An AI-based historical building digital three-dimensional reconstruction method

Abstract

The application relates to the field of artificial intelligence technology, and provides a history building digitization three-dimensional reconstruction method based on AI, which relates to multi-source point cloud and image fusion, component identification, defect detection and three-dimensional complete modeling; the method adopts RandLA-Net and a full connection network to construct a component identification model, combines relative elevation and density and the like enhanced features to realize component defect area positioning; a two-dimensional spline surface fitting with L1 norm constraint is used to extract a terrain trend, and the defect identification precision is improved; a GAN model guided by three joint losses is constructed, three-dimensional form complete reconstruction of building components is completed, finally, a parameterized BIM model is generated in combination with texture images and point cloud data, and a digital building archive is output; the method realizes fine modeling and automatic reconstruction of a history building component level, improves the integrity and accuracy of digital management, and is suitable for cultural relic protection, building repair and heritage digitization scenes.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to an AI-based method for digital 3D reconstruction of historical buildings. Background Technology

[0002] With the rapid development of cultural heritage protection and digital modeling technologies, the digital processing of historical buildings has become an important technical means for the protection, restoration, and management of architectural heritage. By collecting and modeling the geometric form, structural features, and material information of historical buildings with high precision, it is possible not only to achieve the long-term preservation and research of precious cultural relics and buildings, but also to provide reliable data support for subsequent restoration design, structural assessment, and public display. Existing methods for digitizing historical buildings mainly rely on technologies such as laser scanning, photogrammetry, or BIM modeling. Among these, laser scanning and UAV photogrammetry can quickly acquire large-scale point cloud data and texture image information. However, due to the complexity of building structures, environmental obstruction, and historical damage, the acquired data often suffers from missing information, noise, and incomplete local structures, making it difficult to directly generate high-fidelity 3D models. Summary of the Invention

[0003] This invention aims to address the problems existing in the digital processing of historical buildings, such as difficulties in point cloud completion, low accuracy in identifying missing components, low quality of 3D modeling, and insufficient automation. It provides an AI-based 3D reconstruction method for historical buildings, achieving refined, automated, and high-fidelity digital modeling. Compared to existing technologies, this invention introduces for the first time a building structure recognition model composed of RandLA-Net and a fully connected network, combining enhanced feature vectors and spatial clustering to achieve high-precision component region division and defect location. It innovatively designs a two-dimensional cubic spline surface fitting method with L1 norm constraints to extract relative elevation features and enhance the geometric representation of point clouds. In the 3D reconstruction stage, a three-term joint loss based on adversarial generation discriminant terms, real sample gradient penalty terms, and generated sample gradient penalty terms is constructed to guide the generative adversarial network to complete high-fidelity defect completion, effectively improving the realism, robustness, and consistency of the discriminant spatial response of the reconstructed model. Through the above integrated design, this invention achieves refined restoration of missing components of historical buildings and generation of parametric BIM models, significantly improving the intelligence and automation level of digital modeling.

[0004] This invention provides an AI-based method for digital 3D reconstruction of historical buildings, which includes the following steps:

[0005] Step S1: Use drone equipment, high-resolution cameras, and 3D laser scanners to acquire multi-source point cloud data and texture image data of historical buildings;

[0006] Step S2: Use a registration algorithm to fuse multi-source point cloud data to generate fused point cloud data;

[0007] Step S3: Use the historical building structure identification model to identify component regions in the fused point cloud data, and locate the missing and non-missing areas of building components; the historical building structure identification model includes the RandLA-Net network and a fully connected network;

[0008] Step S4: Use a GAN model guided by a three-term joint loss to perform three-dimensional morphological reconstruction of the missing areas of building components, generating high-fidelity reconstructed component data; the GAN model guided by the three-term joint loss includes a generator and a discriminator;

[0009] Step S5: Map and register the fused point cloud data, high-fidelity reconstructed component data, and texture image data to generate a parametric BIM model. Output a complete digital building archive file containing geometric shape information, material attribute information, and historical restoration record information based on the parametric BIM model, so as to realize the refined digital management of historical buildings.

