Manufacturing or controlling a technical system using an op timized parameter set

The present invention relates to a computer-implemented meth od, an apparatus and a computer program product for generat ing an optimized parameter set for manufacturing or control ling a technical system.

Before manufacturing or controlling a product or a technical system determining suitable production, control and/or con struction parameters are required. During e.g. a design phase a computerized optimization method is often used to optimize parameter settings for manufacturing the product or control ling the technical system. Such an optimization procedure is typically run in iterations and can therefore be time and/or resource consuming, as the parameter values obtained from the optimization need to be evaluated. The evaluation is usually performed by using a computerized simulation of the product or technical system, e.g. a finite-element-analysis (FEA) or a computational-fluid dynamics (CFD) simulation. However, a typical optimizer usually has insufficient information about such a computerized model, such that the optimizer can gener ate parameters which cannot be evaluated and/or do not satis fy system or product constraints, resulting in useless param eter sets.

EP 3246 769 A1 describes a controller for simulation and op timization of operations of a power plant, wherein a dynamic model of the power plant based on geometrical and operation data is created, and a surrogate model for a specific perfor mance metric based on the dynamic model is generated, which is incorporated into an optimization procedure to optimize operations of the power plant for the specific performance metric. It is therefore an objective of the present invention to im prove a generation of an optimized parameter set for manufac turing or controlling a product or technical system.

The object of the invention is achieved by the features of the independent claims. The dependent claims contain further developments of the invention.

The invention provides according to the first aspect a com puter-implemented method for manufacturing or controlling a technical system, comprising the method steps:

(a) inputting a sample of parameter sets suitable for manu facturing or controlling the technical system, with a feasi bility identifier assigned to each of the parameter sets, wherein the feasibility identifier marks each parameter set either as technically feasible, if it meets given system cri teria of the technical system, or technically non-feasible, if does not meet the given system criteria, or erroneous, if an evaluation of the parameter set using a computerized simu lation results in an error, in terms of manufacturing or con trolling the technical system,

(b) generating a computerized surrogate model for the tech nical system based on the respective parameter sets of the sample, which are marked as technically feasible but not on erroneous and not on non-feasible parameter sets, by means of a regression method,

(c) determining an optimized parameter set based on the sur rogate model by means of a computerized optimization method, and

(d) outputting the optimized parameter set for manufacturing or controlling the technical system.

If not indicated differently the terms "calculate", "per form", "computer-implemented", "compute", "determine", "gen erate", "configure", "reconstruct", and the like, preferably are related to acts and/or processes and/or steps which change and/or generate data, wherein data can particularly be presented as physical data, and which can be performed by a computer or processor. The term "computer" can be interpreted broadly and can be a personal computer, server, mobile compu ting device, or a processor such as a central processing unit (CPU) or microprocessor.

An important advantage of the present invention is the usage of a surrogate model representing and/or approximating at least part of the technical system. The surrogate model is preferably a computerized model. Preferably, the surrogate model is less complex than a detailed computerized simulation model of the technical system, such as a CFD simulation mod el. The surrogate model can for example be stored in a data structure. It is generated by means of a regression method and is generated depending on the input sample of parameter sets for manufacturing or controlling the technical system. The surrogate model is generated based on at least one tech nically feasible parameter set and corresponding evaluation results of this parameter set.

A parameter set comprises parameter values defining at least a property or a functionality of the technical system. A pa rameter set can be understood as a data set. A parameter set can for example comprise construction, production, control, operation and/or design parameters. A construction parameter set comprises for example parameters which can specify a size, a weight, a material, etc. of at least a component of the technical system. Furthermore, a construction or control parameter set can comprise system- or product-specific set ting parameters for a manufacturing unit or machine. A con trol parameter set can for example comprise system- specific control parameters for setting a control unit of the tech nical system for controlling the technical system.

