Historical building digital three-dimensional reconstruction method based on AI

Abstract

The invention relates to the technical field of artificial intelligence, provides an AI-based historical building digital three-dimensional reconstruction method, and relates to multi-source point cloud and image fusion, component identification, defect detection and three-dimensional completion modeling. According to the method, a component identification model is constructed by adopting RandLA-Net and a full-connection network, and component defect area positioning is realized in combination with enhanced features such as relative elevation and density; the terrain trend is extracted by using two-dimensional spline surface fitting with L1 norm constraint, and the defect recognition precision is improved; a three-item joint loss guided GAN model is constructed, and three-dimensional form completion reconstruction of the building component is completed; and finally, combining the texture image and the point cloud data to generate a parameterized BIM model, and outputting a digital building file. According to the method, fine modeling and automatic reconstruction of a historical building component level are realized, the completeness and accuracy of digital management are improved, and the method is suitable for cultural relic protection, building restoration and heritage digital scenes.

Description

TECHNICAL FIELD

[0001] The application relates to the field of artificial intelligence, in particular to an AI-based historical building digital three-dimensional reconstruction method. BACKGROUND

[0002] With the rapid development of cultural heritage protection and digital modeling technology, the digital processing of historical buildings has become an important technical means for building heritage protection, repair and management. Through high-precision collection and modeling of the geometric shape, structural characteristics and material information of historical buildings, not only can the long-term preservation and research of precious cultural relic buildings be realized, but also reliable data support can be provided for subsequent repair design, structure evaluation and public display. Existing historical building digital methods mainly rely on laser scanning, photogrammetry or BIM modeling technical means, among which laser scanning and unmanned aerial vehicle photogrammetry can quickly obtain large-scale point cloud data and texture image information. However, due to the complexity of building structures, environmental obstruction and historical damage, the obtained data often has the problems of missing, noise and incomplete local structure, which is difficult to directly generate a high-fidelity three-dimensional model. SUMMARY

[0003] The application aims to solve the problems of point cloud missing completion difficulty, low component damage identification accuracy, low three-dimensional modeling quality and insufficient automation level in existing historical building digital processing, and provides an AI-based historical building digital three-dimensional reconstruction method to realize fine, automated and high-fidelity digital modeling of historical buildings. Compared with the prior art, the application first introduces a building structure identification model composed of RandLA-Net and a fully connected network, combines enhanced feature vectors and spatial clustering to realize high-precision component region division and damage positioning; innovatively designs a two-dimensional cubic spline surface fitting method with L1 norm constraint for extracting relative elevation features and enhancing the geometric expression ability of point clouds; in the three-dimensional reconstruction stage, a three-term joint loss based on an adversarial generation discriminator, a real sample gradient penalty term and a generated sample gradient penalty term is constructed to guide the generative adversarial network to complete high-fidelity damage completion, effectively improving the authenticity, robustness and discriminant space response consistency of the reconstructed model; through the above integrated design, the application realizes fine restoration and parameterized BIM model generation of damaged components of historical buildings, significantly improving the intelligent level and automation degree of digital modeling.

[0004] The application provides an AI-based historical building digital three-dimensional reconstruction method, which comprises the following steps:

[0005] Step S1: acquiring multi-source point cloud data and texture image data of the historical building by using a unmanned aerial vehicle device, a high-resolution camera and a three-dimensional laser scanner;

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

[0007] Step S3: identify the component region of the fused point cloud data using a historical building structure identification model to locate the building component defect region and the building component non-defect region; the historical building structure identification model includes a RandLA-Net network and a fully connected network;

[0008] Step S4: use a three-item joint loss guided GAN model to reconstruct the three-dimensional morphology of the building component defect region, and generate high-fidelity reconstructed component data; the three-item joint loss guided GAN model includes a generator and a discriminator;

[0009] Step S5: map and register the fused point cloud data, the high-fidelity reconstructed component data, and the texture image data to generate a parameterized BIM model, output a complete digital building archive file containing geometric morphology information, material attribute information, and historical repair record information from the parameterized BIM model, and realize fine digital management of historical buildings.