[0010] Furthermore, the process of using a historical building structure recognition model to identify component regions from the fused point cloud data and locate the missing areas of building components specifically includes the following steps:

[0011] Step S31: Extract the basic feature data and enhanced feature data of the fused point cloud data, combine them to form an extended feature vector, input the fully connected network, and map it to an 8-dimensional feature space through nonlinear projection to obtain 8-dimensional fused feature data; the enhanced feature data includes relative elevation features, point density features, height variance features, flatness features, and normal vertical component features;

[0012] Step S32: Input the 8-dimensional fused feature data into the RandLA-Net network, and use its sparse attention mechanism and lightweight residual structure to perform feature extraction and point-level classification to obtain the probability distribution of building component categories;

[0013] Step S33: Extract the maximum category probability label of each point in the probability distribution of building component categories to generate a point cloud data set of the same category; perform spatial clustering processing on the point cloud data set of the same category to obtain the spatial clustering result;

[0014] Step S34: Generate a structural component label map based on the probability distribution of building component categories, divide the component regions by combining spatial clustering results, calculate the local point density sparsity index and geometric consistency index of each component region, perform defect detection on the component regions, and locate the defective and non-defective areas of building components.

[0015] Furthermore, the process of extracting relative elevation features from the fused point cloud data specifically includes the following steps:

[0016] Step B1: Project the fused point cloud data horizontally onto a two-dimensional regular grid. In each grid cell, define a neighborhood region, select ground candidate points in the neighborhood region, and construct an initial sparse ground point set.

[0017] Step B2: Based on the initial sparse set of ground points, a two-dimensional cubic spline surface fitting method with L1 norm constraints is used to construct the ground height function;

[0018] Step B3: For each point in the fused point cloud data, call the ground height function to calculate the terrain elevation of the point's location, and subtract the terrain elevation from the point's original elevation to obtain the relative elevation feature value of the point. This completes the extraction of relative elevation features for all points in the fused point cloud, resulting in the relative elevation features.

[0019] Furthermore, the process of using a GAN model guided by a three-term joint loss to perform three-dimensional morphological reconstruction of missing areas in building components and generate high-fidelity reconstructed component data includes the following steps:

[0020] Step S41: Construct a complete component sample set based on the non-damaged areas of the building components; construct an initial pseudo-sample set based on the damaged areas of the building components;

[0021] Step S42: Input the initial pseudo-sample set into the generator to generate pseudo-samples for the completion process; input the complete component sample set and the pseudo-samples for the completion process into the discriminator to perform the true / false discrimination operation;

[0022] Step S43: Construct a joint loss function to optimize the generator's output and constrain the gradient change magnitude of the discriminator in the real and fake sample space; the joint loss function includes an adversarial generation discriminant term, a real sample gradient penalty term, and a generated sample gradient penalty term;

[0023] Step S44: Based on the joint loss function, the Adam optimizer is used to alternately update and train the generator and discriminator to obtain the trained generator model;

[0024] Step S45: Voxelize and scale-normalize the missing areas of the building components to obtain sparse three-dimensional tensor data; input the sparse three-dimensional tensor data into the trained generator model to perform three-dimensional morphological completion and reconstruction of the missing areas of the building components, and generate high-fidelity reconstructed component data.

[0025] By adopting the above solution, the beneficial effects achieved by the present invention are as follows:

[0026] This invention integrates UAV photography, high-resolution image acquisition, and laser scanning to construct a unified data foundation of multi-source point clouds and texture images. A registration algorithm is then used to achieve high-precision fusion, enabling comprehensive perception and unified representation of the complex geometric structures of historical buildings, providing a stable input source for subsequent component identification and reconstruction. The historical building structure recognition model, constructed by combining the RandLA-Net network and a fully connected network, improves the accuracy of identifying multi-scale, heterogeneous components and locating defective areas, solving the problem of traditional rule-based modeling methods struggling to handle large-scale unstructured point cloud data. Furthermore, the model incorporates enhanced features such as relative elevation, point density, and normal components, and extracts component boundaries through spatial clustering, enhancing the ability of defect detection to distinguish fine-grained components and effectively adapting to complex structural defects commonly found in historical buildings, such as localized erosion, fractures, and occlusions.