A parameter set can for example serve as input for a manufac turing unit wherein the manufacturing unit is configured to manufacture the technical system depending on the respective construction parameter set. The sample of parameter sets is hence a plurality of parameter sets wherein each parameter set is suitable for manufacturing the technical system. Each parameter set is labeled with a feasibility identifier allow ing to identify a more suitable parameter set.

An optimized parameter set is determined based on the surro gate model, accelerating the optimization process. As the surrogate model is generated based preferably on feasible, but not on erroneous and preferably also not on non-feasible, parameter sets. The optimization is then run on a pre constrained model. In other words, the optimization procedure for determining the optimized parameter set takes information about e.g. system constraints of the technical system into account. Therefore, a more suitable optimized parameter set for manufacturing or controlling the technical system can be generated. The parameter optimization is therefore accelerat ed compared to standard methods.

According to an embodiment of the computer-implemented meth od, a feasibility identifier of the optimized parameter set can be determined by means of a computerized evaluation meth od.

The computerized evaluation method can for example be trained based on a machine learning method, such that it reproduces a feasibility identifier distribution of a training data set, wherein the training data set comprises parameter sets which are labeled with respective feasibility identifiers. The com puterized evaluation method is for example an artificial neu ral network or the like. Using such a computerized evaluation method, information embedded in the data, i.e. in a parameter set, can be used to identify whether the parameter set is feasible, non-feasible or erroneous.

For example, the outputted optimized parameter set can be evaluated by means of such a trained computerized evaluation method to determine whether it is technically feasible or technically non-feasible or erroneous in terms of manufactur ing or controlling the technical system. According to a further embodiment of the computer-implemented method, the optimized parameter set can be determined by means of the computerized optimization method and depending on its feasibility identifier.

The value of the respective feasibility identifier, i.e. technically feasible or technically non-feasible or errone ous, of an outputted optimized parameter set can further be evaluated by the optimization method as an additional optimi zation constraint. With such an additional information, the optimization method allows to generate a parameter set avoid ing particularly erroneous parameter combinations and/or un necessary computational iterations.

According to a further embodiment of the computer-implemented method, the optimized parameter set can be added to the sam ple of parameter sets and the method steps (b) to (d) of claim 1 are repeated.

By adding the outputted optimized parameter set to the sample of parameter sets an optimized surrogate model can be gener ated, further improving the generation of an optimized param eter set.

According to a further embodiment of the computer-implemented method, the inputted sample of parameter sets can be generat ed and outputted by means of a second computerized optimiza tion method which is configured to determine at least one pa rameter set suitable for manufacturing or controlling the technical system.

The input sample of parameter sets is preferably generated by means of a second computerized optimization method and evalu ated using a computerized simulation model of the technical system in order to determine the technical feasibility of each parameter set. The second computerized optimization method can for example be the same as the aforementioned op timization method.

According to a further embodiment of the computer-implemented method the regression method can be an incremental local Gaussian regression method.

For generating the surrogate model an incremental local gaussian regression algorithm (ILGR) can be adopted to train the model. This algorithm combines the advantage of locally weighted regression (LWR) and Gaussian Process Regression

(GPR). It uses localizing function basis and approximates in ference techniques to build a Gaussian process regression al gorithm of increasingly local nature and similar computation al complexity to LWR. It allows for example streaming da tasets of unknown size, such that it is especially suitable to generate a surrogate model during an iterative optimiza tion process. The model generation can be done incrementally while new parameter sets are available.

According to a further embodiment of the computer-implemented method, the technical system can be manufactured or con trolled using the outputted optimized parameter set.

The outputted optimized parameter set can for example be transferred to a manufacturing unit or machine for manufac turing the technical system according to this construction parameter set. Alternatively, the outputted optimized parame ter set can be transferred to a control unit for controlling the technical system.

