A cable-membrane structure digital twin model reconstruction method based on deep feature fusion

Abstract

The application discloses a kind of based on deep feature fusion's cable membrane structure digital twin model reconstruction method, including survey entity cable membrane space structure, arrangement acquisition station, acquisition cable membrane space structure three-dimensional point cloud data;The three-dimensional point cloud of cable membrane space structure is preprocessed and constructs dataset;RPMNet network architecture is improved, and spatial attention mechanism method is applied to the point cloud feature extraction layer of RPMNet network architecture;The three-dimensional point cloud of cable membrane space structure is finely registered;Based on the three-dimensional registration point cloud of cable membrane space structure, fine model reconstruction is carried out;Deviation analysis is carried out to the fine model reconstruction of cable membrane space structure and theoretical design model.The method realizes the accurate registration of three-dimensional point cloud data of cable membrane space structure, completes the high-fidelity digital twin model reconstruction of entity cable membrane structure space form, has the characteristics of high efficiency, low cost and wide coverage, has good applicability to cable membrane space structure, and can be widely applied to engineering practice.

Description

Technical Field

[0001] This invention relates to the fields of health monitoring of cable-membrane building structures and artificial intelligence 3D point cloud technology, specifically to a method for reconstructing a digital twin model of a cable-membrane structure based on deep feature fusion. Background Technology

[0002] Membrane structures, a novel lightweight and flexible spatial structure system that emerged in the mid-to-late 20th century, possess numerous advantages such as large span, light weight, diverse artistic designs, environmental friendliness and energy efficiency, good mechanical properties, strong self-cleaning ability, short construction period, and excellent economic efficiency. They have long been favored by designers, construction personnel, and university researchers. Domestic research on cable-membrane structures began in the late 1990s. Over the past nearly 30 years, researchers have conducted a series of studies on the material properties, computational theories, computational software development, design methods, structural system innovation and application, and construction methods of membrane structures, achieving many results. However, compared to the rapid development and application of cable-membrane structure design theory and construction technology, the detection, inspection, evaluation, and appraisal technologies for in-service cable-membrane spatial structures are relatively lagging behind, making it difficult to provide long-term and effective technical guarantees for the safe operation of such engineering structures.

[0003] Currently, achieving precise simulation and mapping of form and performance between virtual digital space and real physical space based on digital twin technology has become a hot research method for building structural health monitoring and predictive maintenance. The primary task in the practical application of digital twins is to create an accurate digital twin reconstruction model of the application object. Furthermore, accurate building structural twin models have become indispensable tools for assessing structural load-bearing capacity, safety, and maintenance throughout the entire service life of the structure. The key to safety testing and assessment of cable-membrane space structures lies in establishing a precise digital twin model of the actual service cable-membrane structure, which is a reverse-engineered representation of the solid structure, to facilitate load effect assessment and safety testing.

[0004] However, a common problem exists in the design and construction of cable-membrane spatial structures in China: a disconnect exists between construction and structural design, leading to significant accumulation of construction errors. Furthermore, during service, cable-membrane structures experience environmental erosion, material aging, load fatigue, and sudden overload effects due to external environmental loads, resulting in damage accumulation and resistance decay, significantly impacting their spatial morphology. Therefore, when conducting structural performance analysis and safety evaluation of physical cable-membrane spatial structures, theoretical design models are insufficient to reflect the actual spatial morphology. Digital simulation results based on theoretical models differ significantly from the actual structures, making it difficult to construct high-fidelity models and achieve a true "virtual-real" correspondence. Therefore, it is necessary to refine and reconstruct models of physical cable-membrane spatial structures to achieve digital twin monitoring and safety inspection.

