A Next-Generation Product Design Method Based on Nominal Digital Twins

Abstract

This invention proposes a next-generation product design method based on nominal digital twins. A nominal digital twin is defined here for the first time as a composite digital twin of multiple independent digital twins corresponding to multiple physical products in use. Notional digital twins can build a more accurate approximate model not only from the data and information of a single product in use, but also from the performance and behavior of multiple physical products in a wide range of application scenarios in the physical world. This model can establish a random forest meta-model based on Bayesian optimization. This better model can collaboratively explore the optimal design solutions for the next generation of product design and improvement through multi-objective optimization on the nominal digital twin, and better understand the current products in use through comprehensive information, thereby making the next generation of product design more adaptable.

Description

Technical Field

[0001] The present invention relates to a product design method, and in particular to a next-generation product design method based on nominal digital twins. Background Art

[0002] In today’s increasingly competitive market, the digitalization of manufacturing encourages the use of virtual product models, also known as digital mockups (DMUs) and digital twins (DTs).

[0003] In the early stages of product development, the defined views link between the product digital model and configuration management, providing accurate 3D design data for each product configuration or variant, including 3D geometry, product structure, and properties. During the product development phase, the design is the successor to the definition. The functional digital mockup (FDMU) is proposed as a vehicle for simulating hypothetical scenarios across various engineering domains and models and saving all simulation results to fully demonstrate the product's behavioral description, including geometry, behavior, and visualization of simulation results. It is important to note that during the design phase, the defined and designed views may not yet have corresponding 3D physical models (sometimes, 3D prototypes may be built for concept evaluation). Therefore, the defined and designed views cannot be considered digital twins, but rather digital models. After each individual component or assembly is manufactured from its designed view and all parts are assembled, its digital model is called the industrial digital mockup (iDMU), containing more detailed manufacturing information for each component and assembly. After a single product goes through sales, logistics transportation and installation, its digital model is defined as the current product c-Pn, and more detailed transportation, installation and location information is provided before it is put into use. This will become the basis for establishing a digital twin model of the corresponding product.

[0004] Therefore, the concept of a digital twin for a product lifecycle can be confusing. However, when we view product design and development from the perspective of a single product generation, the digital twins of many previous-generation products can be incorporated into the new concept of a digital mockup of the next-generation product (s-DMU), which consists of the previous-generation products (physical entities), their digital twins, and a defined and designed view of the current product under the next-generation design. Each digital twin can provide "real" operating conditions and performance data for the specific previous-generation product in use. How to integrate this big data and information from these digital twins into the next-generation product design is a new research problem that has not been previously addressed.

[0005] The main challenge in next-generation product design is finding the right approach to integrate the various feedback information from the later stages of the digital twin lifecycles of numerous existing products into the early digital models of next-generation product design. This allows for effective and efficient data and information-derived product design from the manufacturing, operation, and maintenance of each physical product digital twin. This requires synthesizing a nominal digital twin model from the digital twins of many individual products to support the updated design of next-generation products. In summary, existing research on product design based on digital twins is limited to:

[0006] (1) Digital twin theory emphasizes the description of the real state of the product. Although traditional product data management systems can record, share, and manage design drawings, models, and documents, they only establish a static, idealized product information model that may change with the actual state of each product. Deviations exist in dynamic instance data such as processing, assembly, and inspection. How to establish a next-generation design model based on digital twins to more accurately describe and manage the real manufacturing and operation data of each instance product and combine it with the designed ideal product information model are two urgent problems to be solved.

[0007] (2) The core issue is the transition from multiple digital twins of existing products to digital models of next-generation products. Data on current product manufacturing, operation, and maintenance needs to be dynamically fed back into the design of next-generation products to establish an information model that accurately reflects the manufacturing, operation, and maintenance status of the product. Summary of the Invention

[0008] In order to achieve the above-mentioned object of the invention, the present invention provides the following technical solutions:

[0009] A next-generation product design method based on nominal digital twins includes the following steps:

[0010] Step S1: constructing a nominal digital twin model; Step S2: constructing a reverse optimization design model based on the nominal digital twin;

[0011] Wherein, step S1 includes the following steps:

[0012] Step S11: Establish a product behavior model; Step S12: Establish a forward and reverse design combination method; Step S13: Establish a meta-model of multi-instance products; Step S14: Establish an original random forest model; Step S15: Construct a nominal digital twin meta-model based on the Bayesian optimization hyperparameter method;

[0013] Wherein, step S2 includes the following steps:

[0014] Step S21: Establishing a solvable multi-objective optimization mathematical model; Step S22: Establishing a multi-objective optimization mathematical problem model; Step S23: Generating and verifying next-generation product design parameters based on the inverse optimization design model of the nominal digital twin;