The invention provides according to the second aspect an ap paratus for manufacturing or controlling a technical system, comprising :

(a) an input unit configured to input a sample of parameter sets suitable for manufacturing or controlling the technical system, with a feasibility identifier assigned to each of the parameter sets, wherein the feasibility identifier marks each parameter set either as technically feasible, if it meets given system criteria of the technical system, or technical ly non-feasible, if does not meet the given system criteria, or erroneous if an evaluation of the parameter set using a computerized simulation results in an error, in terms manu facturing or controlling the technical system,

(b) a generator configured to generate a computerized surro gate model for the technical system based on the parameter sets, which are marked as technically feasible but not on er roneous and not on non-feasible parameter sets, by means of a regression method,

(c) an optimizer configured to determine an optimized parame ter set based on the generated surrogate model by means of a computerized optimization method, and

(d) an output unit configured to output the optimized parame ter set for manufacturing or controlling the technical sys tem.

The apparatus and/or at least one of its units can further comprise at least one processor or computer to perform the method steps according to the invention. A respective unit may be implemented in hardware and/or in software. If said unit is implemented in hardware, it may be embodied as a de vice, e.g. as a computer or as a processor or as a part of a system. If said unit is implemented in software it may be em bodied as a computer program product, as a function, as a routine, as a program code or as an executable object. The output unit preferably provides a data structure comprising the optimized parameter set. Such data structure can for ex ample be transferred to a manufacturing unit or a control unit.

The apparatus can further comprise a manufacturing unit con figured to manufacture the technical system using the output ted optimized parameter set. Alternatively, the apparatus can be connected to a manufacturing unit, e.g. via a network. The apparatus can further comprise a control unit configured to control the technical system using the outputted optimized parameter set. Alternatively, the apparatus can be connected to a control unit, e.g. via a network.

The invention further comprises a computer program product directly loadable into the internal memory of a digital com puter, comprising software code portions for performing the steps of the said method when said product is run on a com puter.

A computer program product, such as a computer program means, may be embodied as a memory card, USB stick, CD-ROM, DVD or as a file which may be downloaded from a server in a network.

The invention will be explained in more detail by reference to the accompanying figures.

Fig. 1 shows a flow chart including method steps involved in an embodiment of the method for manufacturing or controlling a technical system;

Fig. 2 shows a schematic representation of an embodiment of the method for manufacturing or controlling a technical system; and

Fig. 3 shows a schematic representation of an embodiment of an apparatus for manufacturing or controlling a technical system.

Equivalent parts in the different figures are labeled with the same reference signs.

Figure 1 shows a flow chart illustrating the method steps of a computer-implemented method according to the invention for manufacturing or controlling a technical system using an op timized parameter set. The technical system can for example be a machine, like a turbine or motor, a hardware component of a machine, a mechanical part, a product, or similar. Fur thermore, the technical system can be a complex system com prising multiple individual parts, e.g., a transportation network for water, oil, packages or the like.

In the first step SI a sample of parameter sets suitable for manufacturing or controlling the technical system is in putted. Depending of the respective application, i.e. for ex ample either manufacturing or controlling the technical sys tem, the parameter sets comprise construction or production parameters or control parameters.

Each parameter set is labeled with a feasibility identifier marking it either as technically feasible or technically non- feasible or erroneous in terms of manufacturing or control ling the technical system. The respective feasibility identi fier can for example be determined by means of a classifica tion or evaluation method and/or by means of an analysis method evaluating the respective parameter set e.g. using a computerized simulation of the technical system.