[0027] In the 3D completion process, this invention designs a Generative Adversarial Network (GAN) model guided by a three-term joint loss function, which significantly improves the reconstruction realism and discriminative consistency of the missing region. By introducing adversarial generation and discriminant terms, the generator is guided to output completion results that are closer to the shape of the real component. By using gradient penalty terms between real samples and generated samples, the smoothness and structural stability of the generation process are constrained, solving problems such as surface inconsistency and shape drift in traditional reconstruction algorithms. This completion mechanism completes the joint optimization of the generator and discriminator during the training phase, and can directly perform voxelization encoding and scale standardization of the missing region during the inference phase, realizing fast and high-fidelity component completion and reconstruction, and enhancing the adaptability and automation of the reconstruction process.

[0028] Ultimately, this invention unifies the registration of the completed component data with the original point cloud and texture images, constructs a parametric BIM model, and outputs a complete digital building archive file containing multi-dimensional information such as geometry, materials, and historical restoration records. This enables digital management of historical buildings throughout the entire process from perception, analysis, reconstruction to structured expression. This method not only improves the quality and accuracy of digital modeling but also supports standardized cultural relic archiving, digital asset management, and multi-platform sharing. It solves the problem of lacking structural-level completion and parametric-level expression in traditional methods, providing an efficient, intelligent, and scalable technical path for the protection, restoration design, and heritage information systems of historical buildings. Attached Figure Description

[0029] Figure 1 This is a comparison chart of the fitting results of the ground height function under different regularization constraints proposed in Example 4;

[0030] Figure 2 This is a trend chart showing the changes during the training process of the joint loss function proposed in Example 5.

[0031] Figure 1In the diagram, (a) only the L1 data fidelity term is used, resulting in a lot of surface jitter and poor noise resistance; (b) adding the second-order smoothing term makes the surface continuous and smooth, but local abrupt changes still exist; (c) adding the minimum surface area regularization term significantly suppresses abrupt changes in slope, making the fitted surface continuous and natural, which is particularly suitable for the terrain reconstruction needs of complex structures such as domes and sloping roofs in historical buildings; the left side represents the elevation value, and the bottom side represents the location.

[0032] Figure 2 In the diagram, the horizontal axis represents the number of training epochs, and the vertical axis represents the loss value. The black curve represents the joint loss, the blue curve represents the adversarial generation and discrimination term, the green curve represents the gradient penalty term for real samples, and the orange curve represents the gradient penalty term for generated samples. As the number of training epochs increases, the various losses gradually converge, and the generator and discriminator achieve good collaborative optimization under the constraint of the three joint losses. Detailed Implementation

[0033] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0034] Example 1: This invention provides an AI-based method for digital 3D reconstruction of historical buildings, which includes the following steps:

[0035] Step S1: Use drone equipment, high-resolution cameras, and 3D laser scanners to acquire multi-source point cloud data and texture image data of historical buildings;

[0036] Regarding historical building A, a brick-and-wood structure from the Qing Dynasty:

[0037] A 3D laser scanner acquires 3D point cloud data of a building with a resolution of 5mm, covering an area of ​​approximately 800m², and collecting approximately 23 million original point clouds.

[0038] A high-resolution camera captured texture images, taking a total of 286 images with a resolution of 8688×5792.

[0039] The drone equipment collects supplementary point cloud data and oblique photogrammetric images of high-altitude structures such as rooftops and eaves, with an accuracy of 10cm.

[0040] Step S2: Use a registration algorithm to fuse multi-source point cloud data to generate fused point cloud data;

[0041] Step S3: Use the historical building structure identification model to identify component regions in the fused point cloud data, and locate the missing and non-missing areas of building components; the historical building structure identification model includes the RandLA-Net network and a fully connected network;

[0042] Step S4: A three-term joint loss-guided GAN model is used to perform 3D morphological reconstruction of the missing areas of building components, generating high-fidelity reconstructed component data. The three-term joint loss-guided GAN model includes a generator and a discriminator. The construction process of the three-term joint loss-guided GAN model is as follows: A GAN model is constructed by introducing three joint loss functions: adversarial generation discriminant term, real sample gradient penalty term, and generated sample gradient penalty term. The training process of the generator and discriminator is optimized, and the consistency of the response and gradient smoothness of generated samples and real samples in the discriminator feature space are constrained, thereby obtaining the three-term joint loss-guided GAN model.