[0015] Wherein, step S11: establishing a product behavior model;

[0016] The DMU forward design model is established as shown in formula (1):

[0017] Y=f(X,Q) (1)

[0018] Wherein, the system response Y=(y1,y2,…,y l ) is the initial design index of the product; Q=(q1,q2,…,q m ) is the system state parameter; X=(x1,x2,...,x n ) is the system design variable;

[0019] Optimal value of system design variable X * Based on the input of Q, the expected output is Y expected The optimal value of the system design variable X * get;

[0020] Based on the forward design model, the product digital twin model captures various uncertainties related to the product in the physical world. Combined with the influence of the uncertainty ε of the manufacturing error, the product design model forms d physical product instances. The product behavior model is established as shown in formula (2):

[0021]

[0022] Among them, Y d is the actual output response of the product design indicator of the d-th product instance; f(X d , Q d ) is the initial system response of the product design indicator of the d-th product instance; ε(d) is the uncertainty impact value of the d-th product instance; the mapping relationship between X, Q, and Y is described by a meta-model based on machine learning;

[0023] Step S12: establishing a forward and reverse design combination method;

[0024] Step S12 includes:

[0025] Step S121: Mining product operation data; Step S122: Forming a design decision closed loop;

[0026] Wherein, step S121: mining product operation data;

[0027] Mining product operation data to generate actual system response Y related to product design variables X actual Corresponding actual system usage information, so that the actual system response corresponds to the product design;

[0028] Step S122: forming a design decision closed loop;

[0029] While following the forward design process, reverse design imposes a feedback loop on the forward design; using the actual system response Y actual and the optimal value of the system design variable X * The inverse relationship between the optimal system parameter setting Q * ; Q * |Y actual Bring it back to the forward design process to replace the original assumption, that is, Q * →Q assumed , forming a closed loop of design decision-making and forming an improved design specifically tailored to the personalized needs of individual users;

[0030] Step S13: establishing a meta-model of a multi-instance product;

[0031] The actual design parameters X when using multiple instance products d The actual design parameters of a single product are constantly changing. d Represented as a vector of multiple digital twins, the actual system parameters Q of a single product d It is also represented as a vector of multiple digital twins, and then based on the actual design parameter vector X of each instance d(actual) =(x d1 ,x d2 ,……,x dn ) and the actual system parameter vector Q d(actual) =(q d1 ,q d2 ,……,q dm ), the corresponding design outputs of the optimal design of each instance are obtained respectively, and the relational meta-model is established as shown in formula (3):

[0032]

[0033] Among them, DT d is the digital twin model of the d-th instance product; Y d (actual) is the actual behavior model output value of the d-th instance product; f(X d (actual), Q d (actual)) is the actual system response; ε(d) is the uncertainty impact value caused by the manufacturing error factor corresponding to the d-th instance product;

[0034] Step S14: Establishing the original random forest model;

[0035] Based on the established meta-model of multi-instance products, data sampling is first performed to obtain the training set for establishing each decision tree; then, a decision tree is constructed based on the CART node segmentation algorithm, and multiple decision trees are combined to form a random forest model; finally, the average of the predicted values of all decision trees is solved and used as the predicted value of the random forest model; for regression problems, the average of the results of k decision trees is calculated as the final result, which is expressed as formula (4):

[0036]

[0037] Where k is the number of decision trees in the random forest model; T i (x) is the result of the i-th decision tree in the random forest; R(x) is the average output of the results of k decision trees;

[0038] Step S15: constructing a nominal digital twin model based on the Bayesian optimization hyperparameter method;

[0039] Step S15 includes the following steps: Step S151: combining different parameters of the original random forest model; Step S152: optimizing the parameters in the original random forest using the Bayesian optimization hyperparameter method;

[0040] Step S151: combining different parameters of the original random forest model;

[0041] The random forest model in the Python SciKit-learn learning module is used to set parameters and continuously adjust the combination. The Gaussian process is used to implement the combination process of different parameters of the original random forest model. The linear combination of any finite number of samples is expressed as a joint Gaussian distribution, as shown in formula (5):

[0042] f(x)~gp(m(x),w(x,x′)) (5)

[0043] Where f(x) is the output of the joint Gaussian distribution; m(x) = E(f(x)) is the mathematical expectation of f(x); w(x, x′) is the covariance function of x; gp is a Gaussian process;

[0044] Input the data into the Gaussian model to obtain its mean and variance, and construct the Gaussian distribution of the function; by increasing the amount of data, the gap between the predicted distribution and the true distribution is narrowed;

[0045] Step S152: Optimizing the parameters in the original random forest using the Bayesian hyperparameter optimization method; Step S152 includes S1521-S1524;

