Part digital twin surface model rapid characterization method based on virtual-real fusion

Abstract

The invention discloses a part digital twin surface model rapid characterization method based on virtual-real fusion, and the method comprises the steps: inputting a part design CAD model and a manufacturing solid model, and obtaining the surface feature information of the CAD model and the high-precision point cloud data of the corresponding surface of the solid model; a mechanism modeling method and a data analysis method are adopted to obtain a standard skin shape model and an authenticated skin shape model respectively; after the two skin shape models are preprocessed, a training sample set is formed through virtual-real fusion, and a deep learning model based on a generative adversarial network is constructed and used for generating a surface three-dimensional deviation form unit model meeting functional requirements; and establishing correlation mapping between the CAD ideal surface model of the part and the three-dimensional deviation form of the CAD ideal surface model to obtain a training skin surface shape model, and taking the training skin surface shape model as the representation of the digital twin surface model of the part. According to the method, the precision and efficiency of part surface feature skin surface model expression are improved, and the method can be suitable for assembly precision analysis and prediction simulation of digital twin-driven precision mechanical and electrical products.

Description

Technical Field

[0001] The present invention belongs to the technical field related to digital design and simulation, and specifically relates to a method for quickly characterizing a digital twin surface model of a part based on virtual-real fusion. Background Art

[0002] In the context of improving precision as one of the eternal goals of mechanical manufacturing, as complex precision electromechanical products continue to pursue higher precision, the accuracy requirements of assembly precision analysis and predictive simulation are also constantly increasing. We know that the connection interface of the part assembly surface is an important factor that cannot be ignored in the assembly precision analysis and predictive simulation of complex precision electromechanical products. The modeling and characterization methods of the part surface model will directly determine the results of the product assembly precision analysis and predictive simulation. Since the actual machined part surface is inevitably different from the ideal geometric surface, the traditional assembly precision analysis based on the ideal geometric surface can no longer accurately reflect the impact of the part surface morphology characteristics on high-precision assembly. Therefore, more and more scholars have begun to pay attention to the error modeling and characterization methods of the non-ideal surface model of parts.

[0003] The non-ideal surface model of parts has a unified concept under the new generation of product geometric technical specification standard system - Skin Model (literally translated as "skin surface model"). In recent years, researchers at home and abroad have carried out in-depth research and practice on the modeling technology of the skin surface model of parts. As early as 2013, some scholars proposed Skin ModelShapes (literally translated as "skin surface shape model") based on discrete geometry for the unified expression of the non-ideal surface model of parts, and realized the error modeling and visualization expression of the non-ideal surface model by superimposing geometric defect errors (including systematic errors and random errors) on the basis of the theoretical model. The initial skin surface shape model error modeling accuracy depends entirely on the modeling capability of the parametric CAD system and the mathematical model of the given systematic error and random error. For example, the second-order shape, modal basis function shape, Bezier curve, non-uniform rational B-spline (NURBS), parameter space envelope, etc. are used to simulate the systematic error, and the random error is obtained by sampling methods (such as one-dimensional Gaussian, multi-dimensional Gaussian, Gaussian / non-Gaussian random field, etc.). Although the above methods can realize the digital modeling and expression of the skin surface shape model to a certain extent, they ignore the actual part processing and manufacturing information and the micro-geometric defect errors of the surface morphology. In order to improve the high fidelity of the skin surface shape model, many scholars have further modeled and characterized the non-ideal rough surface of the parts on the basis of considering the morphology of part processing and manufacturing errors, simulated the three-dimensional surface micro-morphology with finite statistical parameters, and obtained the real surface micro-morphology of the parts through surface morphology measuring instruments, which provides strong technical support for the refined characterization of the skin surface shape model of the parts and the high-quality surface modeling.

[0004] However, most of the current modeling methods for non-ideal surface models of parts use Euclidean or fractional dimension geometry theory, as well as methods based on statistical parameter simulation, rough surface measurement reconstruction, etc. However, they are limited by the lack of knowledge-based setting of manufacturing error morphological parameters and the inability to fully utilize a large amount of rich virtual simulation and actual measurement data. As a result, there is a contradiction between the efficiency of part surface modeling and refined characterization in the above methods, resulting in the inability to quickly and accurately realize the construction and expression of non-ideal surface features of parts, and it is difficult to meet the computational requirements of high-fidelity analysis of assembly accuracy of complex precision electromechanical products and the assembly rhythm requirements in actual production. Summary of the invention

[0005] To solve the above problems, the present invention introduces digital twin technology in the modeling of non-ideal surface models of parts, and discloses a method for rapid characterization of digital twin surface models of parts based on virtual-reality fusion. By deeply fusing the virtual simulation of part surface features with actual measurement data, a deep learning model based on a generative adversarial network is used to generate digital twin surface models of parts, thereby achieving rapid, accurate, and high-fidelity digital characterization of product parts, which can be used for high-precision assembly analysis and simulation of complex and precision electromechanical products.