[0043] Step S5: Map and register the fused point cloud data, high-fidelity reconstructed component data, and texture image data to generate a parametric BIM model. Output a complete digital building archive file containing geometric shape information, material attribute information, and historical restoration record information based on the parametric BIM model, so as to realize the refined digital management of historical buildings.

[0044] Example 2, based on Example 1, describes the process of using a historical building structure recognition model to identify component regions from fused point cloud data and locate missing areas of building components. The specific steps include:

[0045] Step S31: Extract the basic feature data and enhanced feature data of the fused point cloud data, combine them to form an extended feature vector, input the fully connected network, and map it to an 8-dimensional feature space through nonlinear projection to obtain 8-dimensional fused feature data; the enhanced feature data includes relative elevation features, point density features, height variance features, flatness features, and normal vertical component features;

[0046] Step S32: Input the 8-dimensional fused feature data into the RandLA-Net network, and use its sparse attention mechanism and lightweight residual structure to perform feature extraction and point-level classification to obtain the probability distribution of building component categories;

[0047] Step S33: Extract the maximum category probability label of each point in the probability distribution of building component categories to generate a point cloud data set of the same category; perform spatial clustering processing on the point cloud data set of the same category to obtain the spatial clustering result;

[0048] Step S34: Generate a structural component label map based on the probability distribution of building component categories, divide the component regions by combining spatial clustering results, calculate the local point density sparsity index and geometric consistency index of each component region, perform defect detection on the component regions, and locate the defective and non-defective areas of building components.

[0049] Example 3, based on Example 1, describes the process of identifying component regions and locating missing areas of building components using fused point cloud data. The specific steps include:

[0050] Step E1: Extract the basic feature data of the fused point cloud data and input it into a fully connected network. Map the data to an 8-dimensional feature space through nonlinear projection to obtain 8-dimensional fused feature data.

[0051] Step E2: Input the 8-dimensional fused feature data into the RandLA-Net network, and use its sparse attention mechanism and lightweight residual structure to perform feature extraction and point-level classification to obtain the probability distribution of building component categories;

[0052] Step E3: Extract the maximum category probability label for each point in the probability distribution of building component categories to generate a point cloud dataset of the same category; perform spatial clustering on the point cloud dataset of the same category to obtain the spatial clustering results;

[0053] Step E4: Generate a structural component label map based on the probability distribution of building component categories, divide the component regions by combining spatial clustering results, calculate the local point density sparsity index and geometric consistency index for each component region, perform defect detection on the component regions, and locate the defective and non-defective areas of building components.

[0054] Example 4, according to Figure 1 This embodiment is based on Embodiment 2. In this embodiment, the process of extracting the relative elevation features from the fused point cloud data specifically includes the following steps:

[0055] Step B1: Project the fused point cloud data horizontally onto a two-dimensional regular grid. In each grid cell, define a neighborhood region, select ground candidate points in the neighborhood region, and construct an initial sparse ground point set to reflect the ground trend of the building area.

[0056] The horizontal projection grid size was set to 0.5m × 0.5m, resulting in a total of 3415 grids;

[0057] The initial set of sparse ground points is denoted as:

[0058] ;

[0059] in, Indicates the first Two-dimensional coordinates of one ground candidate point; No. The elevation values ​​corresponding to each ground candidate point; This represents the total number of candidate ground points;

[0060] This set of points serves as the control benchmark for the ground fitting curve;

[0061] Step B2: Based on the initial sparse set of ground points, a two-dimensional cubic spline surface fitting method with L1 norm constraints is used to construct the ground height function;

[0062] The process of constructing the ground height function using a two-dimensional cubic spline surface fitting method with L1 norm constraints:

[0063] Constructing L1 data fidelity items:

[0064] ;

[0065] in, Represents the fitting function At point The predicted output value at that location represents the fitted terrain elevation corresponding to that plane position. To represent absolute error, the L1 norm is used to enhance robustness; Indicates L1 data fidelity item;

[0066] This option is used to ensure that the fitted surface is as close as possible to the actual ground height at the control points, thus suppressing the influence of outliers.

