the computational workload can be significantly reduced without significantly compromising optimization accuracy.

[0013] Specific embodiments of the present invention are defined by the dependent claims.

[0014] According to a preferred embodiment of the present invention, the machine learning module may include or implement a surrogate model, a Bayesian machine learning model, a Bayesian neural network, and / or a Gaussian process model. For the aforementioned models, many efficient learning methods are known, which allow training individual models to generate predictions and their corresponding uncertainties in a consistent and / or uniform manner.

[0015] According to an advantageous embodiment of the invention, the training of the machine learning module can continue by using a second performance signal as training data. Utilizing the second performance signal, the training of the machine learning module can be improved or updated by accurate simulation results.

[0016] According to another embodiment of the invention, the predictive and / or simulated performance of the design variant may depend on the achievement of one or more design objectives and / or on the compliance of the design variant with one or more design constraints. Such objectives may relate to the product's power, efficiency, speed, accuracy, resource consumption, stability, wear, and / or lifespan. For example, the objectives for a turbine may relate to the flow efficiency, temperature efficiency, and / or cooling efficiency of the turbine or its components. Various constraints may relate to, for example, temperature limits, pressure limits, speed limits, and / or other restrictions. The corresponding objectives or constraints may be specified by configuration data used to configure the simulation and / or training of the machine learning module.

[0017] According to another embodiment of the invention, the prediction uncertainty can be represented by a probability distribution, a discrete probability distribution, statistical variance, statistical standard deviation, confidence interval, and / or error interval. Such a probability distribution can be considered as a distribution of plausible predictions. In particular, the probability distribution can be represented by the mean and standard deviation (or variance) of the distribution.

[0018] According to another embodiment of the invention, the decision to run or skip the corresponding simulation can depend on a comparison of the corresponding prediction uncertainty with a threshold. If the threshold is exceeded, the corresponding simulation can be run. In this way, the uncertain prediction can be replaced by a more accurate simulation result.

[0019] According to another preferred embodiment of the invention, the decision to run or skip the corresponding simulation can take into account the corresponding first performance signal. Since the product design is optimized for its performance, higher uncertainty may be more acceptable when the predicted performance is lower than when the predicted performance is higher. In this way, the number of simulations can be further reduced.

[0020] According to another preferred embodiment of the invention, the decision to run or skip the corresponding simulation can take into account the deviation between the predicted performance quantized by the corresponding first performance signal and the predicted performance quantized by the previously generated first performance signal. In this way, the current predicted performance can be evaluated relative to the previous predicted performance.

[0021] Specifically, an upper performance threshold can be determined for the predicted performance quantized from the corresponding first performance signal and the corresponding prediction uncertainty. Similarly, a lower performance threshold can be determined for the predicted performance quantized from the previously generated first performance signal and the corresponding prediction uncertainty. Then, the decision to run or skip the corresponding simulation can be made by comparing the upper and lower performance thresholds. It seems reasonable that, during performance optimization, current design variants with upper performance thresholds lower than the lower performance thresholds of previous design variants can be ignored. Therefore, simulations can be skipped.

[0022] Furthermore, the upper performance threshold can be ranked within one or more lower performance thresholds determined for one or more previously generated first performance signals. This ranking can then be taken into account when deciding whether to run or skip the corresponding simulation. Specifically, the quantiles of the ranked performance thresholds can be specified. If the ranking of the upper performance threshold is higher than a given quantile, the simulation can be run; otherwise, the simulation is skipped.

[0023] Specific embodiments of the present invention are described below with reference to the accompanying drawings. These drawings are illustrated in schematic form:

[0024] Figure 1: Design system for controlling production equipment to manufacture products.

[0025] Figure 2: Creative design system in the training phase.

[0026] Figure 3: Data-driven forecasting and its forecast uncertainty, and

[0027] Figure 4: Product design optimization by a creative design system.

[0028] In the accompanying drawings, the same reference numerals denote the same or corresponding entities, which are preferably represented as described in various places.

[0029] Figure 1 A design system DS is schematically illustrated, which is coupled to and controls production equipment PP to manufacture product P. Production equipment PP may be or include manufacturing equipment, robots, machining tools, and / or other devices for manufacturing or processing products based on design data. The product to be manufactured may be a robot, electric motor, turbine, turbine blade, internal combustion engine, machining tool, vehicle, or a component thereof.

[0030] The design or design variant of the product P to be manufactured is specified by design data in the form of a Design Data Record (DR). In particular, the Design Data Record (DR) may specify the geometry, structure, properties, manufacturing steps, materials, components, and / or parts of the product P.