[0006] To achieve the above object, the technical solution of the present invention is as follows:

[0007] A method for quickly characterizing a surface model of a digital twin of a part based on virtual-reality fusion drive includes the following steps:

[0008] Step (1): Input the part design CAD model and the manufacturing solid model, obtain the surface feature information of the CAD model and the high-precision point cloud data of the corresponding surface of the solid model, and use mechanism modeling and data analysis methods to obtain the standard and certified surface shape models respectively;

[0009] Step (2): After preprocessing the above two skin surface shape models, a training sample set is formed by virtual-real fusion, and a deep learning model based on a generative adversarial network is constructed to generate a surface three-dimensional deviation morphological unit model that meets the functional requirements;

[0010] Step (3): Establish a correlation mapping between the ideal CAD surface model of the part and its three-dimensional deviation shape, obtain a training surface shape model, and use it as a representation of the digital twin surface model of the part.

[0011] As one of the preferred embodiments of the present invention, in step (1), the part surface refers to a critical plane and / or critical cylindrical surface having geometric dimension tolerance design requirements and structural function requirements.

[0012] Furthermore, the specific implementation method of step (1) is as follows:

[0013] (11) For the part design CAD model, firstly, the geometric dimension tolerance data of the part surface is obtained through the model-based definition method of the computer-aided design software; secondly, the nominal surface of the part design CAD model is extracted and meshed by adopting feature segmentation and meshing algorithms or using computer-aided design and finite element analysis software tools; then, the position, direction and shape errors formed during the part surface processing are simulated by mechanism modeling, combining the surface characteristics of the part and the corresponding geometric dimension tolerance data. Among them, the position and direction errors are modeled by the small displacement spinor method, and the position and direction errors are randomly generated under the constraints of the small displacement spinor parameters, while the shape error is generated by the systematic error and random error of the processing process. The mechanism error is modeled, and the basis function method and the non-Gaussian random field method are used to construct the system error and the random error respectively. The shape error is generated after the two are superimposed, and it is detected whether it exceeds the upper and lower limits of the specified shape tolerance band. If it does not exceed the limit, the subsequent operation is continued, otherwise the scaling operation is performed to make it meet the shape tolerance design requirements before continuing the subsequent operation; finally, the above position and direction errors and shape errors are superimposed to detect whether they exceed the upper and lower limits of the specified position tolerance band. If it does not exceed the limit, the generated skin surface shape model is valid, otherwise the operation of generating position and direction errors and shape errors by mechanism modeling is repeated until an effective skin surface shape model is generated, and this is used as the standard skin surface shape model that meets the geometric dimension tolerance design requirements;

[0014] (12) For the manufacturing of solid models of parts, firstly, high-precision point cloud data of the corresponding surface of the parts is obtained by using high-precision measuring equipment such as a three-dimensional coordinate measuring machine, a three-dimensional laser scanner, and a cylindricity meter. Secondly, the high-precision point cloud data is pre-processed by using point cloud noise reduction, data reduction algorithms, or point cloud processing software tools. Then, the processed point cloud data is subjected to multi-scale surface morphology decomposition operations with the help of data analysis methods. The Mallat algorithm with multi-scale analysis function in wavelet analysis theory is used to realize the two-dimensional discrete wavelet transform of the height field of the surface morphology point cloud data, and the point cloud data is processed according to the zero-scale method. The frequency band division principle of the surface morphology error components of the part is followed, and the power spectral density method is used to obtain the frequency range of the decomposed shape error, waviness and roughness scale components; finally, the inverse process of the above decomposition is reconstructed to generate a skin surface shape model under the shape error scale to detect whether it meets the geometric dimension tolerance design requirements. If so, the surface certification of the part manufacturing solid model is qualified and it is used as the certified skin surface shape model. Otherwise, the part surface is deemed unqualified and it is necessary to readjust the high-precision measuring equipment or re-process the part before inspection and certification until the surface certification is qualified.

[0015] Furthermore, the specific implementation process of step (2) is divided into the following three stages:

[0016] (21) Skin surface shape model training set data preparation stage: The standard and certified skin surface shape models generated in step (1) are preprocessed using a dimensionality reduction method to obtain a series of standard two-dimensional images. By fusing the virtual and real image data, a "real" training sample set is formed, thereby completing the skin surface shape model training set data preparation work;

[0017] (22) Deep learning model construction and training phase: further constructing the generator, discriminator and loss function in the generative adversarial network, and performing model training based on the established generative adversarial network model, including training the generator and discriminator, and realizing the construction of a deep learning model based on the generative adversarial network by inputting a Gaussian random noise vector in the latent space and adjusting the loss function;

[0018] (23) Surface three-dimensional deviation morphological unit model generation stage: The constructed deep learning model based on the generative adversarial network is used to generate a two-dimensional image of the part surface, and the deviation pattern mapping relationship between the two-dimensional image and the three-dimensional model is used to obtain a three-dimensional deviation morphological unit model of the corresponding part surface that meets the functional requirements.