[0067] Constructing a second-order smooth term (curvature control):

[0068] ;

[0069] in, This indicates the domain, which is the entire area covered by the point cloud data; Represents the horizontal coordinates in the point cloud. Indicates the depth coordinates in the point cloud; Indicates the fitted function in The second derivative of the direction; Indicates the fitted function in The second derivative of the direction; express and The mixed derivative; Represents a double integral; Represents a second-order smooth term;

[0070] This item is used to control the curvature of the surface to avoid producing discontinuous or jagged boundaries;

[0071] Constructing the minimum surface regularization term (area control):

[0072] ;

[0073] in, express gradient vector, Represents the square of the gradient magnitude; Represents the local area element of a surface (derived from the theory of minimum surfaces); Represents the regularization term for the minimum surface;

[0074] This feature is used to suppress drastic slope changes, ensuring surface continuity and physical geometric stability, and is particularly suitable for fitting pitched roofs and dome structures in historical buildings.

[0075] Construct the overall optimization objective function:

[0076] ;

[0077] in, This represents the fitted ground height function. This represents the balance coefficient, used to adjust the relative weights of fidelity and smoothness; The weighting coefficients of the area regularization term control the strength of the constraint on the change of surface slope. Represents the fitted function Find the solution that minimizes the objective function, and use it to construct the optimal terrain elevation function under sparse ground point constraints;

[0078] The ground height function constructed using the above formula achieves adaptive constraints on surface smoothness, geometric continuity, and slope changes while maintaining fitting accuracy. It is suitable for refined processing of point clouds of complex terrain and historical buildings. This function can be used in subsequent applications such as calculating relative elevation features and locating missing building components.

[0079] Figure 1 The image shows a comparison of the fitting results of the ground height function under different regularization constraints. In (a), only the L1 data fidelity term is used, resulting in more surface jitter and poor noise resistance. In (b), the addition of the second-order smoothing term makes the surface continuous and smooth, but local abrupt changes still exist. In (c), the addition of the minimum surface area regularization term significantly suppresses abrupt changes in slope, and the fitted surface is continuous and natural, which is particularly suitable for the terrain reconstruction needs of complex structures such as domes and pitched roofs in historical buildings. The left side represents the elevation value, and the bottom side represents the location.

[0080] Step B3: For each point in the fused point cloud data, call the ground height function to calculate the terrain elevation of the point's location, and subtract the terrain elevation from the point's original elevation to obtain the relative elevation feature value of the point. This completes the extraction of relative elevation features for all points in the fused point cloud, resulting in the relative elevation features.

[0081] Example 5, according to Figure 2 This embodiment is based on Embodiment 4. In this embodiment, a three-term joint loss-guided GAN model is used to perform three-dimensional morphological reconstruction of the missing areas of building components and generate high-fidelity reconstructed component data. The specific steps include:

[0082] Step S41: Construct a complete component sample set based on the non-damaged areas of the building components; construct an initial pseudo-sample set based on the damaged areas of the building components;

[0083] Step S42: Input the initial pseudo-sample set into the generator. The generator uses a residual structure consisting of bilinear upsampling, 1×1 convolution, and 3×3 sets of convolutions to reconstruct the features of the input data and generate pseudo-samples for the completion process. Input the complete component sample set and the pseudo-samples for the completion process into the discriminator. The discriminator uses a shallow fully convolutional structure and has a 4×4 depth convolution module and a linear mapping module at the output end to compress the features of the input samples, perform the real and fake discrimination operation, and realize the function of distinguishing between generated samples and real samples.

[0084] Step S43: Construct a joint loss function to optimize the generator's output and constrain the gradient change magnitude of the discriminator in the real and fake sample spaces; the joint loss function includes an adversarial generation discriminant term, a real sample gradient penalty term, and a generated sample gradient penalty term, and the formula used is as follows:

[0085] Adversarial generation criteria:

[0086] ;

[0087] in, Represents generator The parameters, Discriminator The parameters, Indicates from potential space random variables sampled in the middle , The generator is based on the parameters. and random variables The output generated samples (pseudo-samples). Indicates the distribution of real data Real samples from the middle, This indicates that the discriminator is effective against real samples. The output, This represents the discriminator's output on the generated samples; Indicates the expected symbol; Indicates the adversarial generation discriminant;

[0088] Adversarial generation discriminant: Based on the relative score between generated samples and real samples, the generator is guided to output more realistic completion results;

[0089] Gradient penalty term for real samples:

[0090] ;

[0091] in, The weights of the gradient penalty term are represented by the coefficients. Indicates the real sample Find the gradient of the discriminator output to measure the smoothness of the discriminant function; The regularization weights represent the consistency constraint terms. Represents the reference mean of the real sample set; The square of the Euclidean norm is used to quantify the degree of sample deviation. This represents the gradient penalty term for the real sample.