[0031] According to the present invention, the design system DS should be able to automatically generate a design data record DR, which is optimized for one or more given performance criteria of product P. In this respect, the term "optimization" should also include the meaning of being closer to the optimal. Such performance criteria may be related to the power, yield, speed, runtime, accuracy, error rate, oscillation tendency, resource consumption, aerodynamic efficiency, energy efficiency, pollutant emissions, stability, wear, lifespan and / or other design objectives or design constraints of product P.

[0032] The design system DS generates performance-optimized design data records (DRs) and transmits them to the production equipment PP. Using the performance-optimized design data records DRs, the production equipment PP is controlled to manufacture the performance-optimized product P specified by the performance-optimized design data records DRs. Many efficient computer-operated manufacturing tools are available for manufacturing the product specified by the design data.

[0033] Figure 2 illustrates a creative design system DS in the training phase. The design system DS includes one or more processors PROC for performing the method steps of the present invention and one or more data storage devices MEM for storing relevant data.

[0034] The design system DS is coupled to a database DB and also includes a machine learning module BNN, which is implemented as a Bayesian neural network. Such a Bayesian neural network can be considered a statistical estimator. The latter is used to determine statistical estimates for, for example, the mean, variance, standard deviation, and / or probability distribution, from empirical data of a sample of a statistical population. The Bayesian estimator can specifically optimize the posterior expected value of a loss function, cost function, reward function, or utility function.

[0035] Using known machine learning methods and sample data from a statistical population, a Bayesian neural network (BNN) can be trained to estimate the probability distribution of a given new data record for the statistical population. As a complement or alternative to the Bayesian neural network, the machine learning module BNN may include or implement a Gaussian process model, which can also be trained to estimate the probability distribution of a given new data record for the statistical population.

[0036] A probability distribution can be represented by its mean, variance, and / or standard deviation. In this case, the variance or standard deviation can be considered as the uncertainty of the mean. Additionally or alternatively, the uncertainty of a probability distribution can be specified by other measures such as the error interval, confidence interval, or the width of the probability distribution. In this sense, Bayesian neural networks and Gaussian process models can be viewed as uncertainty-aware machine learning models.

[0037] Some relevant machine learning methods for training uncertainty-aware machine learning models are described, for example, in “Pattern Recognition and Machine Learning” by Christopher M. Bishop, Springer 2011.

[0038] According to this embodiment, the Bayesian neural network (BNN) should be trained to generate a first performance signal that quantifies the predicted performance (i.e., expected performance) of a design variant of a given product P, as well as the prediction uncertainty of the predicted performance, from the design data record of the design variant.

[0039] Here, the term "training" generally means that the mapping from the input data of a machine learning module to the output data of that module is optimized during the training phase with respect to predetermined and / or learned criteria. In the present case, these criteria include estimating the performance of the design variant as accurately as possible, along with the uncertainty of the estimate. The mapping can be optimized by adjusting the mapping parameters of the machine learning module. In the case of artificial neural networks, the connection structure of its neurons and / or the weights of the connections between neurons can be altered to optimize the mapping. For such optimization, various numerical standard methods, such as gradient descent, particle swarm optimization, or genetic algorithms, are available.

[0040] To train the Bayesian Neural Network (BNN), a large amount of training data stored in the database DB is fed into the design system DS. The training data includes numerous Training Design Data Records (TDRs), each specifying a design variant of the product. Assigned to the corresponding TDR, the training data also includes a corresponding Training Performance Value (TPF). The latter quantifies the actual performance of the design variant specified by the corresponding TDR. As indicated above, performance can be related to the product's power, yield, speed, runtime, accuracy, error rate, oscillation tendency, resource consumption, aerodynamic efficiency, energy efficiency, pollutant emissions, stability, wear, lifespan, and / or other design objectives or constraints.

[0041] The Training Design Data Record (TDR) is input into the Bayesian Neural Network (BNN). Based on the corresponding TDR, the BNN generates an output signal PF, which is intended to quantify the performance of the design variant specified by the TDR. Furthermore, the BNN outputs the uncertainty UC of the output signal PF. The output signal PF is compared with the training performance value TPF corresponding to the TDR. The deviation D between the training performance value TPF and the performance quantized by the output signal PF is then fed back to the BNN, as indicated by the dashed arrow in Figure 2. Thus, the BNN is trained to minimize the deviation D and output the uncertainty UC, which reflects the actual statistics of the training data.

[0042] After training, the Bayesian Neural Network (BNN) can be used as a statistical estimator. Specifically, the trained BNN is capable of generating an output signal PF from design data records of design variants of a given product. This output signal PF quantifies the predictive performance of the design variant and can therefore be used as a first performance signal for the design data records. Furthermore, the predictive uncertainty UC of the predictive performance is generated in parallel with the first performance signal PF.

[0043] Figure 3 shows a graph where the predicted performance quantized by the first performance signal PF and the predicted uncertainty UC of that performance are schematically plotted for various design variants DV. The predicted performance is plotted as a solid line, with the corresponding uncertainty UC indicated by error bars. For clarity, the figure label UC is used to indicate only one error bar. Furthermore, the actual performance of the design variant DV is plotted as a dashed line.