[0019] Furthermore, the specific implementation method of the skin surface shape model training set data preparation stage in step (2) is:

[0020] The surface of the part is divided into two categories according to plane and cylindrical surface. Through step (1), multiple simulations and test certifications are performed to generate n1 standard skin surface shape models and n2 certified skin surface shape models as data sources. The Hessian local linear embedding algorithm is used to reduce the data dimension of the skin surface shape model, transform it into a two-dimensional coordinate system, and normalize the height field deviation data to the interval [-1,1]. Then, it is mapped and transformed according to the 256-level pixel intensity in the grayscale image to form a standard image data set with a matrix size of N×N: Where N represents the image pixel resolution used for deep learning model training, x i represents the standard image sample, i represents the number of the standard image sample, n represents the number of the standard image samples, and the specific value is n=n1+n2, R d Represents a real number set of d dimensions, thereby completing the preparation of the skin surface shape model training set data.

[0021] Furthermore, the specific implementation method of the deep learning model construction and adjustment training phase in step (2) is:

[0022] The deep learning model uses the original generative adversarial network framework and is divided into two parts: a generator G and a discriminator D. The structure of the generator G is a multi-layer neural network, which is used to generate image data {G(z)} with the same size as the standard image in the skin surface shape model training set data, where z represents a Gaussian random noise vector in the potential input space; the structure of the discriminator D is also a multi-layer neural network, which is used to distinguish the properties of the input standard image: it is the "real" image data {x i}, or the "forged" image data {G(z)} given by the generator G; the multi-layer neural network training method of the generator G and the discriminator D adopts the Adam method, and the number of hidden layers, the selection of activation function, and the selection of learning rate depend on the specific problem; generative adversarial training is performed by alternately training the discriminator D and the generator G, and the binary cross entropy loss function for updating the parameters of the discriminator D and the generator G is: Loss(D,G):=E x {D(x)}+E x {1-D(G(z))}, where Loss(D,G) represents the name of the loss function of the generative adversarial network model, E x Denotes the expected calculation for x, D(G(z)) denotes the degree of confidence of the discriminator D in G(z), and the loss function is used to update the parameters of the discriminator D and the generator G. When the discriminator D cannot distinguish whether the input standard image is the "real" image data of the skin surface shape model training set, or the "fake" image data given by the generator G, it indicates that the generator G and the discriminator D have reached the Nash equilibrium point. At this time, the training of the generative adversarial network model is terminated; at this time, the G(z) given by the generator G is the valid output of the model.

[0023] Furthermore, the specific implementation method of step (3) is:

[0024] (31) First, the key planes and / or key cylindrical surfaces of the parts are separated and discretized from the CAD model for preprocessing;

[0025] (32) Secondly, the principal component analysis method is used to obtain the eigenvalues ​​and corresponding eigenvectors of the ideal surface model of the part and the three-dimensional deviation morphological unit model obtained in step 2, and the eigenvalues ​​are sorted from large to small, and are used as the first principal axis, the second principal axis and the third principal axis in sequence, wherein the first principal axis is the X-axis direction vector, the second principal axis is the Y-axis direction vector, and the third principal axis is the Z-axis direction vector, and then the principal axes of the above two surface models are matched one by one;

[0026] (33) Then, through a series of operations such as rotation, translation, and scaling, the ideal surface model of the part is converted into the size of a three-dimensional deviation morphological unit model, and the three-dimensional deviation morphological pattern is associated and mapped onto the ideal surface model through an interpolation method to realize the deviation morphological construction of the ideal surface model of the part, and the offset size of its ideal discrete point is determined by the deviation value of the corresponding point of the three-dimensional deviation morphological unit model in the normal direction;

[0027] (34) Finally, the part surface model after discrete point offset is triangulated and subjected to the corresponding inverse operation to obtain the final training skin surface shape model of the part surface, which is used as the representation of the part digital twin surface model.

[0028] The present invention also discloses an implementation framework formed by using the method for rapid characterization of a part digital twin surface model based on virtual-real fusion, wherein the implementation framework includes a physical entity layer, a virtual-real data layer, and a surface twin layer;

[0029] The physical entity layer is a collection of entities that objectively exist in the physical world to build a digital twin surface model of a part, and mainly includes three parts: computer system entities, physical scene entities, and high-precision measurement equipment entities. Among them, the computer system entity is a computer system device with relevant software such as computer-aided design, finite element analysis, and point cloud processing; the physical scene entity includes part entities, processing equipment, tooling, fixtures, and related operators; and the high-precision measurement equipment entity includes equipment such as a three-coordinate measuring machine, a three-dimensional laser scanner, and a cylindricity meter;