[0092] Generate sample gradient penalty term:

[0093] ;

[0094] in, This represents the gradient of the discriminator output over the generated samples; This represents the gradient penalty term for generating samples;

[0095] ;

[0096] in, Represents the joint loss function;

[0097] Step S44: Based on the joint loss function, the Adam optimizer is used to alternately update and train the generator and discriminator to obtain the trained generator model; the optimization objective of the discriminator is to maximize the joint loss function, and the objective of the generator is to minimize the joint loss function.

[0098] Steps S41 to S44 above constitute the training phase; step S45 constitutes the inference phase.

[0099] Figure 2The graph shows the trend of the joint loss function during training. The horizontal axis represents the number of training epochs, and the vertical axis represents the loss value. The black curve represents the joint loss, the blue curve represents the adversarial generation and discrimination term, the green curve represents the gradient penalty term for real samples, and the orange curve represents the gradient penalty term for generated samples. As the number of training epochs increases, the various losses gradually converge, and the generator and discriminator achieve good collaborative optimization under the constraints of the three joint losses.

[0100] Step S45: Enter the inference stage, perform voxel encoding and scale standardization on the missing areas of the building components to obtain sparse three-dimensional tensor data; input the sparse three-dimensional tensor data into the trained generator model, perform three-dimensional morphological completion and reconstruction of the missing areas of the building components, and generate high-fidelity reconstructed component data.

[0101] Example 6, based on Example 4, describes the process of performing three-dimensional morphological reconstruction of missing areas of building components to generate high-fidelity reconstructed component data. The specific steps include:

[0102] Step R1: Construct a complete component sample set based on the non-damaged areas of the building components; construct an initial pseudo-sample set based on the damaged areas of the building components;

[0103] Step R2: Input the initial pseudo-sample set into the generator to generate pseudo-samples for the completion process; input the complete component sample set and the pseudo-samples for the completion process into the discriminator to perform the real / fake discrimination operation, thereby realizing the function of distinguishing between generated samples and real samples;

[0104] Step R3: Construct the standard adversarial training objective function for the generator and discriminator. The standard adversarial training objective function includes the following two parts:

[0105] The discriminator's loss function is used to maximize the accuracy of classifying real samples as "true" and generated samples as "false", specifically the cross-entropy loss function;

[0106] The generator's loss function is used to minimize the probability that the generated sample is judged as "fake," guiding the generator to output more realistic completion process results;

[0107] Step R4: Based on the joint loss function, the Adam optimizer is used to alternately update and train the generator and discriminator to obtain the trained generator model;

[0108] Steps R1 to R4 constitute the training phase; step R5 constitutes the inference phase.

[0109] Step R5: Enter the inference stage, perform voxel encoding and scale standardization on the missing areas of the building components to obtain sparse three-dimensional tensor data; input the sparse three-dimensional tensor data into the trained generator model, perform three-dimensional morphological completion and reconstruction of the missing areas of the building components, and generate high-fidelity reconstructed component data.

[0110] The present invention and its embodiments have been described above. This description is not restrictive. The accompanying drawings are only one embodiment of the present invention, and the actual structure is not limited thereto. In short, if a person skilled in the art is inspired by this description and designs a similar structure and embodiment without departing from the spirit of the present invention, such design should fall within the protection scope of the present invention.