[0044] Design variant DV is grouped into design ranges B1 and B2. It is assumed that the Bayesian neural network BNN is primarily trained on training data covering design range B1, while design range B2 is only slightly covered by training data. Therefore, it can be expected that the performance predictions of the Bayesian neural network BNN will be more accurate and / or less uncertain in design range B1 than in design range B2. This behavior is also reflected in Figure 3. The figure also shows that the actual performance lies within the uncertainty UC generated by the Bayesian neural network BNN. This is due to the inherent ability of Bayesian neural networks to automatically estimate the actual uncertainty of their predictions. In practice, the estimated uncertainty is typically higher in regions with poor training data coverage and / or containing a large amount of random data.

[0045] Therefore, the first performance signal PF and the corresponding uncertainty UC output by the trained Bayesian neural network BNN can indicate possible limitations or thresholds to the actual performance of the design variant, even in design regions where accurate prediction is not allowed.

[0046] Figure 4 illustrates the optimization of the design of product P by the design system DS. The design system DS includes a Bayesian neural network (BNN) trained as described above.

[0047] The design system DS also includes a design generator GEN, which generates synthetic design data records DR, each of which specifies a design variant of product P.

[0048] The design data record DR generated from the design generator GEN is fed as input data into a trained Bayesian neural network BNN. The trained Bayesian neural network BNN generates a first performance signal PF1 from the corresponding design data record DR, quantifying the predicted performance of the corresponding design variant, and generates the prediction uncertainty UC of the predicted performance in parallel.

[0049] The first performance signal PF1, the uncertainty UC, and the corresponding synthetic design data record DR are fed into the decision module DC of the design system DS. The decision module DC decides whether to run or skip the simulation of the corresponding design variant based on the corresponding uncertainty UC.

[0050] To perform such simulations, the design system DS includes a simulator SIM, which also receives the synthetic design data record DR from the design generator GEN. The simulator SIM is capable of performing physical simulations of design variants based on the corresponding design data record DR. Many efficient methods, such as the finite element method, are known for such physical simulations.

[0051] In the current context, the simulator SIM specifically determines the simulation performance of the corresponding simulated design variants. Performance can often be determined very accurately through simulation. However, such accurate physical simulations typically require substantial computational resources. Therefore, if performance can be reliably determined elsewhere, skipping expensive simulations seems advantageous.

[0052] According to the present invention, the uncertainty UC of the first performance signal PF1 predicted by the trained Bayesian neural network (BNN) is used to determine whether to run or skip the simulation. If the prediction uncertainty UC is low, the corresponding performance predicted by the trained Bayesian neural network (BNN) may be accurate enough, and therefore the simulation can be skipped. On the other hand, if the prediction uncertainty UC is high, the corresponding predicted performance may not be reliable enough. In this case, the simulation can be run to produce a more accurate performance prediction.

[0053] To make decisions in the simulation, the decision module DC also considers the first performance signal PF1. Specifically, the decision module DC uses the current first performance signal PF1 along with one or more previously generated first performance signals. An upper performance threshold UPF is calculated based on the current first performance signal PF1. Preferably, the upper performance threshold UPF should be greater than the actual performance with, for example, a predetermined probability of 95%. Such an upper limit can be easily derived from the corresponding uncertainty UC. In the simplest case, the uncertainty UC (possibly multiplied by a given factor) is added to the corresponding performance to produce a possible upper performance threshold UPF. Similarly, a lower performance threshold LPF is calculated based on the previously generated first performance signal and its corresponding uncertainty. Specifically, the corresponding lower performance threshold LPF can be chosen as the lower boundary of, for example, a 95% confidence interval for the corresponding performance. In the simplest case, the corresponding uncertainty (possibly multiplied by a given factor) is subtracted from the corresponding performance to produce a possible lower performance threshold LPF.

[0054] According to this embodiment, the upper performance threshold (UPF) is then ranked within the lower performance threshold (LPF). For this purpose, the performance thresholds UPF and LPF can be sorted. If the ranking of the upper performance threshold (UPF) is higher than a given quantile of, for example, 75% of the overall ranking, the decision module DC decides to run the simulation. Otherwise, the decision module DC decides to skip the simulation. If the performance of the current design variant is likely to be less than 25% of the performance of previously generated design variants, the current design variant can be ignored during performance optimization.

[0055] If the decision module DC decides to run the simulation, it sends a trigger signal TR to the simulator SIM. The trigger signal TR causes the simulator SIM to simulate the design variant specified by the corresponding design data record DR. As a result of the simulation, the simulator SIM generates a second performance signal PF2, which quantifies the simulated performance of the design variant. The second performance signal PF2 is then sent from the simulator SIM to the decision module DC.