[0005] Currently, the inspection of building structure shapes is often carried out through visual observation, traditional ruler measurement, and laser total station measurement. This method is not only inefficient but also labor-intensive, costly, and has limited coverage, failing to meet the accuracy requirements for the shape inspection of solid cable-membrane spatial structures. Furthermore, traditional BIM reverse modeling software cannot effectively model the irregular curved surfaces of cable-membrane spatial structures. When modeling the target structure using 3D point cloud data, point cloud data registration is required. Traditional methods for processing this data mainly involve ICP and its improved algorithms, but these methods involve numerous empirical parameter selections, are cumbersome, and are not universally applicable. Summary of the Invention

[0006] To address the aforementioned technical shortcomings, the present invention aims to provide a method for reconstructing a digital twin model of a cable-membrane structure based on deep feature fusion. This method proposes using a high-precision, high-resolution 3D laser scanner to acquire 3D point cloud data of the cable-membrane spatial structure in a non-contact manner. A deep learning-based 3D point cloud registration RPMNet network architecture is employed to intelligently register the acquired 3D point cloud model of the cable-membrane spatial structure. This algorithm is end-to-end, where all feature parameters and iterations previously requiring manual calculation are now completed iteratively by the neural network. Furthermore, Geomagic Studio software is used to reconstruct the registered point cloud model of the cable-membrane spatial structure into a digital model. Further, GeomagicControl software is used to perform 3D measurement and inspection on the reconstructed digital model and the theoretical calculation model of the cable-membrane spatial structure. This results in the construction of a refined reconstruction model of the physical cable-membrane spatial structure, providing assurance for the digital twin monitoring and safety inspection of cable-membrane spatial structures.

[0007] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:

[0008] This invention provides a method for reconstructing a digital twin model of a cable-membrane structure based on deep feature fusion, comprising the following steps:

[0009] Step 1: Investigate the actual cable membrane spatial structure, set up 3D laser scanner acquisition sites, and sequentially acquire 3D point cloud data of the cable membrane spatial structure;

[0010] Step 2: Preprocess the 3D point cloud of the cable membrane spatial structure and construct a dataset;

[0011] Step 3: Improve the RPMNet network architecture by applying the spatial attention mechanism to the point cloud feature extraction layer of the RPMNet network architecture;

[0012] Step 4: Train the 3D point cloud of the cable membrane spatial structure based on the improved RPMNet network model and complete the fine registration;

[0013] Step 5: Reconstruct a refined model based on the 3D registration point cloud of the cable membrane spatial structure;

[0014] Step 6: Perform a deviation analysis between the refined model reconstruction of the cable membrane spatial structure and the theoretical design model.

[0015] Preferably, in step 1, when collecting data on the tension membrane surface and supporting components of the cable-membrane space structure, it is necessary to perform a three-dimensional point cloud scan of the cable-membrane space structure in batches, around the cable-membrane structure, with each scan occurring at 20°-30° intervals.

[0016] Preferably, in step 1, when scanning the cable membrane structure, the device needs to be kept horizontally fixed throughout the process, without shaking or moving, to ensure that the collected cable membrane spatial structure point cloud data have the same shape and size, only with different coordinate positions. This avoids non-rigid transformations between the scanned point cloud data, which would affect the rigid transformation registration effect of the deep learning improved RPMNet model on the point cloud data.

[0017] Preferably, the method for preprocessing and constructing the dataset in step 2 is as follows:

[0018] Step 21: Use 3D point cloud processing software to manually delete noise points in the 3D point cloud data of the tension membrane surface and supporting components of the cable membrane space structure, including non-observed cable membrane target point cloud, noise point cloud with poor reflection of structural materials, and noise point cloud of moving targets in the detection environment;

[0019] Step 22: Use 3D point cloud processing software to perform secondary sampling on the point cloud data of the tension membrane surface and supporting components of the processed cable membrane space structure. While retaining the original acquisition features, reduce the amount of data in the initial acquisition text, thereby reducing the computational load of the improved RPMNet model and accelerating the model training convergence speed.

[0020] The three-dimensional point cloud data of the cable membrane spatial structure from different angles after data preprocessing are saved as follows: (x, y, z) 3D point cloud coordinate position information and (N) 3D normal vector information. x N y N z A .txt text file containing 6 columns of data;

[0021] Step 23: Construct a 3D point cloud registration dataset for the cable membrane structure based on the preprocessed point cloud data of the tension membrane surface and supporting components, and divide the dataset into training and testing sets according to a 9:1 ratio.

