A method for automatically building large-scale digital twin models

Abstract

The present invention discloses a method for automatically constructing a large-scale digital twin model, comprising the following steps: S1: using a three-dimensional model AI automatic recognition system with a self-attention mechanism to perceive the three-dimensional model, the three-dimensional model AI automatic recognition system comprising a downsampling module, an upsampling module, and a 3D Transformer extraction module; S2: using a downsampling module to downsample the input data. The present invention belongs to the field of digital twin model technology, and the technical effects achieved are: effectively reducing the damage to the local detail features of the model caused by the lightweight process, retaining the key information of the model as much as possible, solving problems in the storage, transmission, and rendering of the three-dimensional model, and effectively improving the rendering efficiency and performance of the model in large-scale digital twin scenarios. In addition, a variety of lightweight methods are provided to users according to different model scenarios to improve the controllability of lightweighting.

Description

Technical Field

[0001] The present invention relates to the technical field of digital twin models, and in particular to a method for automatically constructing a large-scale digital twin model. Background Art

[0002] Digital twin technology is a new type of processing technology that is now used in aerospace, industrial production, oil field mining and other fields. By building a model for the system, the business system can be monitored in real time.

[0003] The defects of the prior art are as follows:

[0004] During the lightweighting process, the local detailed features of the model are easily destroyed, and less key information of the model is retained, resulting in problems such as the inability to store, transmit and render the three-dimensional model, resulting in low rendering efficiency and performance of the model in the digital twin scenario. Summary of the Invention

[0005] To this end, the present invention provides a method for automatically constructing a large-scale digital twin model to solve the above-mentioned problems in the prior art.

[0006] In order to achieve the above object, the present invention provides the following technical solutions:

[0007] According to a first aspect of the present invention, a method for automatically constructing a large-scale digital twin model includes the following steps:

[0008] S1: The downsampling module in the 3D model AI automatic recognition system is used to downsample the input 3D model;

[0009] S2: The upsampling module in the 3D model AI automatic recognition system is used to upsample the downsampled 3D model, and the Transition Up module in the upsampling module is used to generate new points;

[0010] S3: Use the 3D Transformer Extraction module in the 3D Model AI Automatic Recognition System to position encode the new points generated by the Transition Up module and establish a 3D Transformer layer;

[0011] S4: Use AI automatic recognition technology to batch map the position-encoded 3D Transformer layers to obtain a new digital twin model.

[0012] Furthermore, the step S2 specifically includes the following steps:

[0013] M1: Use the farthest point sampling to select a subset P2 from the input 3D model point set P1 as the downsampling result, P2 belongs to P1;

[0014] M2: In P2, find its nearest k neighboring points in the original point set P1 through k-nearest neighbor search;

[0015] M3: For each sampled point, the features of its k nearest neighbors are merged together through maximum pooling;

[0016] M4: The aggregated features are further refined through the linear layer.

[0017] Furthermore, the step S3 specifically includes the following steps:

[0018] N1: Assign features to the points generated by the Transition Up module through interpolation and feature propagation, thereby restoring or enhancing local details;

[0019] N2: The Transition Up module is used in conjunction with the Transition Down module to form an encoding-decoding collaboration to fully extract and utilize the features of point cloud data.

[0020] Furthermore, the position coding in S4 specifically includes the following steps:

[0021] Q1: A multi-layer perceptron is used, and the dimension of its output vector is consistent with the dimension of the point's feature vector;

[0022] Q2: This allows location information to be effectively integrated into subsequent calculations;

[0023] Q3: The position encoding function is defined as follows:

[0024] δ=θ(p i -p j )

[0025] Q4: During the process, the model adjusts the parameters of position encoding according to task requirements.

[0026] Furthermore, the position coding in S4 specifically includes the following steps:

[0027] E1: Based on vector attention, using subtraction relationship, the position encoding is added to the attention vector and the transformed features;

[0028] E2: Define as follows:

[0029]

[0030] Among them, φ(x i ),Ψ(x j ) and α(x j) refer to Query, Key and Value in the attention mechanism respectively, γ is MLP, ρ is softmax, δ is position encoding, and ⊙ is the dot product operation.