[0046] S1521 is: within the range of the number of random forest hyperparameters, randomly generate initialization parameters, and input the initialization parameters into the Gaussian model, then input the test sample into the fitting model to obtain the model output, and then modify the model output to make the model closer to the true distribution of the function;

[0047] S1522 is: using an extraction function to extract parameter combination points that need to be evaluated in the next step from the modified Gaussian model;

[0048] S1523 is: When the parameter combination error meets the target accuracy requirement, the algorithm ends and exits, outputting the appropriate parameter combination and the model prediction error (x i ,f(x i ));

[0049] S1524 is: if f(x) does not meet the predetermined accuracy requirement, add (x i ,f(x i )) Make changes, and then repeat S1522 and S1523 until the predetermined accuracy requirements are met; use the mean square error MSE, mean absolute error MAE and determination coefficient R 2 Evaluate the correctness of the metamodel, as shown in Equations (6)-(8);

[0050]

[0051] in: The actual value of the test set minus the predicted value; m is the number of test samples.

[0052] Step S21 includes steps 211 to 213;

[0053] Wherein, step S211: determining the optimization target, optimization parameters and optimization constraints of each instance condition by analyzing the specific instance conditions;

[0054] Wherein, step S212: determining the importance weight of each instance condition for different actual research objects;

[0055] Among them, step S213: use genetic algorithm to perform single-objective optimization and calculate the ideal value of each single objective under different working conditions; based on the constructed nominal digital twin meta-model, a multi-objective optimization mathematical model of multi-instance fusion is established, as shown in formula (9):

[0056]

[0057] Among them, x is the design variable, and the lower limit of its parameter value is x u , the upper limit is x w ;max / minY1(X)-max / minY d(X) is the optimization problem for different instances of working conditions; y dn (x) is the value of each performance indicator in the specific example; y dn The constraint range is (a r ,a s ).

[0058] In step S22, the multi-instance multi-objective optimization model is transformed using the ideal point method and the multi-objective programming method to establish a mathematical model that can solve the multi-objective optimization; the performance index y under different instance working conditions is transformed using the ideal point method. 11 (x),y 21 (x),…y d1 (x) is transformed into a single-objective problem; y 11 * ,y 21 * ,…y d1 * represents the ideal optimal solution of a single objective obtained under the same constraint conditions; according to the above formula (9), the optimization mathematical solution of multiple instance conditions is transformed into a multi-objective optimization mathematical problem:

[0059]

[0060] Among them, y1(X), y2(X), ..., y n (X) are the output performance indicators; y d1 (X),y d2 (X),……,y dn (X) are the performance index values under different instance conditions; ω d is the weight coefficient corresponding to the importance of the objective function; y * d1 ,y * d2 ,……,y * dn It is the ideal optimal solution of a single objective obtained under the constraints of different working conditions.

[0061] In step S22, after the optimization mathematical solution of the multi-instance conditions is converted into a multi-objective optimization mathematical problem, the parameters of the optimization model are input into Python for solution, and then the improved NSGA-II genetic algorithm is used for multi-objective intelligent optimization.

[0062] Step S23 includes steps S231 to S233;

[0063] In step S231: input key design variables X k and the original hypothesis Q assumed ;

[0064] In step S232: In the expected space of the next generation product, based on the DMU forward design model, under the given original assumption Q assumed Under the optimal value of the system design variable X*, the expected next-generation product design output Y is obtained. expected ;

[0065] In step S233: based on the previously established multi-objective inverse optimization model of the nominal digital twin of the current product, the improved NSGA-II genetic algorithm is used to solve the design scheme of the next generation product;

[0066] In step S234: Develop simulation scenarios, perform simulation tests on prototypes, predict the performance of the example product in use, identify design defects, and improve the design solution.

[0067] Compared with the prior art, the present invention has the following beneficial effects:

[0068] (1) The existing technology has not yet solved the technical problem of communication, collaboration and co-evolution between a physical object and its digital model (DMU). The present invention introduces an extended nominal digital twin (NDT), which integrates the digital twin information of multiple instantiation products and supports closed-loop iterative design. It not only helps to fully understand the relationship between the digital model and the digital twin, but also supports the update of the digital model view of the next generation product definition and design driven by the digital twin, thereby shortening the production process of the next generation product and improving product quality.

[0069] (2) The inventors discovered that the next-generation product development model and method still have deficiencies in practice. The present invention proposes a new next-generation product development model based on nominal digital twins, which dynamically constructs a next-generation product design model based on machine learning using multiple digital twins. This forms a closed loop of design decisions to produce an improved design, providing better product family design requirements for grouped users or usage scenarios.