[0030] The virtual-real data layer includes two parts: virtual data and actual data, wherein the virtual data depends on the standard surface shape model data of the corresponding surface obtained from the ideal part design model through the mechanism modeling analysis method, and the actual data depends on the certified surface shape model data of the corresponding surface obtained from the actual part entity model through the measurement data processing method;

[0031] The surface twin layer uses virtual-real fusion data as a training sample set, and generates a three-dimensional deviation morphological unit model of the corresponding part surface through the construction and adjustment of a deep learning model based on a generative adversarial network. Then, the correlation mapping relationship between the ideal surface model of the part and its three-dimensional deviation morphology is used to obtain the training skin surface shape model of the corresponding part surface, which is used as the representation of the digital twin surface model of the part, thereby realizing fast and high-fidelity digital mirror expression of the surface of the physical part.

[0032] The rapid characterization method of the digital twin surface model of parts based on virtual-reality fusion fully integrates the virtual simulation of the part surface features with the actual measurement data, and uses the trained deep learning model based on the generative adversarial network to quickly and accurately obtain the training surface shape model of the corresponding part surface. It provides an effective and feasible technical approach to achieve rapid and high-fidelity characterization of the digital twin surface model of product parts, and has important engineering significance for realizing high-precision assembly simulation of complex precision electromechanical products and improving assembly quality consistency and assembly efficiency.

[0033] The beneficial effects of the present invention are:

[0034] The method of the present invention not only solves the problems of uncertainty and poor adaptability in modeling the standard surface shape model of parts, but also overcomes the problems of time-consuming, labor-intensive and complicated operation in modeling the certified surface shape model of parts. It also avoids the problem of inconsistency between the data sources of virtual and real part surface models. The method deeply integrates the virtual simulation of the surface features of the parts with the actual measurement data, and uses a deep learning model based on a generative adversarial network to generate a digital twin surface model of the parts, thereby realizing rapid, accurate and high-fidelity digital representation of product parts, which can be used for high-precision assembly analysis and simulation of complex and precision electromechanical products. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 It is a schematic flow chart of the method of the present invention;

[0036] Figure 2 is a detailed flow chart of a method for rapid characterization of a digital twin surface model of a part based on virtual-real fusion in an embodiment;

[0037] Figure 3 2 is a schematic diagram of deep learning model calibration and training based on a generative adversarial network in an embodiment;

[0038] Figure 4 Schematic diagram of the mapping relationship between the part training surface shape model and the part entity model in the embodiment;

[0039] Figure 5 It is a structural block diagram of the system for implementing the present invention. DETAILED DESCRIPTION

[0040] The present invention will be further explained below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the following specific embodiments are only used to illustrate the present invention and are not used to limit the scope of the present invention.

[0041] like Figure 1-4 As shown, the method for rapid characterization of the surface model of a digital twin of a part based on virtual-real fusion in this embodiment includes the following three steps:

[0042] Step (1): Input the part design CAD model and the manufacturing solid model, obtain the surface feature information of the CAD model and the high-precision point cloud data of the corresponding surface of the solid model, and use the mechanism modeling and data analysis methods to obtain the standard and certified surface shape models respectively, wherein the part surface refers to the key plane and / or key cylindrical surface with geometric dimension tolerance design requirements and structural function requirements. The specific implementation method of step (1) is:

[0043] (11) For the part design CAD model, firstly, the geometric dimension tolerance data of the part surface is obtained through the model-based definition method of the computer-aided design software; secondly, the nominal surface of the part design CAD model is extracted and meshed by adopting feature segmentation and meshing algorithms or using computer-aided design and finite element analysis software tools; then, the position, direction and shape errors formed during the part surface processing are simulated by mechanism modeling, combining the surface characteristics of the part and the corresponding geometric dimension tolerance data. Among them, the position and direction errors are modeled by the small displacement spinor method, and the position and direction errors are randomly generated under the constraints of the small displacement spinor parameters, while the shape error is generated by the systematic error and random error of the processing process. The mechanism error is modeled, and the basis function method and the non-Gaussian random field method are used to construct the system error and the random error respectively. The shape error is generated after the two are superimposed, and it is detected whether it exceeds the upper and lower limits of the specified shape tolerance band. If it does not exceed the limit, the subsequent operation is continued, otherwise the scaling operation is performed to make it meet the shape tolerance design requirements before continuing the subsequent operation; finally, the above position and direction errors and shape errors are superimposed to detect whether they exceed the upper and lower limits of the specified position tolerance band. If it does not exceed the limit, the generated skin surface shape model is valid, otherwise the operation of generating position and direction errors and shape errors by mechanism modeling is repeated until an effective skin surface shape model is generated, and this is used as the standard skin surface shape model that meets the geometric dimension tolerance design requirements;