Claims

1. An AI-based digital three-dimensional reconstruction method for historical buildings, characterized by, The method comprises the following steps: Step S1: obtaining multi-source point cloud data and texture image data of a historical building; Step S2: performing fusion processing on the multi-source point cloud data to generate fused point cloud data; Step S3: performing component region identification on the fused point cloud data by using a historical building structure identification model to locate a building component defect region and a building component non-defect region; the historical building structure identification model comprises a RandLA-Net network and a fully connected network; Step S4: performing three-dimensional morphological completion reconstruction on the building component defect region by using a three-loss joint loss guided GAN model to generate high-fidelity reconstructed component data; the three-loss joint loss guided GAN model comprises a generator and a discriminator; the joint loss function in the three-loss joint loss guided GAN model comprises an adversarial generation discrimination term, a real sample gradient penalty term and a generated sample gradient penalty term; Step S5: performing mapping registration on the fused point cloud data, the high-fidelity reconstructed component data and the texture image data to generate a parameterized BIM model, outputting a complete digital building archive file according to the parameterized BIM model to realize fine digital management of the historical building; The process of locating the building component defect region by using the historical building structure identification model comprises the following steps: Step S31: extracting basic feature data and enhanced feature data of the fused point cloud data to combine to form extended feature data input into the fully connected network to obtain 8-dimensional fused feature data; Step S32: inputting the 8-dimensional fused feature data into the RandLA-Net network to perform feature extraction and point-level classification to obtain a building component category probability distribution; Step S33: extracting a maximum category probability label of each point in the building component category probability distribution to generate a same-category point cloud data set; performing spatial clustering processing on the same-category point cloud data set to obtain a spatial clustering result; Step S34: generating a structure component label atlas according to the building component category probability distribution, combining the spatial clustering result to divide a component region, performing defect detection and locating a building component defect region; The enhanced feature data comprises relative height features, point density features, height variance features, flatness features and normal vertical component features; The process of extracting the relative height features in the fused point cloud data comprises the following steps: Step B1: projecting the fused point cloud data in a horizontal direction to construct an initial sparse ground point set; Step B2: based on the initial sparse ground point set, a two-dimensional cubic spline surface fitting method with L1 norm constraint is used to construct a ground height function; Step B3: for each point in the fused point cloud data, the ground height function is called to obtain a relative height feature value of the point, and the relative height feature values of all points in the fused point cloud data are extracted to obtain the relative height features.

2. The AI-based historical building digital three-dimensional reconstruction method according to claim 1, characterized in that: The process of generating the high-fidelity reconstructed component data comprises the following steps: Step S41: constructing a complete component sample set according to the building component non-defect region; constructing an initial pseudo sample set according to the building component defect region; Step S42: input the initial pseudo sample set to the generator to generate a completed process pseudo sample; input the complete component sample set and the completed process pseudo sample to the discriminator to perform a true or false discrimination operation; Step S43: construct a joint loss function to optimize the output of the generator and constrain the gradient variation amplitude of the discriminator; Step S44: based on the joint loss function, the generator and the discriminator are alternately updated and trained by using an Adam optimizer to obtain a trained generator model; Step S45: voxelize and scale the building component damage area to obtain sparse three-dimensional tensor data; input the sparse three-dimensional tensor data into the trained generator model to perform three-dimensional morphological completion reconstruction of the building component damage area and generate high-fidelity reconstructed component data.

3. The AI-based historical building digital three-dimensional reconstruction method according to claim 2, characterized in that: The joint loss function includes an adversarial generation discrimination term, a real sample gradient penalty term and a generated sample gradient penalty term, and the formula is as follows: Adversarial generation discrimination term: ； wherein denotes the parameters of the generator , denotes the parameters of the discriminator , denotes a random variable sampled from the latent space , , denotes a generated sample, i.e. a fake sample, output by the generator according to the parameters and the random variable , denotes a real sample sampled from the real data distribution , denotes the output of the discriminator for the real sample , denotes the output of the discriminator for the generated sample; denotes the expectation symbol; denotes the adversarial generative discriminative term; The adversarial generation discrimination term is based on the relative score between the generated sample and the real sample to guide the generator to output a more realistic completion result; Real sample gradient penalty term: ； wherein, denotes a weight coefficient of the gradient penalty term, denotes a gradient of the discriminator output, solves the gradient of the discriminator output, measures the smoothness of the discriminant function; denotes a regularization weight of the consistency constraint term, denotes a reference mean value of the real sample set; denotes the square of the Euclidean norm, used to quantify the degree of deviation of the sample; denotes the real sample gradient penalty term; Generated sample gradient penalty term: ； wherein, denotes the gradient of the discriminator output with respect to the generated sample; denotes the gradient penalty term for the generated sample. ； wherein, denotes the joint loss function.

Images