[0070] (3) The inventors discovered in practice that the current digital twin is only applied to a single physical product related to manufacturing, transportation, installation, working scenarios, performance and behavior. Based on the introduction of nominal digital twin (NDT), the present invention proposed nominal digital twin modeling, which establishes a model by integrating different models of the real world with uncertainty in multiple cyberspaces. When developing new products, the design and design knowledge reuse are based not only on the current product, but also on the previous generation of products, and on the same lineage.

[0071] (4) The inventors found in practice that current digital twins mainly use cyberspace to predict product behavior, and lack understanding of existing design solutions and next-generation designs. The new nominal digital twin model proposed in this invention can perform rapid parallel learning and search through an incremental development method using multiple DTs, which not only helps to fully understand existing design solutions, but also provides profound insights into next-generation designs.

[0072] (5) The inventors discovered in practice that the digital model has defects in forward design, and the conversion from multiple digital twins to the digital model of the next-generation product is a reverse design problem. The forward design and reverse design models need to interact through feedback iteration. On this basis, the present invention proposes a forward and reverse design combination method. This method analyzes digital twin data to find the characteristic patterns of instance products under different working conditions or environments, establishes and integrates the meta-modeling of multiple instance products into the meta-modeling of nominal digital twins, and provides support for the reverse optimization of the next-generation product development.

[0073] (6) The inventors discovered in practice that it is difficult to construct a nominal digital twin model. To address this problem, the present invention proposes a new method for constructing a nominal digital twin model using a random forest model. The random forest model has the characteristics of being easy to implement, low in computational cost, and good in scalability, and effectively solves the errors caused by data loss during the operation of the model. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1 Schematic diagram of a design framework for next-generation products driven by notional digital twins. DETAILED DESCRIPTION

[0075] To make the purpose, technical solutions and advantages of the embodiments of the present invention more clear, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them.

[0076] Therefore, the following detailed description of the embodiments of the present invention is not intended to limit the scope of the claimed invention, but merely represents some embodiments of the present invention. All other embodiments derived by persons of ordinary skill in the art based on the embodiments of the present invention without creative effort shall fall within the scope of protection of the present invention.

[0077] It should be noted that, in the absence of conflict, the embodiments of the present invention and the features and technical solutions therein may be combined with each other.

[0078] A next-generation product design method based on nominal digital twins includes the following steps:

[0079] Step S1: Construct a nominal digital twin model; Step S2: Construct a reverse optimization design model based on the nominal digital twin.

[0080] Preferably, step S1 includes the following steps:

[0081] Step S11: Establish a product behavior model; Step S12: Establish a forward and reverse design combination method; Step S13: Establish a meta-model of multi-instance products; Step S14: Establish an original random forest model; Step S15: Construct a nominal digital twin meta-model based on the Bayesian optimization hyperparameter method.

[0082] Preferably, step S11: establishing a product behavior model.

[0083] Establish a DMU forward design model:

[0084] Y=f(X,Q) (1)

[0085] Wherein, the system response Y=(y1,y2,…,y l ) is the initial design index of the product; Q=(q1,q2,…,q m ) is the system state parameter; X=(x1,x2,…,x n ) are system design variables.

[0086] Preferably, the optimal value of the system design variable X * Based on the input of Q, the expected output is Y expected The optimal value of the system design variable X * get.

[0087] Based on the forward design model, the product digital twin model captures various uncertainties associated with the product in the physical world. Combined with the uncertainty (ε) of factors such as manufacturing errors, the product design model forms d physical product instances. The product behavior model is established as follows:

[0088]

[0089] Among them, Y d is the actual output response of the product design indicator of the d-th product instance; f(X d , Q d ) is the initial system response of the product design indicator for the dth product instance; ε(d) is the uncertainty impact value of the dth product instance. The mapping relationship between X, Q, and Y is described by a machine learning-based metamodel (such as the random forest model described later).

[0090] Step S12: Establishing a forward and reverse design combination method.

[0091] Preferably, step S12 includes: step S121: mining product operation data; step S122: forming a design decision closed loop.

[0092] Preferably, step S121: mining product operation data;

[0093] Mining the actual system response (Y) generated by product operation data and product design variables (X) actual ) corresponding to the actual system usage information, so that the actual system response corresponds to the product design (ie X|Y actual ), while generating actual observations of the system parameters (i.e., Q actual );

[0094] Preferably, step S122: forming a design decision closed loop;

[0095] While following the forward design process, reverse design imposes a feedback loop on the forward design. actual and design parameter X * The inverse relationship between the optimal system parameter setting Q * (i.e. Q * |Y actual ). * |Y actual Bring it back into the forward design process to replace the original hypothesis (i.e. Q * →Q assumed ), forming a closed loop of design decision-making and forming an improved design specifically for the personalized needs of individual users.

[0096] Step S13: Establish a meta-model of multi-instance products.