[0044] (12) For the manufacturing of solid models of parts, firstly, high-precision point cloud data of the corresponding surface of the parts is obtained by using high-precision measuring equipment such as a three-dimensional coordinate measuring machine, a three-dimensional laser scanner, and a cylindricity meter. Secondly, the high-precision point cloud data is pre-processed by using point cloud noise reduction, data reduction algorithms, or point cloud processing software tools. Then, the processed point cloud data is subjected to multi-scale surface morphology decomposition operations with the help of data analysis methods. The Mallat algorithm with multi-scale analysis function in wavelet analysis theory is used to realize the two-dimensional discrete wavelet transform of the height field of the surface morphology point cloud data, and the point cloud data is processed according to the zero-scale method. The frequency band division principle of the surface morphology error components of the part is followed, and the power spectral density method is used to obtain the frequency range of the decomposed shape error, waviness and roughness scale components; finally, the inverse process of the above decomposition is reconstructed to generate a skin surface shape model under the shape error scale to detect whether it meets the geometric dimension tolerance design requirements. If so, the surface certification of the part manufacturing solid model is qualified and it is used as the certified skin surface shape model. Otherwise, the part surface is deemed unqualified and it is necessary to readjust the high-precision measuring equipment or re-process the part before inspection and certification until the surface certification is qualified.

[0045] Step (2): After preprocessing the above two skin surface shape models, a training sample set is formed by virtual-real fusion, and a deep learning model based on a generative adversarial network is constructed to generate a surface three-dimensional deviation morphological unit model that meets the functional requirements. The specific implementation process of step (2) is divided into three stages:

[0046] (21) Skin surface shape model training set data preparation stage: The standard and certified skin surface shape models generated in step (1) are preprocessed by dimensionality reduction method to obtain a series of standard two-dimensional images. By integrating the virtual and real image data, a "real" training sample set is formed, thereby completing the skin surface shape model training set data preparation work. Furthermore, the specific implementation method of this stage is as follows: the part surface is divided into two categories according to plane and cylindrical surface, and multiple simulations and inspections are performed through step (1) to generate n1 standard skin surface shape models and n2 certified skin surface shape models as data sources; the Hessian local linear embedding algorithm is used to perform data dimensionality reduction processing on the skin surface shape model, transform it into a two-dimensional coordinate system, and normalize the height field deviation data to the interval [-1,1], and then map and transform it according to the 256-level pixel intensity in the grayscale image to form a standard image data set with a matrix size of N×N: Where N represents the image pixel resolution used for deep learning model training, x i represents the standard image sample, i represents the number of the standard image sample, n represents the number of the standard image samples, and the specific value is n=n1+n2, R dRepresents a real number set of d dimensions, thereby completing the preparation of the skin surface shape model training set data.

[0047] (22) Deep learning model construction and training phase: further construct the generator, discriminator and loss function in the generative adversarial network, and perform model training based on the established generative adversarial network model, including training the generator and discriminator, and realizing the construction of a deep learning model based on the generative adversarial network by inputting a Gaussian random noise vector in the latent space and adjusting the loss function. Furthermore, the specific implementation method of this phase is as follows: the deep learning model uses the original generative adversarial network framework and is divided into two parts: the generator G and the discriminator D; the structure adopted by the generator G is a multi-layer neural network, which is used to generate image data {G(z)} of the same size as the standard image in the skin surface shape model training set data, where z represents the Gaussian random noise vector in the latent input space; the structure adopted by the discriminator D is also a multi-layer neural network, which is used to distinguish the properties of the input standard image: it is the "real" image data {x i}, or the "forged" image data {G(z)} given by the generator G; the multi-layer neural network training method of the generator G and the discriminator D adopts the Adam method, and the number of hidden layers, the selection of activation function, and the selection of learning rate depend on the specific problem; generative adversarial training is performed by alternately training the discriminator D and the generator G, and the binary cross entropy loss function for updating the parameters of the discriminator D and the generator G is: Loss(D,G):=E x {D(x)}+E x {1-D(G(z))}, where Loss(D,G) represents the name of the loss function of the generative adversarial network model, E x Denotes the expected calculation for x, D(G(z)) denotes the degree of confidence of the discriminator D in G(z), and the loss function is used to update the parameters of the discriminator D and the generator G. When the discriminator D cannot distinguish whether the input standard image is the "real" image data of the skin surface shape model training set, or the "fake" image data given by the generator G, it indicates that the generator G and the discriminator D have reached the Nash equilibrium point. At this time, the training of the generative adversarial network model is terminated; at this time, the G(z) given by the generator G is the valid output of the model.

[0048] (23) Surface three-dimensional deviation morphological unit model generation stage: The constructed deep learning model based on the generative adversarial network is used to generate a two-dimensional image of the part surface, and the deviation pattern mapping relationship between the two-dimensional image and the three-dimensional model is used to obtain a three-dimensional deviation morphological unit model of the corresponding part surface that meets the functional requirements.