[0022] Preferably, in step 3, the feature extraction layer in the RPMNet point cloud registration model uses the PointNet++ feature extraction module; by introducing the SAM attention module, the feature extraction layer in the RPMNet point cloud registration model is improved, and the multi-angle pooling method is used to generate adaptive attention weights to generate spatial attention of local point cloud spatial features in the point cloud data, thereby realizing adaptive integration of local point features, strengthening the network architecture in applying point-by-point information from local spatial regions to enhance the connection between features, and achieving accurate registration of the three-dimensional point cloud data of the cable membrane spatial structure.

[0023] Preferably, the method for improving the feature extraction layer in the RPMNet point cloud registration model by introducing the SAM attention module is as follows: In the PointNet network structure, by introducing max pooling, only the maximum value is retained on all channels of the point cloud mapping high-dimensional features, thus obtaining high-dimensional point cloud features with 1 channel. The SAM attention module is further introduced to fuse the point cloud input channel features with parallel max pooling and average pooling respectively on the previously obtained high-dimensional point cloud features with 1 channel, thereby generating two channel feature representations. Convolution is used to train the weights, and the channel feature dimension of the aggregated features is trained using a shared parameter MLP to generate attention weights. Then, the activation function is used to activate the weights, and a dot product is performed with the original high-dimensional point cloud features with 1 channel to generate new high-dimensional point cloud features with 1 channel.

[0024] Preferably, in step 4, the improved RPMNet network model is trained based on the training set of three-dimensional point cloud data of the cable membrane spatial structure to obtain a three-dimensional point cloud registration pre-trained model of the cable membrane spatial structure. Then, the two point clouds of the cable membrane spatial structure to be registered are input to complete the fine point cloud registration.

[0025] Preferably, in step 5, the processed membrane structure point cloud model is imported into Geomagic Studio software, non-connection points and external isolated points are manually deleted, and then encapsulation is performed. The encapsulated model is viewed and quickly smoothed and edge burrs are removed. The cable-membrane spatial structure model reconstructed by Geomagic Studio software is exported as a .stl format file.

[0026] Preferably, in step 6, the cable membrane spatial structure reconstruction model.stl file and the theoretical design model.dwg file are imported into the same coordinate system in Geomagic Control software to analyze the shape deviation between the physical cable membrane spatial structure and the theoretical design model, and to correct the cable membrane spatial structure based on the theoretical design model, thereby constructing a refined reconstruction model of the physical cable membrane spatial structure.

[0027] The beneficial effects of this invention are as follows:

[0028] (1) This invention proposes to address the problem of 3D point cloud registration of cable membrane spatial structures by embedding the SAM spatial attention mechanism into the feature extraction layer of the NPMNet network architecture, which can extract significant spatial features from local point clouds, thereby enhancing the semantic information of point cloud features and achieving accurate registration of 3D point cloud data of cable membrane spatial structures.

[0029] (2) By improving the architecture of the RPMNet deep learning model, it is possible to achieve end-to-end accurate registration of three-dimensional point cloud data. Compared with traditional manual measurement and total station measurement methods, it can save a lot of manpower, material resources and financial resources, and improve the reconstruction efficiency of digital twin models of physical cable membrane spatial structures.

[0030] (3) By using the three-dimensional point cloud model of the cable membrane spatial structure based on precise registration, the model reconstruction and three-dimensional monitoring of Geomagic Studio and Geomagic Control software, the spatial geometric dimensions of the physical cable membrane structure can be obtained efficiently and accurately. A precise physical model can be created using the model reconstruction method. By comparing it with the theoretical calculation model, the deviation analysis between the physical cable membrane spatial structure and the theoretical design model can be realized.