[0097] The actual design parameters X when using multiple instance products d The actual design parameters X of a single product are expressed as vectors of multiple digital twins, and the actual system parameters Q of a single product are expressed as vectors of multiple digital twins. d It is also represented as a vector of multiple digital twins, and then based on the actual design parameter vector X of each instance d(actual) =(x d1 ,x d2 ,……,x dn ) and the actual system parameter vector Q d(actual) =(q d1 ,q d2 ,……,q dm ), the corresponding design outputs of the optimal design of each instance are obtained, and the relational meta-model is established as follows:

[0098]

[0099] Among them, DT dis the digital twin model of the d-th instance product; Y d (actual) is the actual behavior model output value of the d-th instance product; f(X d (actual), Q d (actual)) is the actual system response (i.e., the actual design indicator response of the product); ε(d) is the uncertainty impact value caused by the manufacturing error factor corresponding to the d-th instance product.

[0100] Step S14: Establish the original random forest model.

[0101] Based on the previously established meta-model of multiple instance products, data sampling is first performed to obtain a training set for building each decision tree. Then, a decision tree is constructed based on the CART node segmentation algorithm, and multiple decision trees are combined to form a random forest model. Finally, the average of the predicted values of all decision trees is solved and used as the predicted value of the random forest model. For regression problems, the average of the results of k decision trees is calculated as the final result, which is expressed as formula (4):

[0102]

[0103] Where k is the number of decision trees in the random forest model; T i (x) is the result of the i-th decision tree in the random forest; R(x) is the average output of the results of k decision trees.

[0104] Step S15: Construct a nominal digital twin meta-model based on the Bayesian optimization hyperparameter method, that is, use the random forest model to solve the meta-model construction problem of NDT.

[0105] Preferably, step S15 includes the following steps: step S151: combining different parameters of the original random forest model; step S152: optimizing various parameters in the original random forest using the Bayesian optimization hyperparameter method.

[0106] Preferably, step S151: combining different parameters of the original random forest model;

[0107] The random forest model in the Python SciKit-learn learning module is used to set parameters and continuously adjust the combination. The Gaussian process is used to implement the combination process of different parameters of the original random forest model. The linear combination of any finite number of samples is expressed as a joint Gaussian distribution, as shown in formula (5):

[0108] f(x)~gp(m(x),w(x,x′)) (5)

[0109] Wherein, f(x) is the output of the joint Gaussian distribution; m(x)=E(f(x)) is the mathematical expectation of f(x); w(x, x′) is the covariance function of x; and gp is a Gaussian process.

[0110] Input the data into the Gaussian model to obtain its mean and variance, and construct the Gaussian distribution of the function; by increasing the amount of data, the gap between the predicted distribution and the true distribution is narrowed.

[0111] Preferably, step S152: using the Bayesian optimization hyperparameter method to optimize various parameters in the original random forest; step S152 includes S1521-S1524.

[0112] Preferably, S1521 is: randomly generating initialization parameters within the range of the number of random forest hyperparameters, and inputting the initialization parameters into the Gaussian model, then inputting the test sample into the fitting model to obtain the model output, and then modifying the model output to make the model closer to the true distribution of the function;

[0113] Preferably, S1522 is: using an extraction function to extract parameter combination points to be evaluated in the next step from the modified Gaussian model. This step can make the modified Gaussian model approximate the true distribution of the objective function faster and more accurately than other candidate set combinations;

[0114] Preferably, S1523 is: when the parameter combination error meets the target requirement, the algorithm ends and exits, outputting the appropriate parameter combination and the prediction error of the model (x i ,f(x i ));

[0115] Preferably, S1524 is: if f(x) does not meet the predetermined accuracy requirement, then add (x i ,f(x i )) and repeat S1522 and S1522 until the predetermined accuracy requirement is met. The mean square error (MSE), mean absolute error (MAE) and coefficient of determination (R-squared, R 2 ) to evaluate the correctness of the meta-model, as shown in Equations (6)-(8).

[0116]

[0117] in: is the actual value of the test set minus the predicted value; m is the number of test samples;.

[0118] The optimal parameters are obtained through Bayesian optimization in step S152, thereby finding the optimal parameter combination of the random forest prediction model.

[0119] The inventors discovered a series of technical problems, including defects in the forward design of digital models, large amounts of data in the nominal digital twin model, modeling problems requiring a high degree of data fitting, and difficulty in describing the mapping relationship between parameters X, Q, and Y with uncertainty in the design process.