[0049] Step (3): Establish an association mapping between the ideal surface model of the part CAD and its three-dimensional deviation shape, obtain a training surface shape model, and use it as a representation of the surface model of the digital twin of the part. Furthermore, the specific implementation method of step (3) is as follows:

[0050] (31) First, the key planes and / or key cylindrical surfaces of the parts are separated and discretized from the CAD model for preprocessing;

[0051] (32) Secondly, the principal component analysis method is used to obtain the eigenvalues ​​and corresponding eigenvectors of the ideal surface model of the part and the three-dimensional deviation morphological unit model obtained in step 2, and the eigenvalues ​​are sorted from large to small, and are used as the first principal axis, the second principal axis and the third principal axis in sequence, wherein the first principal axis is the X-axis direction vector, the second principal axis is the Y-axis direction vector, and the third principal axis is the Z-axis direction vector, and then the principal axes of the above two surface models are matched one by one;

[0052] (33) Then, through a series of operations such as rotation, translation, and scaling, the ideal surface model of the part is converted into the size of a three-dimensional deviation morphological unit model, and the three-dimensional deviation morphological pattern is associated and mapped onto the ideal surface model through an interpolation method to realize the deviation morphological construction of the ideal surface model of the part, and the offset size of its ideal discrete point is determined by the deviation value of the corresponding point of the three-dimensional deviation morphological unit model in the normal direction;

[0053] (34) Finally, the part surface model after discrete point offset is triangulated and the corresponding inverse operation is performed to obtain the final training skin surface shape model of the part surface, and this is used as the representation of the part digital twin surface model. Through the above processing, the generated training skin surface shape model is used as the part digital twin surface model, which can realize the rapid and accurate simulation of the non-ideal surface of the simulated part, and is used to replace the part CAD model, solid model or other skin surface shape model, providing an efficient, high-fidelity, high-quality data model foundation for the assembly precision analysis and predictive simulation of complex precision electromechanical products, and then used to improve the assembly quality consistency of complex precision electromechanical products.

[0054] like Figure 5 As shown, the implementation framework of the method for rapid characterization of the surface model of a digital twin of a part based on virtual-real fusion in this embodiment includes a physical entity layer, a virtual-real data layer, and a surface twin layer. The specific contents are as follows:

[0055] The physical entity layer is a collection of entities that objectively exist in the physical world to build the digital twin surface model of parts. It mainly includes three parts: computer system entities, physical scene entities, and high-precision measurement equipment entities. Among them, the computer system entity is a computer system device with relevant software such as computer-aided design, finite element analysis, and point cloud processing. The physical scene entity includes part entities, processing equipment, tooling, fixtures, and related operators, and the high-precision measurement equipment entity includes equipment such as three-coordinate measuring machines, three-dimensional laser scanners, and cylindricity meters.

[0056] The virtual-real data layer includes virtual data and actual data. The virtual data depends on the standard surface shape model data of the corresponding surface obtained from the ideal part design model through the mechanism modeling analysis method, and the actual data depends on the certified surface shape model data of the corresponding surface obtained from the actual part solid model through the measurement data processing method.

[0057] The surface twin layer uses virtual-reality fusion data as a training sample set, and builds and adjusts a deep learning model based on a generative adversarial network to generate a three-dimensional deviation morphological unit model of the corresponding part surface. Then, the correlation mapping relationship between the ideal surface model of the part and its three-dimensional deviation morphology is used to obtain the training skin surface shape model of the corresponding part surface, which is used as the representation of the digital twin surface model of the part, thereby realizing fast and high-fidelity digital mirror expression of the surface of the physical part.

[0058] It should be noted that the above content only illustrates the technical idea of ​​the present invention and cannot be used to limit the protection scope of the present invention. For ordinary technicians in this technical field, several improvements and modifications can be made without departing from the principle of the present invention. These improvements and modifications all fall within the protection scope of the claims of the present invention.

Claims

1. A method for rapid characterization of the surface model of a digital twin of a part based on virtual-reality fusion drive, characterized in that: The following steps are involved: Step (1): Input the part design CAD model and the manufacturing solid model, obtain the surface feature information of the CAD model and the high-precision point cloud data of the corresponding surface of the solid model, and use mechanism modeling and data analysis methods to obtain the standard and certified surface shape models respectively; Step (2): After preprocessing the above two skin surface shape models, a training sample set is formed by virtual-real fusion, and a deep learning model based on a generative adversarial network is constructed to generate a surface three-dimensional deviation morphological unit model that meets the functional requirements; Step (3): Establish a correlation mapping between the ideal CAD surface model of the part and its three-dimensional deviation shape, obtain a training surface shape model, and use it as a representation of the digital twin surface model of the part.

2. According to claim 1, a method for rapid characterization of a digital twin surface model of a part based on virtual-real fusion drive is characterized in that: In the step (1), the part surface refers to a critical plane and / or critical cylindrical surface having geometric dimension tolerance design requirements and structural function requirements.