[0031] (4) This invention is the first to combine three-dimensional laser scanning point cloud, improved deep learning feature fusion RPMNet network model, Geomagic Studio reverse modeling and Geomagic Control three-dimensional detection technology. It has the characteristics of high efficiency, low cost and wide coverage, and has good applicability to cable membrane spatial structures. It can be widely used in engineering practice.

[0032] (5) The digital twin model reconstruction method of cable membrane spatial structure based on deep learning feature fusion proposed in this invention is applicable to the reconstruction of the whole life cycle model of physical cable membrane spatial structure, and innovatively provides a model reconstruction basis for realizing the whole life cycle health monitoring and detection of physical cable membrane spatial structure. Attached Figure Description

[0033] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0034] Figure 1 A flowchart illustrating a method for reconstructing a digital twin model of a cable-membrane structure based on deep feature fusion, provided in an embodiment of the present invention;

[0035] Figure 2This invention provides a network model architecture based on the NPMNet model with the SAM attention mechanism added.

[0036] Figure 3 The model architecture of the SAM attention mechanism based on local point cloud spatial dimension features provided in the embodiments of the present invention;

[0037] Figure 4 The actual effect of the method for reconstructing a digital twin model of a cable membrane spatial structure based on deep learning feature fusion provided in the embodiments of the present invention. Detailed Implementation

[0038] The technical solutions of 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 some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0039] To address the issue that the theoretical design model of a cable-membrane structure can no longer accurately reflect its actual spatial form due to the accumulation of errors during actual construction and the continuous external environmental loads during service, thus hindering performance analysis and safety evaluation of the structure, this embodiment provides a method for reconstructing a digital twin model of a cable-membrane structure based on deep feature fusion. The specific steps include:

[0040] Step 1: Investigate the actual cable membrane spatial structure, set up 3D laser scanner acquisition sites, and sequentially acquire 3D point cloud data of the cable membrane spatial structure;

[0041] Step 2: Preprocess the 3D point cloud of the cable membrane spatial structure and construct a dataset;

[0042] Step 3: Improve the RPMNet network architecture by applying the spatial attention mechanism to the point cloud feature extraction layer of the RPMNet network architecture;

[0043] Step 4: Perform fine registration of the 3D point cloud of the cable membrane spatial structure based on the improved RPMNet network model;

[0044] Step 5: Reconstruct a refined model based on the 3D registration point cloud of the cable membrane spatial structure;

[0045] Step 6: Perform a deviation analysis between the refined model reconstruction of the cable membrane spatial structure and the theoretical design model.

[0046] This invention achieves precise registration of 3D point cloud data for cable-membrane spatial structures, completing the reconstruction of a high-fidelity digital twin model of the spatial morphology of the physical cable-membrane structure. This further lays the foundation for realizing a high-fidelity digital twin model for monitoring and safety inspection of cable-membrane spatial structures, ensuring a seamless integration of virtual and real elements. Furthermore, this invention combines 3D laser scanning point clouds, deep learning models, Geomagic Studio reverse modeling, and Geomagic Control 3D detection technology, offering high efficiency, low cost, and broad coverage. It has good applicability to cable-membrane spatial structures and can be widely applied in engineering practice.

[0047] like Figure 1 The flowchart below illustrates a method for reconstructing a digital twin model of a cable-membrane structure based on deep feature fusion, as provided in an embodiment of the present invention.

[0048] The specific implementation process for surveying the spatial structure of the cable membrane and setting up the acquisition points for the 3D laser scanner is as follows:

[0049] In practical engineering applications, cable-membrane spatial structures are generally characterized by large spans and diverse artistic designs, exhibiting irregular shapes. When using 3D scanners to collect data on the tension membrane surface and supporting components (steel frames, steel columns, or steel cables, etc.) of these structures, issues such as viewing angle and occlusion prevent the acquisition of complete structural information in a single scan. Therefore, it is necessary to perform 3D point cloud scanning of the cable-membrane structure in batches, around a ring shape and at regular intervals. Furthermore, during the scanning process, the 3D scanner must be kept horizontally and fixed throughout the entire process, without any shaking or movement. This ensures that the collected point cloud data of the cable-membrane spatial structure has the same shape and size, differing only in coordinate position. This avoids non-rigid transformations between the scanned point cloud data, which could affect the rigid transformation registration effect of the RPMNet model improved by deep learning on the point cloud data.