[0120] In response to the technical problems discovered above, in step S1, the inventor creatively proposed a method for constructing a nominal digital twin metamodel using a random forest model based on the Bayesian optimization hyperparameter method. Specifically, five steps were adopted: establishing a product behavior model, establishing a forward and reverse design combination method, establishing a metamodel of a multi-instance product, establishing a nominal digital twin original random forest model, and constructing a nominal digital twin metamodel based on the Bayesian optimization hyperparameter method. This method solves the problems of data loss in the construction of the nominal digital twin metamodel and its operation, as well as the difficulty in maintaining a high degree of fit when the data volume is large. It better describes the uncertainty distribution of parameters X, Q, and Y, while forming a design decision-making closed loop for user personalized needs.

[0121] Step S2: Construct a reverse optimization design model based on the nominal digital twin, that is, establish a multi-objective optimization comprehensive model from multi-instance fusion.

[0122] Preferably, step S2 includes the following steps: step S21: establishing a solvable multi-objective optimization mathematical model; step S22: establishing a multi-objective optimization mathematical problem model; step S23: generating and verifying the next generation product design parameters based on the inverse optimization design model of the nominal digital twin.

[0123] Preferably, step S21: establishing a solvable multi-objective optimization mathematical model. Step S21 includes steps 211 to 213. Preferably, step S211: determining the optimization goal, optimization parameters and optimization constraints of each instance condition by analyzing specific instance conditions.

[0124] Preferably, step S212: For different actual research objects, the importance weight of each instance condition is determined based on the professional judgment and experience of experts or by using methods such as CRITIC, independence weight, and information weight. The ideal point obtained by single-objective optimization is the result of single-objective optimization of each performance indicator in different instance conditions. When determining the correlation coefficient of multi-objective optimization, the correlation of the instance conditions and the importance of the dynamic performance indicators are taken into account, so it is necessary to determine the importance weight of each instance condition.

[0125] Step S213: Use the genetic algorithm to perform single-objective optimization and calculate the ideal value of each single objective under different working conditions. Based on the constructed nominal digital twin meta-model, a multi-objective optimization mathematical model with multi-instance fusion is established (this model is the reverse optimization design model based on the nominal digital twin and will be used in the reverse optimization process of the next generation product design), as shown in formula (9):

[0126]

[0127] Among them, x is the design variable, and the lower limit of its parameter value is x u , the upper limit is x w ;max / minY1(X)-max / minY d (X) is the optimization problem for different instances of working conditions; y dn (x) is the value of each performance indicator in a specific instance; each performance indicator in a specific instance has a constraint range, such as y dn The constraint range is (a r ,a s ).

[0128] Step S22: Establish a multi-objective optimization mathematical problem model.

[0129] The ideal point method and multi-objective programming method are used to transform the multi-instance multi-objective optimization model and establish a mathematical model that can solve multi-objective optimization. 11 (x),y 21 (x),…y d1 (x) is transformed into a single objective problem. 11 * ,y 21 * ,…y d1 * represents the ideal optimal solution of a single objective obtained under the same constraints. According to the above formula (9), the optimization mathematical solution of multiple instance conditions is transformed into a multi-objective optimization mathematical problem:

[0130]

[0131] Among them, y1(X), y2(X), ..., y n (X) are the output performance indicators; y d1 (X),y d2 (X),……,y dn (X) are the performance index values under different instance conditions; ω d is the weight coefficient corresponding to the importance of the objective function; y * d1 ,y *d2 ,……,y * dn It is the ideal optimal solution of a single objective obtained under the constraints of different working conditions.

[0132] Preferably, the parameters of the optimization model are first input into Python for solution, and then an optimization algorithm (such as the improved NSGA-II genetic algorithm) is used for multi-objective intelligent optimization. In this way, a non-inferior solution set of the optimization model is established, realizing reverse optimization of the next generation product design.

[0133] Step S23: Generate and verify next-generation product design parameters based on the reverse optimization design model of the nominal digital twin. Step S23 includes steps S231 to S234.

[0134] Preferably, in step S231: input key design variables X k and the original hypothesis Q assumed ;

[0135] Preferably, in step S232: in the expected space of the next generation product, based on the DMU forward design model, under the given original assumption Q assumed Under this condition, the optimal value of the system design variable X is obtained. * The corresponding expected next-generation product design output Y expected (i.e. X * |Y expected );

[0136] Preferably, in step S233: based on the previously established nominal digital twin multi-objective inverse optimization model of the current product, an intelligent optimization algorithm (such as an improved NSGA-II genetic algorithm) is selected to solve the design scheme of the next generation product (i.e., a set of all design parameters that meet multiple working conditions);

[0137] Preferably, in step S234: a simulation scenario is developed, a simulation test is performed on the prototype, and the performance of the example product in use is predicted, so as to quickly adjust and accurately identify design defects and effectively improve the design solution.

[0138] In practice, the inventors discovered the difficulty in optimizing design parameters in next-generation product design. To address this, they proposed, in step S2, a method for constructing an inverse optimization design model based on a nominal digital twin. Specifically, this method employs the step of establishing a comprehensive multi-objective optimization model through multi-instance fusion, solving the problem of determining optimal design parameters for products under different operating conditions in next-generation product design.