[0031] The present invention has the following advantages: effectively reducing the damage to the local details of the model during the lightweight process, preserving the key information of the model as much as possible, solving problems in the storage, transmission and rendering of three-dimensional models, and effectively improving the rendering efficiency and performance of the model in large-scale digital twin scenarios;

[0032] In addition, a variety of lightweighting methods are provided to users according to different model scenarios to improve the controllability of lightweighting. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 This is a flowchart of the downsampling module in the automatic construction method of a large-scale digital twin model provided by the present invention.

[0034] Figure 2 This is a flowchart of the upsampling module in the automatic construction method of a large-scale digital twin model provided by the present invention.

[0035] Figure 3 This is a flowchart of the 3D Transformer in the automatic construction method of a large-scale digital twin model provided by the present invention.

[0036] Figure 4 A flowchart of automatic batch mapping in the automatic construction method of a large-scale digital twin model provided by the present invention.

[0037] Figure 5 This is the overall architecture diagram provided by the present invention.

[0038] Figure 6 This is the architectural rendering provided by the present invention.

[0039] Figure 7 The present invention provides a 3D Transformer three-dimensional model based on the self-attention mechanism.

[0040] Figure 8 This is a flowchart of the AI-based automatic batch replacement technology for 3D models.

[0041] Figure 9 It is a flow chart of the model’s multi-objective lightweight technology.

[0042] Figure 10 It is the diagram of the quadratic error calculation formula. DETAILED DESCRIPTION

[0043] The following describes the implementation of the present invention through specific embodiments. People skilled in the art can easily understand other advantages and effects of the present invention from the contents disclosed in this specification.

[0044] Example 1

[0045] like Figures 1 to 9 As shown, a method for automatically constructing a large-scale digital twin model in an embodiment of the first aspect of the present invention includes the following steps:

[0046] S1: The downsampling module in the 3D model AI automatic recognition system is used to downsample the input 3D model;

[0047] S2: The upsampling module in the 3D model AI automatic recognition system is used to upsample the downsampled 3D model, and the Transition Up module in the upsampling module is used to generate new points;

[0048] S3: Use the 3D Transformer Extraction module in the 3D Model AI Automatic Recognition System to position encode the new points generated by the Transition Up module and establish a 3D Transformer layer;

[0049] S4: Use AI automatic recognition technology to batch map the position-encoded 3D Transformer layers to obtain a new digital twin model.

[0050] In the above embodiments, it should be noted that the standard classification, feature extraction, automatic recognition, and intelligent construction of multi-source three-dimensional models are achieved, the quality and efficiency of model construction are improved, and the construction needs of digital twin scenes in complex scenarios are met;

[0051] The technical effects achieved by the above embodiments are: Using a 3D transformer algorithm with a self-attention mechanism, the 3D scene of the industrial digital twin is perceived, enabling automatic AI recognition of different types of 3D models in the scene. Different types of models in the digital twin have different feature information. Based on the self-attention mechanism, the correlation between various model features is adaptively calculated and feature aggregation is performed to achieve accurate recognition of the 3D model.

[0052] For the multi-source 3D models produced by various professional 3D design software, the optimized 3D models are integrated with static attribute data, production dynamic data, simulation data, and AI data through automatic model reconstruction and automatic optimization processing methods to generate digital twin models of various equipment and facilities, realizing dynamic visualization management and control of the entire life cycle of equipment and facilities.

[0053] Example 2

[0054] like Figures 1 to 4 As shown, a method for automatically constructing a large-scale digital twin model includes all the contents of Example 1. In addition, S2 specifically includes the following steps:

[0055] M1: Use the farthest point sampling to select a subset P2 from the input 3D model point set P1 as the downsampling result, P2 belongs to P1;

[0056] M2: In P2, find its nearest k neighboring points in point set P1 through k-nearest neighbor search;

[0057] M3: For each sampled point, the features of its k nearest neighbors are merged together through maximum pooling;

[0058] M4: The aggregated features are further refined through the linear layer.