[0050] Specifically, for the umbrella-shaped tensile membrane structure of the survey laboratory, Leica 3D laser scanner measurement stations were set up around the supporting frame of the umbrella-shaped membrane structure at approximately 20° intervals. The Leica 3D laser scanner was used to perform 3D point cloud scanning on the umbrella-shaped tensile membrane structure in batches. After the scanning was completed, the 3D point cloud of the umbrella-shaped membrane structure was exported as a .txt file, for a total of 18 .txt files.

[0051] The specific implementation process for preprocessing the 3D point cloud of the cable membrane spatial structure and constructing the dataset is as follows:

[0052] Step 1: Use 3D point cloud processing software to manually delete noise points in the 3D point cloud data of the tension membrane surface and supporting components (steel frame, steel column or steel cable, etc.) of the cable membrane space structure. This mainly includes point clouds of non-observed cable membrane structure targets (such as trees, ground and other buildings), noise point clouds with poor reflection of structural materials, and other noise point clouds such as moving targets in the detection environment (such as pedestrians, vehicles, etc.).

[0053] Specifically, noise points in the 3D point cloud data of the umbrella-shaped tensile membrane structure acquired by the Leica 3D laser scanner were manually deleted using CloudCompare software, because the measurement environment was a laboratory environment, and the noise point cloud mainly consisted of the ground, other buildings, etc.

[0054] Step 2: Using 3D point cloud processing software, the point cloud data of the tension membrane surface and supporting components of the cable-membrane space structure are resampled. While retaining the original acquisition features, the amount of data in the initial acquisition text is reduced, thereby reducing the computational load of the improved RPMNet model and accelerating the model training convergence speed. The 3D point cloud data of the cable-membrane space structure from different angles after data preprocessing are saved as separate files containing the 3D coordinate position information (x, y, z) and 3D normal vector information (N). x N y N z A .txt text file containing 6 columns of data;

[0055] Specifically, CloudCompare software was used to perform secondary sampling on the point cloud of the denoised umbrella-shaped membrane structure. The sampling distance was set to 0.0015. After secondary sampling, the amount of text data of the displayed membrane point cloud was reduced from more than 3 million to more than 90,000. The preprocessed point cloud data of the umbrella-shaped membrane structure was saved, including the three-dimensional coordinate position information (x, y, z) and the three-dimensional normal vector information (N). x N y N z A .txt text file containing 6 columns of data;

[0056] Step 3: Construct a 3D point cloud registration dataset for the cable membrane structure based on the preprocessed point cloud data of the tension membrane surface and supporting components, and divide the dataset into training and testing sets according to a certain allocation ratio.

[0057] Specifically, a three-dimensional point cloud registration dataset for the umbrella-shaped tensile membrane structure was constructed based on the preprocessed point cloud data of the membrane surface, and the dataset was divided into a training set: test set ratio of 9:1.

[0058] The specific implementation process of improving the RPMNet network architecture by applying the spatial attention mechanism to the point cloud feature extraction layer of the RPMNet network architecture is as follows:

[0059] In the RPMNet point cloud registration model, the feature extraction layer mainly uses the PointNet++ feature extraction module, but the PointNet++ network's performance in local feature extraction still needs improvement. For example... Figure 2 As shown, the feature extraction layer in the RPMNet point cloud registration model is improved by introducing the SAM attention module. The multi-angle pooling method is used to generate adaptive attention weights to produce spatial attention of local point cloud spatial features in the point cloud data, realize adaptive integration of local point features, and strengthen the network architecture in applying point-by-point information from local spatial regions to enhance the connection between features, thereby achieving accurate registration of 3D point cloud data of cable membrane spatial structures.