3. A method for rapid characterization of a digital twin surface model of a part based on virtual-real fusion drive according to claim 1 or 2, characterized in that: The specific implementation method of step (1) is: (11) For the part design CAD model, firstly, the geometric dimension tolerance data of the part surface is obtained through the model-based definition method of the computer-aided design software; secondly, the nominal surface of the part design CAD model is extracted and meshed by adopting feature segmentation and meshing algorithms or using computer-aided design and finite element analysis software tools; then, the position, direction and shape errors formed during the part surface processing are simulated by mechanism modeling, combining the surface characteristics of the part and the corresponding geometric dimension tolerance data. Among them, the position and direction errors are modeled by the small displacement spinor method, and the position and direction errors are randomly generated under the constraints of the small displacement spinor parameters, while the shape error is generated by the systematic error and random error of the processing process. The mechanism error is modeled, and the basis function method and the non-Gaussian random field method are used to construct the system error and the random error respectively. The shape error is generated after the two are superimposed, and it is detected whether it exceeds the upper and lower limits of the specified shape tolerance band. If it does not exceed the limit, the subsequent operation is continued, otherwise the scaling operation is performed to make it meet the shape tolerance design requirements before continuing the subsequent operation; finally, the above position and direction errors and shape errors are superimposed to detect whether they exceed the upper and lower limits of the specified position tolerance band. If it does not exceed the limit, the generated skin surface shape model is valid, otherwise the operation of generating position and direction errors and shape errors by mechanism modeling is repeated until an effective skin surface shape model is generated, and this is used as the standard skin surface shape model that meets the geometric dimension tolerance design requirements; (12) For the manufacturing of solid models of parts, firstly, high-precision point cloud data of the corresponding surface of the parts is obtained by using high-precision measuring equipment such as a three-dimensional coordinate measuring machine, a three-dimensional laser scanner, and a cylindricity meter. Secondly, the high-precision point cloud data is pre-processed by using point cloud noise reduction, data reduction algorithms, or point cloud processing software tools. Then, the processed point cloud data is subjected to multi-scale surface morphology decomposition operations with the help of data analysis methods. The Mallat algorithm with multi-scale analysis function in wavelet analysis theory is used to realize the two-dimensional discrete wavelet transform of the height field of the surface morphology point cloud data, and the point cloud data is processed according to the zero-scale method. The frequency band division principle of the surface morphology error components of the part is followed, and the power spectral density method is used to obtain the frequency range of the decomposed shape error, waviness and roughness scale components; finally, the inverse process of the above decomposition is reconstructed to generate a skin surface shape model under the shape error scale to detect whether it meets the geometric dimension tolerance design requirements. If so, the surface certification of the part manufacturing solid model is qualified and it is used as the certified skin surface shape model. Otherwise, the part surface is deemed unqualified and it is necessary to readjust the high-precision measuring equipment or re-process the part before inspection and certification until the surface certification is qualified.

4. A method for rapid characterization of a part digital twin surface model based on virtual-real fusion according to claim 1, 2 or 3, characterized in that: The specific implementation process of step (2) is divided into three stages: (21) Skin surface shape model training set data preparation stage: The standard and certified skin surface shape models generated in step (1) are preprocessed using a dimensionality reduction method to obtain a series of standard two-dimensional images. By fusing the virtual and real image data, a "real" training sample set is formed, thereby completing the skin surface shape model training set data preparation work; (22) Deep learning model construction and training phase: further constructing the generator, discriminator and loss function in the generative adversarial network, and performing model training based on the established generative adversarial network model, including training the generator and discriminator, and realizing the construction of a deep learning model based on the generative adversarial network by inputting a Gaussian random noise vector in the latent space and adjusting the loss function; (23) Surface three-dimensional deviation morphological unit model generation stage: The constructed deep learning model based on the generative adversarial network is used to generate a two-dimensional image of the part surface, and the deviation pattern mapping relationship between the two-dimensional image and the three-dimensional model is used to obtain a three-dimensional deviation morphological unit model of the corresponding part surface that meets the functional requirements.

5. According to the method for rapid characterization of the surface model of a digital twin of a part based on virtual-real fusion according to claim 4, it is characterized in that: The specific implementation method of the skin surface shape model training set data preparation stage is: The surface of the part is divided into two categories according to plane and cylindrical surface. Through step (1), multiple simulations and test certifications are performed to generate n1 standard skin surface shape models and n2 certified skin surface shape models as data sources. The Hessian local linear embedding algorithm is used to reduce the data dimension of the skin surface shape model, transform it into a two-dimensional coordinate system, and normalize the height field deviation data to the interval [-1,1]. Then, it is mapped and transformed according to the 256-level pixel intensity in the grayscale image to form a standard image data set with a matrix size of N×N: Where N represents the image pixel resolution used for deep learning model training, x i represents the standard image sample, i represents the number of the standard image sample, n represents the number of standard image samples, and the specific value is n=n1+n2, R d Represents a real number set of d dimensions, thereby completing the preparation of the skin surface shape model training set data.