[0059] In the above embodiment, it should be noted that the sampling point set is gradually reduced. The downsampling rate of each stage is [1, 4, 4, 4], so the point set generated in each stage is [N, N / 4, N / 16, N / 64, N / 256] of the input point set, where N is the number of input points. Among them, the upsampling and downsampling modules and the point cloud feature extraction module based on point transformer are the key modules of the network;

[0060] Automatic reconstruction of 3D models

[0061] In view of the similar characteristics of a large number of models, instantiation technology, feature matching technology, deep learning technology and other methods are adopted. Through the three-dimensional model AI automatic recognition technology of the self-attention mechanism, the three-dimensional model is perceived, and the standard classification, feature extraction, automatic recognition and intelligent construction of multi-source three-dimensional models are realized, thereby improving the quality and efficiency of model construction and meeting the construction needs of digital twin scenes in complex scenarios.

[0062] Example 3

[0063] like Figures 1 to 4 As shown, a method for automatically constructing a large-scale digital twin model includes all the contents of Example 1. In addition, S3 specifically includes the following steps:

[0064] N1: Assign features to the points generated by the Transition Up module through interpolation and feature propagation, thereby restoring or enhancing local details;

[0065] N2: The Transition Up module is used in conjunction with the Transition Down module to form an encoding-decoding coordination to fully extract and utilize the features of the point cloud data;

[0066] The entire AI neural network architecture, such as Figure 7 As shown in the figure, the network consists of four modules: a multi-layer perceptron (MLP), a downsampling module, an upsampling module, and a 3D transformer. It operates on a progressively smaller set of sampled points in five stages. The downsampling ratio in each stage is [1, 4, 4, 4], resulting in a point set of [N, N / 4, N / 16, N / 64, N / 256] of the input point set, where N is the number of input points. The upsampling and downsampling modules and the point cloud feature extraction module based on the point transformer are key components of the network.

[0067] Example 4

[0068] like Figures 1 to 4 As shown, a method for automatically constructing a large-scale digital twin model includes all the contents of Example 1. In addition, the position encoding in S4 specifically includes the following steps:

[0069] Q1: A multi-layer perceptron is used, and the dimension of its output vector is consistent with the dimension of the point's feature vector;

[0070] Q2: This allows location information to be effectively integrated into subsequent calculations;

[0071] Q3: The position encoding function is defined as follows:

[0072] δ=θ(p i -p j )

[0073] Q4: During the process, the model can adjust the parameters of position encoding according to task requirements.

[0074] In the above examples, it should be noted that in the 3D Transformer layer, each point interacts with other points to calculate attention weights. These weights determine the contribution of each point in the feature aggregation process. In this way, the 3D Transformer is able to learn complex relationships and contextual information between points;

[0075] For industrial digital twin scenarios, we first conduct efficient scene analysis using AI-based automatic recognition of 3D models based on a self-attention mechanism, building a model-type database {model id: type}. Users then set mapping information for different types of models, such as color, highlights, glossiness, and texture images, to build a type-type-map database {type: map information}. We then perform type matching to obtain {model id: map information}. Finally, we traverse each model in the scene file to automatically batch map the 3D models, improving the realism and fidelity of the digital twin scene.

[0076] Example 5

[0077] like Figures 1 to 4 As shown, a method for automatically constructing a large-scale digital twin model includes all the contents of Example 1. In addition, the position encoding in S4 specifically includes the following steps:

[0078] E1: Based on vector attention, using subtraction relationship, the position encoding is added to the attention vector and the transformed features;

[0079] E2: Define as follows:

[0080]

[0081] Among them, φ(x i ),Ψ(x j ) and α(x j ) refer to Query, Key and Value in the attention mechanism respectively, γ is MLP, ρ is softmax, δ is position encoding, and ⊙ is the dot product operation;

[0082] Model multi-objective lightweight technology

[0083] By developing an improved QEM (Quadric Error Mactrics) lightweighting algorithm based on folding costs, the team effectively minimizes the damage to local model details during the lightweighting process, preserving the model's key information as much as possible. This solves issues with 3D model storage, transmission, and rendering, effectively improving the rendering efficiency and performance of models in large-scale digital twin scenarios. Furthermore, the team provides users with a variety of lightweighting methods (such as edge deletion and vertex aggregation) tailored to different model scenarios, enhancing the controllability of lightweighting.