[0010] Further, the process of identifying the component region of the fused point cloud data using the historical building structure identification model to locate the building component defect region includes the following steps:

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

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

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

[0014] Step S34: generate a structure component label atlas according to the building component class probability distribution, combine the spatial clustering result to divide the component region, calculate the local point density sparsity index and geometric consistency index of each component region, detect the component region for defects, and locate the building component defect region and the building component non-defect region.

[0015] Further, the process of extracting the relative elevation feature in the fused point cloud data includes the following steps:

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

[0017] 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;

[0018] Step B3: For each point in the fused point cloud data, the ground height function is called to calculate the terrain elevation at the location of the point, and the relative elevation feature value of the point is obtained by subtracting the terrain elevation from the original height of the point. The relative elevation feature extraction of all points in the fused point cloud is completed, and the relative elevation feature is obtained.

[0019] Further, the three-item joint loss is used to guide the GAN model to perform three-dimensional morphological completion reconstruction on the building component missing area, and a process of generating high-fidelity reconstructed component data is generated, which specifically includes the following steps:

[0020] Step S41: Construct a complete component sample set according to the non-missing area of the building component, and construct an initial pseudo sample set according to the missing area of the building component;

[0021] Step S42: input the initial pseudo sample set into the generator to generate a pseudo sample in the completion process; input the complete component sample set and the pseudo sample in the completion process into the discriminator to perform a true or false discrimination operation;

[0022] Step S43: construct a joint loss function to optimize the output of the generator and constrain the gradient change amplitude of the discriminator in the true or false sample space; the joint loss function includes an adversarial generation discrimination term, a real sample gradient penalty term and a generated sample gradient penalty term;

[0023] Step S44: based on the joint loss function, the generator and the discriminator are alternately updated and trained using the Adam optimizer to obtain a trained generator model;

[0024] Step S45: voxelize and scale the missing area of the building component 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 on the missing area of the building component, and generate high-fidelity reconstructed component data.

[0025] By using the above scheme, the application has the following beneficial effects:

[0026] The application fuses unmanned aerial photography, high-resolution image acquisition and laser scanning, constructs a unified data basis of multi-source point clouds and texture images, and uses a registration algorithm to complete high-precision fusion, realizes comprehensive perception and unified expression of complex geometric structures of historical buildings, and provides a stable input source for subsequent component identification and reconstruction; the historical building structure identification model constructed by combining the RandLA-Net network and the full connection network improves the identification accuracy of multi-scale and heterogeneous components and the positioning accuracy of damaged areas, and solves the problem that the traditional rule-based modeling method is difficult to deal with large-scale unstructured point cloud data; the model further combines relative height, point density and normal component and other enhanced features, and extracts the component boundary through a spatial clustering mechanism, enhances the resolution ability of the damaged detection to the fine-grained components, and effectively adapts to the complex structural defects such as local erosion, fracture and shielding in historical buildings.

[0027] In the three-dimensional completion process, the application designs a generative adversarial network (GAN) model based on a three-item joint loss function guide, which significantly improves the reconstruction authenticity and discrimination consistency of the damaged area; by introducing an adversarial generation discrimination term, the generator is guided to output a more realistic component morphology completion process result; by introducing a gradient penalty term of real samples and generated samples, the smoothness and structural stability of the generation process are constrained, and the problems of surface incoherence and morphology drift in traditional reconstruction algorithms are solved; the completion mechanism completes the joint optimization of the generator and the discriminator in the training stage, and directly encodes and scales the damaged area in the inference stage, realizes fast and high-fidelity component completion reconstruction, and enhances the adaptability and automation degree of the reconstruction process.