6. According to claim 4, a method for rapid characterization of a digital twin surface model of a part based on virtual-real fusion is characterized in that: The specific implementation method of the deep learning model construction and adjustment training stage is: The deep learning model uses the original generative adversarial network framework and is divided into two parts: a generator G and a discriminator D. The structure of the generator G is a multi-layer neural network, which is used to generate image data {G(z)} with the same size as the standard image in the skin surface shape model training set data, where z represents a Gaussian random noise vector in the potential input space; the structure of the discriminator D is also a multi-layer neural network, which is used to distinguish the properties of the input standard image: it is the "real" image data {x i }, or the "forged" image data {G(z)} given by the generator G; the multi-layer neural network training method of the generator G and the discriminator D adopts the Adam method, and the number of hidden layers, the selection of activation function, and the selection of learning rate depend on the specific problem; generative adversarial training is performed by alternately training the discriminator D and the generator G, and the binary cross entropy loss function for updating the parameters of the discriminator D and the generator G is: Loss(D,G):=E x {D(x)}+E x {1-D(G(z))}, where Loss(D,G) represents the name of the loss function of the generative adversarial network model, E x Denotes the expected calculation for x, D(G(z)) denotes the degree of confidence of the discriminator D in G(z), and the loss function is used to update the parameters of the discriminator D and the generator G. When the discriminator D cannot distinguish whether the input standard image is the "real" image data of the skin surface shape model training set, or the "fake" image data given by the generator G, it indicates that the generator G and the discriminator D have reached the Nash equilibrium point. At this time, the training of the generative adversarial network model is terminated; at this time, the G(z) given by the generator G is the effective output of the model.

7. The method for rapid characterization of a digital twin surface model of a part based on virtual-real fusion according to claim 1 or 2, characterized in that: The specific implementation method of step 3 is: (31) First, the key planes and / or key cylindrical surfaces of the parts are separated and discretized from the CAD model for preprocessing; (32) Secondly, the principal component analysis method is used to obtain the eigenvalues ​​and corresponding eigenvectors of the ideal surface model of the part and the three-dimensional deviation morphological unit model obtained in step 2, and the eigenvalues ​​are sorted from large to small, and are used as the first principal axis, the second principal axis and the third principal axis in sequence, wherein the first principal axis is the X-axis direction vector, the second principal axis is the Y-axis direction vector, and the third principal axis is the Z-axis direction vector, and then the principal axes of the above two surface models are matched one by one; (33) Then, through a series of operations such as rotation, translation, and scaling, the ideal surface model of the part is converted into the size of a three-dimensional deviation morphological unit model, and the three-dimensional deviation morphological pattern is associated and mapped onto the ideal surface model through an interpolation method to realize the deviation morphological construction of the ideal surface model of the part, and the offset size of its ideal discrete point is determined by the deviation value of the corresponding point of the three-dimensional deviation morphological unit model in the normal direction; (34) Finally, the part surface model after discrete point offset is triangulated and subjected to the corresponding inverse operation to obtain the final training skin surface shape model of the part surface, which is used as the representation of the part digital twin surface model.

8. An implementation architecture using the method for rapid characterization of a part digital twin surface model based on virtual-real fusion as described in any one of claims 1 to 7, characterized in that: The implementation architecture includes a physical entity layer, a virtual and real data layer, and a surface twin layer; The physical entity layer is a collection of entities that objectively exist in the physical world for constructing the digital twin surface model of parts, including three parts: computer system entities, physical scene entities, and high-precision measurement equipment entities. Among them, the computer system entity is a computer system device with computer-aided design, finite element analysis, and point cloud processing related software tools; the physical scene entity includes part entities, processing equipment, tooling, fixtures, and related operators; the high-precision measurement equipment entity includes three-coordinate measuring machines, three-dimensional laser scanners, and cylindricity testers; The virtual-real data layer includes two parts: virtual data and actual data, wherein the virtual data depends on the standard surface shape model data of the corresponding surface obtained from the ideal part design model through the mechanism modeling analysis method, and the actual data depends on the certified surface shape model data of the corresponding surface obtained from the actual part entity model through the measurement data processing method; The surface twin layer uses virtual-real fusion data as a training sample set, and generates a three-dimensional deviation morphological unit model of the corresponding part surface through the construction and adjustment of a deep learning model based on a generative adversarial network. Then, the correlation mapping relationship between the ideal surface model of the part and its three-dimensional deviation morphology is used to obtain the training skin surface shape model of the corresponding part surface, which is used as the representation of the digital twin surface model of the part, thereby realizing fast and high-fidelity digital mirror expression of the surface of the physical part.

Images