[0084] Example 6

[0085] like Figures 1 to 4 As shown, a method for automatically constructing a large-scale digital twin model includes all the contents of Example 1, and further includes the following steps:

[0086] S6: For industrial digital twin scenarios, we first conduct efficient scenario analysis using AI automatic recognition technology based on a self-attention mechanism for 3D models.

[0087] S7: Build model-type database {model id: type};

[0088] S8: The user sets the mapping information for different types of models and builds a type-mapping database {type: mapping information};

[0089] S9: Perform type matching and obtain {model id: texture information}.

[0090] In the above embodiment, it should be noted that according to the replacement instruction, the replacement parameters given in the replacement model list are retrieved, the new model is loaded, and the corresponding new model is called to scale and rotate it to match the old model. This AI-based model automatic batch replacement technology improves the accuracy, efficiency and automation of the replacement, and improves the realism and fidelity of the digital twin scene.

[0091] The three-dimensional model AI automatic recognition technology based on the self-attention mechanism performs efficient scene analysis, parses the parameters of each model, and builds a model category database, including (model id, type, geometric information). The geometric information includes: coordinates, height, radius, and rotation angle. According to the modification design requirements (original model id, original model type, target model id, target model type) and the model category database, the geometric parameters of the target model are calculated, and a replacement model list is constructed (original model id, original model type, geometric information of the original model, target model id, geometric information of the target model). According to the replacement instruction, the replacement parameters given in the replacement model list are retrieved, the new model is loaded, and the corresponding new model is called to scale and rotate it to match the old model. This AI-based model automatic batch replacement technology improves the accuracy, efficiency, and automation of replacement, and improves the realism and fidelity of the digital twin scene. Figure 5 This is a flowchart of the AI-based automatic batch replacement technology for 3D models.

[0092] Quadratic error calculation formula based on edge deletion and collapse cost Figure 10 As shown, Cvi is the vertex curvature, SLR is the vertex area, and rvi is the vertex flatness.

Claims

1. A method for automatically constructing a large-scale digital twin model, characterized in that: The steps include: S1: The downsampling module in the 3D model AI automatic recognition system is used to downsample the input 3D model; S2: The upsampling module in the 3D model AI automatic recognition system is used to upsample the downsampled 3D model, and the Transition Up module in the upsampling module is used to generate new points; S3: Use the 3D Transformer Extraction module in the 3D Model AI Automatic Recognition System to position encode the new points generated by the TransitionUp module and establish a 3D Transformer layer; S4: Use AI automatic recognition technology to batch map the position-encoded 3D Transformer layers to obtain a new digital twin model; The S3 specifically includes the following steps: N1: Assign features to the points generated by the Transition Up module through interpolation and feature propagation, thereby restoring or enhancing local details; N2: The Transition Up module is used in conjunction with the Transition Down module to form an encoding-decoding coordination to fully extract and utilize the features of the point cloud data; The position coding in S4 specifically includes the following steps: E1: Based on vector attention, using subtraction relationship, the position encoding is added to the attention vector and the transformed features; E2: Define as follows: Among them, φ(x i ),Ψ(x j ) and α(x j ) refer to Query, Key and Value in the attention mechanism respectively, γ is MLP, ρ is softmax, δ is position encoding, and ⊙ is the dot product operation.

2. The method for automatically constructing a large-scale digital twin model according to claim 1, characterized in that: The S2 specifically includes the following steps: M1: Use the farthest point sampling to select a subset P2 from the input 3D model point set P1 as the downsampling result, P2 belongs to P1; M2: In P2, find its nearest k neighboring points in point set P1 through k-nearest neighbor search; M3: For each sampled point, the features of its k nearest neighbors are merged together through maximum pooling; M4: The aggregated features are further refined through the linear layer.

3. The method for automatically constructing a large-scale digital twin model according to claim 1, characterized in that: The position coding in S4 specifically includes the following steps: Q1: A multi-layer perceptron is used, and the dimension of its output vector is consistent with the dimension of the point's feature vector; Q2: This allows location information to be effectively integrated into subsequent calculations; Q3: The position encoding function is defined as follows: δ=θ(p i -p j ) Q4: During the process, the model adjusts the parameters of position encoding according to task requirements.

Images