[0139] The above embodiments are only used to illustrate the present invention and are not intended to limit the technical solutions described in the present invention. Although this specification has described the present invention in detail with reference to the above embodiments, the present invention is not limited to the above specific implementation methods. Therefore, any modification or equivalent replacement of the present invention; and all technical solutions and improvements thereof that do not depart from the spirit and scope of the invention are included in the scope of the claims of the present invention.

Claims

1. A next-generation product design method based on nominal digital twins, characterized by: The following steps are involved: Step S1: constructing a nominal digital twin model; Step S2: constructing a reverse optimization design model based on the nominal digital twin; Wherein, step S1 includes the following steps: Step S11: Establish a product behavior model; Step S12: Establish a forward and reverse design combination method; Step S13: Establish a meta-model of multi-instance products; Step S14: Establish an original random forest model; Step S15: Construct a nominal digital twin meta-model based on the Bayesian optimization hyperparameter method; Wherein, step S2 includes the following steps: Step S21: Establishing a solvable multi-objective optimization mathematical model; Step S22: Establishing a multi-objective optimization mathematical problem model; Step S23: Generating and verifying next-generation product design parameters based on the inverse optimization design model of the nominal digital twin; Wherein, step S11: establishing a product behavior model; The DMU forward design model is established as shown in formula (1): Y=f(X,Q) (1) Wherein, the system response Y=(y1,y2,…,y l ) is the initial design index of the product; Q=(q1,q2,…,q m ) is the system state parameter; X=(x1,x2,...,x n ) is the system design variable; Optimal value of system design variable X * Based on the input of Q, the expected output is Y expected The optimal value of the system design variable X * get; Based on the forward design model, the product digital twin model captures various uncertainties related to the product in the physical world. Combined with the influence of the uncertainty ε of the manufacturing error, the product design model forms d physical product instances. The product behavior model is established as shown in formula (2): Among them, Y d is the actual output response of the product design indicator of the d-th product instance; f(X d , Q d ) is the initial system response of the product design indicator of the d-th product instance; ε(d) is the uncertainty impact value of the d-th product instance; the mapping relationship between X, Q, and Y is described by a meta-model based on machine learning; Step S12: establishing a forward and reverse design combination method; Step S12 includes: Step S121: Mining product operation data; Step S122: Forming a design decision closed loop; Wherein, step S121: mining product operation data; Mining product operation data to generate actual system response Y related to product design variables X actual Corresponding actual system usage information, so that the actual system response corresponds to the product design; Step S122: forming a design decision closed loop; While following the forward design process, reverse design imposes a feedback loop on the forward design; using the actual system response Y actual and the optimal value of the system design variable X * The inverse relationship between the optimal system parameter setting Q * ; Q * |Y actual Bring it back to the forward design process to replace the original assumption, that is, Q * →Q assumed , forming a closed loop of design decision-making and forming an improved design specifically tailored to the personalized needs of individual users; Step S13: establishing a meta-model of a multi-instance product; The actual design parameters X when using multiple instance products d The actual design parameters of a single product are constantly changing. d Represented as a vector of multiple digital twins, the actual system parameters Q of a single product d It is also represented as a vector of multiple digital twins, and then based on the actual design parameter vector X of each instance d(actual) =(x d1 ,x d2 ,……,x dn ) and the actual system parameter vector Q d(actual) =(q d1 ,q d2 ,……,q dm ), the corresponding design outputs of the optimal design of each instance are obtained respectively, and the relational meta-model is established as shown in formula (3): Among them, DT d is the digital twin model of the d-th instance product; Y d (actual) is the actual behavior model output value of the d-th instance product; f(X d (actual), Q d (actual)) is the actual system response; ε(d) is the uncertainty impact value caused by the manufacturing error factor corresponding to the d-th instance product; Step S14: Establishing the original random forest model; Based on the established meta-model of multi-instance products, data sampling is first performed to obtain the training set for establishing each decision tree; then, a decision tree is constructed based on the CART node segmentation algorithm, and multiple decision trees are combined to form a random forest model; finally, the average of the predicted values of all decision trees is solved and used as the predicted value of the random forest model; for regression problems, the average of the results of k decision trees is calculated as the final result, which is expressed as formula (4): Where k is the number of decision trees in the random forest model; T i (x) is the result of the i-th decision tree in the random forest; R(x) is the average output of the results of k decision trees; Step S15: constructing a nominal digital twin model based on the Bayesian optimization hyperparameter method; Step S15 includes the following steps: Step S151: combining different parameters of the original random forest model; Step S152: optimizing the parameters in the original random forest using the Bayesian optimization hyperparameter method; Step S151: combining different parameters of the original random forest model; The random forest model in the Python SciKit-learn learning module is used to set parameters and continuously adjust the combination. The Gaussian process is used to implement the combination process of different parameters of the original random forest model. The linear combination of any finite number of samples is expressed as a joint Gaussian distribution, as shown in formula (5): f(x)~gp(m(x),w(x,x′))(5) Where f(x) is the output of the joint Gaussian distribution; m(x) = E(f(x)) is the mathematical expectation of f(x); w(x, x′) is the covariance function of x; gp is a Gaussian process; Input the data into the Gaussian model to obtain its mean and variance, and construct the Gaussian distribution of the function; by increasing the amount of data, the gap between the predicted distribution and the true distribution is narrowed; Step S152: Optimizing the parameters in the original random forest using the Bayesian hyperparameter optimization method; Step S152 includes S1521-S1524; S1521 is: within the range of the number of random forest hyperparameters, randomly generate initialization parameters, and input the initialization parameters into the Gaussian model, then input the test sample into the fitting model to obtain the model output, and then modify the model output to make the model closer to the true distribution of the function; S1522 is: using an extraction function to extract parameter combination points that need to be evaluated in the next step from the modified Gaussian model; S1523 is: When the parameter combination error meets the target accuracy requirement, the algorithm ends and exits, outputting the appropriate parameter combination and the model prediction error (x i ,f(x i )); S1524 is: if f(x) does not meet the predetermined accuracy requirement, add (x i ,f(x i )) Make changes, and then repeat S1522 and S1523 until the predetermined accuracy requirements are met; use the mean square error MSE, mean absolute error MAE and determination coefficient R 2 Evaluate the correctness of the metamodel, as shown in Equations (6)-(8); in: The actual value of the test set minus the predicted value; m is the number of test samples.