[0028] Finally, the application unifies the completed component data, the original point cloud and the texture image, and constructs a parameterized BIM model, outputs a complete digital building archive file containing multi-dimensional information such as geometry, material and historical repair records, realizes the whole process of digital management of historical buildings from perception, analysis, reconstruction to structured expression; this method not only improves the quality and precision of digital modeling, but also supports standardized cultural relic filing, digital asset management and multi-platform sharing, solves the problem of lack of structure-level completion and parameter-level expression in traditional methods, and provides an efficient, intelligent and scalable technical path for historical building protection, repair design and heritage information system. BRIEF DESCRIPTION OF DRAWINGS

[0029] Figure 1 The fitting results of the ground height function under different regular term constraints proposed in Example Four are compared in the graph;

[0030] Figure 2 The trend graph of the joint loss function training process proposed in Example Five is shown.

[0031] Figure 1In the middle, (a) only L1 data fidelity term, the surface appears more jitter, poor anti-noise; (b) add the second order smooth term surface becomes continuous smooth, but there are still local mutations; (c) plus the minimum surface area regularization term significantly inhibits the slope of the dramatic change, the fitting surface is continuous, natural, especially suitable for historical buildings such as domes, slope roof and other complex structure of the terrain reconstruction needs; The left side represents the elevation value, and the lower side is the position;

[0032] Figure 2 In the middle, the horizontal axis is the training round (epoch), and the vertical axis is the loss value; The black curve represents the joint loss, the blue curve represents the adversarial generation discriminant term, the green curve represents the real sample gradient penalty term, and the orange curve represents the generated sample gradient penalty term; With the increase of the training round, the loss of each item gradually converges, and the generator and the discriminator have good collaborative optimization effect under the constraint of the three joint losses. DETAILED DESCRIPTION

[0033] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the drawings in the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, not all the embodiments; based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor belong to the scope of protection of the present application.

[0034] Embodiment one, the present application provides a kind of historical building digital three-dimensional reconstruction method based on AI, which comprises the following steps:

[0035] Step S1: use unmanned aerial vehicle equipment, high-resolution camera and three-dimensional laser scanner to obtain multi-source point cloud data and texture image data of historical building;

[0036] For the historical building A of Qing Dynasty brick-wood structure:

[0037] The three-dimensional laser scanner acquires building three-dimensional point cloud data, with a resolution of 5mm, covering an area of about 800m², and a total of about 23 million original points are collected;

[0038] The high-resolution camera collects texture images, a total of 286 images are taken, with a resolution of 8688x5792;

[0039] The unmanned aerial vehicle equipment collects supplementary point cloud data and oblique photography images of high-altitude structures such as roof and eaves, with an accuracy of 10cm;

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

[0041] Step S3: identifying the component region of the fused point cloud data using a historical building structure identification model to locate the building component defect region and the building component non-defect region; the historical building structure identification model comprises a RandLA-Net network and a fully connected network;

[0042] Step S4: performing three-dimensional morphological completion reconstruction on the building component defect region using a three-item joint loss guided GAN model to generate high-fidelity reconstructed component data; the three-item joint loss guided GAN model comprises a generator and a discriminator; the construction process of the three-item joint loss guided GAN model: constructing a GAN model, optimizing the training process of the generator and the discriminator by introducing a three-item joint loss function of an adversarial generation discrimination term, a real sample gradient penalty term and a generated sample gradient penalty term, and constraining the response consistency and gradient smoothness of the generated sample and the real sample in the feature space of the discriminator, thereby obtaining the three-item joint loss guided GAN model;

[0043] Step S5: mapping and registering 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 containing geometric morphological information, material attribute information and historical repair record information from the parameterized BIM model, and realizing fine digital management of the historical building.