A parameter set is for example labeled as technically feasi ble if it meets given system criteria of the technical sys tem. A parameter set is for example labeled as technically non-feasible if it does not meet the given system criteria. A parameter set is for example labeled as erroneous if for ex ample an evaluation of the parameter set using a computerized simulation results in an error. Hence, whether a parameter set is considered as technically feasible is determined by the formed optimization problem. If the parameter set vio lates one of the constraints, it is not feasible. If the pa rameter set fulfills all constraints, it is marked as feasi ble. The differentiation between feasible or infeasible is the error case. While the former is determined by the optimi zation already, the latter is a result of faulty evaluations, e.g. error from the evaluating software tool. In the next step S2, a computerized surrogate model for the technical system is generated by means of a regression meth od. Preferably an incremental local Gaussian regression meth od is used, which is adopted to train the surrogate model. A respective parameter set preferably defines a functionality or property of the technical system, such as size, mass, width, height, etc.. The computerized surrogate model is con figured to represent and/or approximate the technical system. Therefore, the surrogate model can be generated based on the input parameter sets which are technically feasible and/or technically non-feasible. Preferably, only the technically feasible parameter sets are used for the surrogate model gen eration. In other words, for the surrogate model generation only part of the sample of parameters, preferably only the technically feasible, is used.

In the next step S3, a computerized optimization method is used to determine an optimized parameter set based on the generated computerized surrogate model. A black-box-optimizer can be used. As only evaluation for the objective and con straints for the variable values, i.e. the respective parame ter set, are available, a black-box optimizer can be used to propose variable values and to evaluate these.

The optimized parameter set is outputted, step S4, for manu facturing or controlling the technical system. The optimized construction parameter set can for example serve as an input for a manufacturing machine for manufacturing the technical system or as input for a production unit for producing a product.

Alternatively, a feasibility identifier of the optimized pa rameter set is determined by means of a computerized evalua tion method, step S6. The computerized evaluation method has been trained to determine whether an inputted parameter set is technically feasible, technically non-feasible or errone ous in terms of manufacturing or controlling the technical system. Therefore, the computerized evaluation method is preferably trained or configured to represent the feasibility identifier distribution of the initial input sample of param eter sets. The value of the feasibility identifier of the op timized parameter set can further be used during optimiza tion, step S3, as a further optimization constraint. There fore, the optimization method can be adopted to additionally evaluate the respective feasibility identifier when determin ing an optimized parameter set. Preferably, an optimized pa rameter set which is evaluated as erroneous is sorted out.

Furthermore, the outputted optimized parameter set, which is preferably evaluated as technically feasible, can be added to the initial input sample of parameter sets and the method steps SI to S4 can be repeated, see arrow REP. Such an itera tion provides an improved optimized parameter set, which can be outputted for manufacturing or controlling the technical system.

Figure 2 shows a schematic representation of an embodiment of the invention. The computer-implemented method for manufac turing or controlling a technical system using an optimized parameter set can comprise two method parts 10, 20.

The initial input data set, i.e., the sample of parameter sets PS1, can for example be provided by an optimization method OPT2. Such optimization method OPT2 can for example generate a sample of parameter sets PS1 for manufacturing the technical system. A feasibility identifier is assigned to each of the sample of parameter sets PS1, marking each param eter set either as technically feasible, technically non- feasible or erroneous. The respective feasibility identifier can for example be determined using a simulation-based evalu ation. In other words, the parameter sets PS1 can be under stood as labelled data sets.

This sample of parameter sets PS1 can for example be used as a training data set in order to train a computerized evalua tion method NN, e.g. an artificial neural network. The com- puterized evaluation method NN is trained such that it repro duces the feasibility identifier distribution of the training data set. The method part 10 of the illustrated method repre sents this data classification.

The labeled data set can for example be divided into the training data set and a test data set, such that the comput erized evaluation method NN can be trained with the training data set and verified with the test data set. Preferably, all parameter sets are used, i.e. the technically feasible, tech nically non-feasible or erroneous parameter sets. The trained evaluation method NN can then be used to determine a feasi bility identifier FI of an outputted parameter set.

Furthermore, the result of the computerized evaluation method NN can be used as an additional constraint for the optimiza tion method, step S3, to determine an optimized parameter set. Such constraint can provide feedback to the optimization method and guides the optimization problem to generate an op timized parameter set according to the feasibility identifier FI, i.e. preferably a parameter set without an error.