[0056] If a simulation is run, the decision module DC selects the simulated performance quantized by the second performance signal PF2 as the performance value PV for the corresponding design variant. Otherwise, if the simulation is skipped, the decision module DC selects the predicted performance quantized by the first performance signal PF1 as the performance value PV for the corresponding design variant. The resulting performance value PV is assigned to the corresponding design data record DR.

[0057] The corresponding design data record (DR) and its corresponding performance value (PV) are transmitted from the decision module (DC) to the optimization module (OPT) of the design system (DS). Based on the received performance value (PV), the optimization module (OPT) selects one or more design data records (DRs) with the highest or a specific performance value (PV). Based on the one or more selected design data records (DRs), the optimization module (OPT) obtains the performance-optimized design data record (ODR) or interpolates it. The performance-optimized design data record (ODR) is then output to control the production equipment (PP) as described above.

[0058] According to an advantageous implementation, the optimization module OPT can influence the design generator GEN to drive the generation of design data records (DRs) toward design variants with higher performance. This influence is indicated by the dashed arrows in Figure 4.

[0059] Furthermore, the second performance signal PF2 can be used to continue training the Bayesian neural network (BNN). Since the simulation results are usually quite accurate, the training state of the BNN can be improved in this way.

Claims

1. A computer-implemented method for controlling production equipment (PP) to manufacture product (P), comprising: a) Provide a machine learning module (BNN) trained to generate the following from design data records (DRs) of design variants of a specified product (P): - First performance signal (PF1), quantifying the predicted performance of the design variant, and -The prediction uncertainty (UC) of the prediction performance. b) Generate various Design Data Records (DRs), each DR specifying a design variant of the product (P); c) For the corresponding Design Data Record (DR) - The machine learning module (BNN) generates a first performance signal (PF1) and the corresponding prediction uncertainty (UC). - Based on the predicted uncertainty (UC), run or skip the simulation (SIM) of the design variant, the simulation (SIM) generating a second performance signal (PF2), the second performance signal (PF2) quantifying the simulated performance of the design variant, and If the simulation (SIM) is run, the performance value (PV) is derived from the second performance signal (PF2); otherwise, the performance value (PV) is derived from the first performance signal (PF1). d) Based on the derived performance value (PV), determine the performance-optimized design data record (ODR) from the various design data records (DRs), and e) Output the performance optimization design data (ODR) record for controlling the production equipment (PP).

2. The method according to claim 1, wherein The machine learning module (BNN) includes or implements surrogate models, Bayesian machine learning models, Bayesian neural networks, and / or Gaussian process models.

3. The method according to claim 1 or 2, wherein The training of the machine learning module (BNN) continues by using the second performance signal (PF2) as training data.

4. The method according to claim 1 or 2, wherein The predictive and / or simulation performance of a design variant depends on the achievement of one or more design objectives and / or the compliance of the design variant with one or more design constraints.

5. The method according to claim 1 or 2, wherein The prediction uncertainty (UC) is represented by the probability distribution, discrete probability distribution, statistical variance, statistical standard deviation, confidence interval, and / or error interval.

6. The method according to claim 1 or 2, wherein To decide whether to run or skip the corresponding simulation (SIM), the corresponding prediction uncertainty (UC) is compared with a threshold, and if it exceeds the threshold, the corresponding simulation is run.

7. The method according to claim 1, wherein To determine whether to run or skip the corresponding simulation (SIM), the corresponding first performance signal (PF1) is taken into account.

8. The method according to claim 1, wherein To determine whether to run or skip the corresponding simulation (SIM), the deviation between the predicted performance quantized by the corresponding first performance signal (PF1) and the predicted performance quantized by the previously generated first performance signal is taken into account.

9. The method according to claim 7 or 8, wherein An upper performance threshold (UPF) is determined for the prediction performance quantized by the corresponding first performance signal (PF1) and the corresponding prediction uncertainty (UC). A lower bound performance threshold (LPF) is determined for the prediction performance quantized from the previously generated first performance signal and the corresponding prediction uncertainty. To determine whether to run or skip the corresponding simulation (SIM), the upper performance threshold (UPF) is compared with the lower performance threshold (LPF).

10. The method of claim 9, wherein The upper performance threshold (UPF) is ranked within one or more lower performance thresholds (LPF) determined for one or more previously generated first performance signals, and The ranking is taken into account when deciding whether to run or skip the corresponding simulation (SIM).

11. A system (DS) for controlling a production facility (PP) to manufacture a product (P), adapted to perform the method according to any one of claims 1-10.

12. A computer program product for controlling production equipment (PP) to manufacture product (P), adapted to perform the method according to any one of claims 1 to 10.

13. A non-transient computer-readable storage medium storing a computer program product according to claim 12.

Images