2. The next-generation product design method based on nominal digital twins according to claim 1, characterized in that: Step S21 includes steps 211 to 213; Wherein, step S211: determining the optimization target, optimization parameters and optimization constraints of each instance condition by analyzing the specific instance conditions; Wherein, step S212: determining the importance weight of each instance condition for different actual research objects; Among them, step S213: use genetic algorithm to perform single-objective optimization and calculate the ideal value of each single objective under different working conditions; based on the constructed nominal digital twin meta-model, a multi-objective optimization mathematical model of multi-instance fusion is established, as shown in formula (9): Among them, x is the design variable, and the lower limit of its parameter value is x u , the upper limit is x w ;max / minY1(X)-max / minY d (X) is the optimization problem for different instances of working conditions; y dn (x) is the value of each performance indicator in the specific example; y dn The constraint range is (a r ,a s ).

3. The next-generation product design method based on nominal digital twins according to claim 2, characterized in that: In step S22, the multi-instance multi-objective optimization model is transformed using the ideal point method and the multi-objective programming method to establish a mathematical model that can solve the multi-objective optimization; the performance index y under different instance working conditions is transformed using the ideal point method. 11 (x),y 21 (x),…y d1 (x) is transformed into a single-objective problem; y 11 * ,y 21 * ,…y d1 * represents the ideal optimal solution of a single objective obtained under the same constraint conditions; according to the above formula (9), the optimization mathematical solution of multiple instance conditions is transformed into a multi-objective optimization mathematical problem: Among them, y1(X), y2(X), ..., y n (X) are the output performance indicators; y d1 (X),y d2 (X),……,y dn (X) are the performance index values under different instance conditions; ω d is the weight coefficient corresponding to the importance of the objective function; y * d1 ,y * d2 ,……,y * dn It is the ideal optimal solution of a single objective obtained under the constraints of different working conditions.

4. The next-generation product design method based on nominal digital twins according to claim 3, characterized in that: In step S22, after the optimization mathematical solution of the multi-instance conditions is converted into a multi-objective optimization mathematical problem, the parameters of the optimization model are input into Python for solution, and then the improved NSGA-II genetic algorithm is used for multi-objective intelligent optimization.

5. The next-generation product design method based on nominal digital twins according to claim 4, characterized in that: Step S23 includes steps S231 to S233; In step S231: input key design variables X k and the original hypothesis Q assumed ; In step S232: In the expected space of the next generation product, based on the DMU forward design model, under the given original assumption Q assumed Under the optimal value of the system design variable X*, the expected next-generation product design output Y is obtained. expected ; In step S233: based on the previously established multi-objective inverse optimization model of the nominal digital twin of the current product, the improved NSGA-II genetic algorithm is used to solve the design scheme of the next generation product; In step S234: Develop simulation scenarios, perform simulation tests on prototypes, predict the performance of the example product in use, identify design defects, and improve the design solution.

Images