[0060] The improvement to the RPMNet network architecture, which introduces the SAM attention module to enhance the feature extraction layer in the RPMNet point cloud registration model, is implemented as follows:

[0061] like Figure 3 As shown, in the PointNet network structure, max pooling is introduced to map the point cloud to high-dimensional features, retaining only the maximum value across all channels, thus obtaining high-dimensional point cloud features with one channel. A SAM attention module is further introduced to fuse the previously obtained high-dimensional point cloud features with one channel using parallel max pooling and average pooling, generating two channel feature representations. Convolutional layers are used for weight training, and a shared-parameter MLP is used to train the channel feature dimensions of the aggregated features to generate attention weights. Then, an activation function is used to activate the weights, and a dot product is performed with the original high-dimensional point cloud features with one channel to generate new high-dimensional point cloud features with one channel.

[0062] The specific implementation process of training and refining the 3D point cloud of the cable membrane spatial structure based on the improved RPMNet network model is as follows:

[0063] The improved RPMNet network model is trained based on the training set of 3D point cloud data of the cable membrane spatial structure to obtain a 3D point cloud registration pre-trained model of the cable membrane spatial structure. Further input of the point cloud data of the two points of the cable membrane spatial structure to be registered completes the fine point cloud registration.

[0064] The detailed implementation process of reconstructing the refined model based on the 3D registration point cloud of the cable membrane spatial structure is as follows:

[0065] like Figure 4As shown, the processed membrane structure point cloud model is imported into Geomagic Studio software. After manually deleting non-connection points and external isolated points, the model is encapsulated. The encapsulated model is then viewed and quickly smoothed and edge burrs are removed. The cable-membrane spatial structure model reconstructed by Geomagic Studio software is exported as a .stl file.

[0066] The specific implementation process of the deviation analysis between the refined model reconstruction of the cable membrane spatial structure and the theoretical design model is as follows:

[0067] like Figure 4 As shown, the .stl file of the cable membrane spatial structure reconstruction model and the .dwg file of the theoretical design model are imported into the same coordinate system of Geomagic Control software to analyze the shape deviation between the physical cable membrane spatial structure and the theoretical design model, and to correct the cable membrane spatial structure based on the theoretical design model, thereby constructing a refined reconstruction model of the physical cable membrane spatial structure.

[0068] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for reconstructing a digital twin model of a cable-membrane structure based on deep feature fusion, characterized in that, Includes the following steps: Step 1: Investigate the actual cable membrane spatial structure, set up 3D laser scanner acquisition sites, and sequentially acquire 3D point cloud data of the cable membrane spatial structure; Step 2: Preprocess the 3D point cloud of the cable membrane spatial structure and construct a dataset; Step 3: Improve the RPMNet network architecture by applying the spatial attention mechanism to the point cloud feature extraction layer of the RPMNet network architecture; Step 4: Train the 3D point cloud of the cable membrane spatial structure based on the improved RPMNet network model and complete the fine registration; Step 5: Reconstruct a refined model based on the 3D registration point cloud of the cable membrane spatial structure; Step 6: Perform a deviation analysis between the refined model reconstruction of the cable-membrane spatial structure and the theoretical design model; In step 3, the feature extraction layer in the RPMNet point cloud registration model uses the PointNet++ feature extraction module; By introducing the SAM attention module, the feature extraction layer in the RPMNet point cloud registration model is improved. The multi-angle pooling method is used to generate adaptive attention weights to produce spatial attention of local point cloud spatial features in the point cloud data, realize adaptive integration of local point features, strengthen the network architecture in applying point-by-point information from local spatial regions to enhance the connection between features, and achieve accurate registration of three-dimensional point cloud data of cable membrane spatial structure. The method for improving the feature extraction layer in the RPMNet point cloud registration model by introducing the SAM attention module is as follows: In the PointNet network structure, max pooling is introduced to map the point cloud to high-dimensional features, retaining only the maximum value on all channels, thus obtaining high-dimensional point cloud features with 1 channel. The SAM attention module is further introduced to fuse the point cloud input channel features with parallel max pooling and average pooling respectively on the previously obtained high-dimensional point cloud features with 1 channel, thereby generating two channel feature representations. Convolution is used to train the weights, and the channel feature dimension of the aggregated features is trained using a shared parameter MLP to generate attention weights. Then, an activation function is used to activate the weights, and a dot product is performed with the original high-dimensional point cloud features with 1 channel to generate new high-dimensional point cloud features with 1 channel.