The method part 20 comprises the surrogate model SM genera tion for objective and constraints and the parameter optimi zation based on the surrogate model. Using preferably only part of the initial sample of parameter sets PS2, comprising only the parameter sets classified as technically feasible, a surrogate model SM for the technical system is generated by means of a regression method, step S2.

The generation of the surrogate model SM preferably comprises the following steps. The first major step is the training process. Here an incremental local gaussian regression algo rithm (ILGR) is adopted to train the surrogate model SM com bining the advantage of locally weighted regression (LWR) and Gaussian Process Regression (GPR). The second step is the cross validation of the generated surrogate model. The gener ated surrogate model SM is validated by a validation set, e.g. one part of the partial sample of parameter sets PS2.

The root-mean-square-error (RMSE) and R-squared values are evaluated with the validation set and the number of sub models of the ILGR - also called local models - are inputs of the third step: hyperparameter optimization. These three terms form an objective function and an optimization problem is solved by a black box optimizer, such as BOBYQA from the NLopt library. Then the surrogate model generation process is restarted using the updated hyperparameters, until a stop criterion is satisfied. Therefore, the hyperparameters for ILGR are determined by minimizing the RMSE and the number of sub-models, and maximizing R ^A 2.

Part 10 and part 20 can be two independent approaches to ac celerate optimization. They can be applied together or sepa rately.

Based on the surrogate model SM, an optimized parameter set PS_opt can be determined by means of an optimization method, step S3. By means of the optimization method, parameters of the surrogate model SM are modified in order to find an opti mum. In case of manufacturing, the optimization method can for example determine an optimized parameter set for a prod uct, yielding an optimized size, weight or other physical property of the product. In other words, the optimization method can determine optimized parameters of the surrogate model with respect to a given optimization goal.

If method part 10 is applied, a parameter set can be evaluat ed based on the computerized evaluation method NN resulting in a feasibility identifier FI. If only part 20 is applied, the optimization method in step S3 will evaluate the surro gate model. If both parts 10, 20 are applied, the optimiza tion method in step S3 includes the additional constraint de pending on the feasibility identifier FI and evaluates the generated surrogate model SM. Taking this additional con straint into account, an optimized parameter set PS_opt is determined. Figure 3 shows a schematic representation of an embodiment of an apparatus for manufacturing or controlling a technical system. The apparatus 100 for manufacturing or controlling a technical system comprises an input unit 101, a generator 102, an optimizer 103 and an output unit 104. The apparatus and/or at least one of its units can further comprise at least one processor or computer to perform the method steps according to the invention.

The input unit 101 is configured to input a sample of parame ter sets suitable for manufacturing or controlling the tech nical system, with a feasibility identifier assigned to each of the parameter sets, wherein the feasibility identifier marks each parameter set either as technically feasible or technically non-feasible or erroneous in terms manufacturing or controlling the technical system.

The generator 102 is configured to generate a computerized surrogate model for the technical system based on the parame ter sets, which are preferably marked as technically feasi ble, by means of a regression method.

The optimizer 103 is configured to determine an optimized pa rameter set based on the generated surrogate model by means of a computerized optimization method.

The output unit 104 is configured to output the optimized pa rameter set for manufacturing or controlling the technical system.

The apparatus preferably comprises an evaluation unit 105 which is configured to determine a feasibility identifier of an inputted parameter set by means of a computerized evalua tion method, such as an artificial neural network or a random forest algorithm. The apparatus further preferably further comprises a manufac turing unit (not shown) or is coupled with a manufacturing unit, wherein the manufacturing unit is configured to manu facture the technical system using the outputted optimized parameter set. In addition or alternatively the apparatus can comprise a control unit (not shown) or is coupled with a con trol unit, wherein the control unit is configured to control the technical system using the outputted optimized parameter set.

Although the present invention has been described in detail with reference to the advantageous embodiment, it is to be understood that the present invention is not limited by the disclosed examples, and that numerous additional modifica- tions and variations could be made thereto by a person skilled in the art without departing from the scope of the invention.