2. The method for reconstructing a digital twin model of a cable-membrane structure based on deep feature fusion as described in claim 1, characterized in that, In step 1, when collecting data on the tension membrane surface and supporting components of the cable-membrane space structure, it is necessary to perform a three-dimensional point cloud scan of the cable-membrane space structure in batches, around the cable-membrane structure, with each scan occurring at 20°-30° intervals.

3. The method for reconstructing a digital twin model of a cable-membrane structure based on deep feature fusion as described in claim 2, characterized in that, In step 1, when scanning the cable membrane structure, the device must be kept horizontal and fixed throughout the process, without shaking or moving, to ensure that the collected point cloud data of the cable membrane spatial structure are exactly the same in shape and size, only with different coordinate positions. This avoids non-rigid transformations between the scanned point cloud data, which would affect the rigid transformation registration effect of the deep learning-improved RPMNet model on the point cloud data.

4. The method for reconstructing a digital twin model of a cable-membrane structure based on deep feature fusion as described in claim 3, characterized in that, The method for preprocessing and constructing the dataset in step 2 is as follows: Step 21: Use 3D point cloud processing software to manually delete noise points in the 3D point cloud data of the tension membrane surface and supporting components of the cable membrane space structure, including non-observed cable membrane target point cloud, noise point cloud with poor reflection of structural materials, and noise point cloud of moving targets in the detection environment; Step 22: Use 3D point cloud processing software to perform secondary sampling on the point cloud data of the tension membrane surface and supporting components of the processed cable membrane space structure. While retaining the original acquisition features, reduce the amount of data in the initial acquisition text, thereby reducing the computational load of the improved RPMNet model and accelerating the model training convergence speed. The three-dimensional point cloud data of the preprocessed cable membrane spatial structure from different angles are saved as separate data containing the three-dimensional coordinate position information of the point cloud. 3D normal vector information A .txt text file containing 6 columns of data; Step 23: Construct a 3D point cloud registration dataset for the cable membrane structure based on the preprocessed point cloud data of the tension membrane surface and supporting components, and divide the dataset into training and testing sets according to a 9:1 ratio.

5. The method for reconstructing a digital twin model of a cable-membrane structure based on deep feature fusion as described in claim 1, characterized in that, In step 4, the improved RPMNet network model is trained based on the training set of 3D point cloud data of the cable membrane spatial structure to obtain the 3D point cloud registration pre-trained model of the cable membrane spatial structure. Then, the point cloud data of the two points of the cable membrane spatial structure to be registered are input to complete the fine point cloud registration.

6. The method for reconstructing a digital twin model of a cable-membrane structure based on deep feature fusion as described in claim 5, characterized in that, In step 5, the processed membrane structure point cloud model is imported into Geomagic Studio software. After manually deleting non-connection points and external isolated points, the model is encapsulated. The encapsulated model is then viewed and quickly smoothed and edge burrs are removed. The cable-membrane spatial structure model reconstructed by Geomagic Studio software is exported as a .stl file.

7. The method for reconstructing a digital twin model of a cable-membrane structure based on deep feature fusion as described in claim 6, characterized in that, In step 6, the .stl file of the cable membrane spatial structure reconstruction model and the .dwg file of the theoretical design model are imported into the same coordinate system in the Geomagic Control software. The shape deviation between the physical cable membrane spatial structure and the theoretical design model is analyzed, and the cable membrane spatial structure based on the theoretical design model is corrected, thereby constructing a refined reconstruction model of the physical cable membrane spatial structure.

Images