[0044] In this embodiment, the process of identifying the component region of the fused point cloud data using a historical building structure identification model to locate the building component defect region comprises the following steps:

[0045] Step S31: extracting the basic feature data and the enhanced feature data of the fused point cloud data, combining to form an extended feature vector inputting a fully connected network, and obtaining 8-dimensional fused feature data through nonlinear projection mapping to an 8-dimensional feature space; the enhanced feature data comprises relative elevation features, point density features, height variance features, flatness features and normal vertical component features;

[0046] Step S32: inputting the 8-dimensional fused feature data into the RandLA-Net network to perform feature extraction and point-level classification through the sparse attention mechanism and the lightweight residual structure thereof, and obtaining a building component class probability distribution;

[0047] Step S33: extracting the maximum class probability label of each point in the building component class probability distribution to generate a same-class point cloud data set; performing spatial clustering processing on the same-class point cloud data set to obtain a spatial clustering result;

[0048] Step S34: generating a structural component label atlas according to the building component category probability distribution, combining the spatial clustering result to divide the component region, calculating the local point density sparsity index and the geometric consistency index of each component region, performing defect detection on the component region, and positioning the building component defect region and the building component non-defect region.

[0049] Embodiment three, based on embodiment one, in this embodiment, the process of component region identification and positioning the building component defect region of the fused point cloud data, specifically includes the following steps:

[0050] Step E1: extracting the basic feature data of the fused point cloud data, and inputting the full connection network to map to an 8-dimensional feature space through nonlinear projection, and obtaining 8-dimensional fused feature data;

[0051] Step E2: inputting the 8-dimensional fused feature data into the RandLA-Net network, and performing feature extraction and point-level classification through the sparse attention mechanism and lightweight residual structure thereof, and obtaining the building component category probability distribution;

[0052] Step E3: extracting the maximum class probability label of each point in the building component category probability distribution, and generating a same-class point cloud data set; performing spatial clustering processing on the same-class point cloud data set to obtain a spatial clustering result;

[0053] Step E4: generating a structural component label atlas according to the building component category probability distribution, combining the spatial clustering result to divide the component region, calculating the local point density sparsity index and the geometric consistency index of each component region, performing defect detection on the component region, and positioning the building component defect region and the building component non-defect region.

[0054] Embodiment four, according to Figure 1 , based on embodiment two, in this embodiment, the process of extracting the relative elevation feature in the fused point cloud data, specifically includes the following steps:

[0055] Step B1: projecting the fused point cloud data to a two-dimensional regular grid in the horizontal direction, in each grid cell, defining a neighborhood region, selecting a ground candidate point in the neighborhood region, and constructing an initial sparse ground point set for reflecting the ground trend of the building region;

[0056] The horizontal projection grid size is set to 0.5m x 0.5m, and a total of 3415 grids are obtained;

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

[0058] ;

[0059] Wherein, represents the two-dimensional coordinates of the ground candidate points; the first elevation value corresponding to the ground candidate point; representing the total number of ground candidate points;

[0060] The point set serves as the control reference for fitting the ground curve;

[0061] 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 the ground height function;

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

[0063] Construct the L1 data fidelity term:

[0064] ;

[0065] wherein, represents the predicted output value of the fitting function at point , representing the fitted terrain elevation corresponding to the plane position, represents the absolute error, using L1 norm to enhance robustness; represents the L1 data fidelity term;

[0066] This term is used to ensure that the fitted surface is as close as possible to the true ground height at the control points, and to suppress the influence of outliers;

[0067] Construct the second-order smooth term (curvature control):

[0068] ;

[0069] wherein, represents the domain, i.e. the entire point cloud data coverage area; represents the horizontal direction coordinate in the point cloud, represents the vertical direction coordinate in the point cloud; represents the second-order derivative of the fitting function in direction; represents the second-order derivative of the fitting function in direction; represents the mixed derivative of and ; represents the double integral; represents the second-order smooth term;

[0070] This term is used to control the bending degree of the surface, avoiding the generation of discontinuous or jagged boundaries;

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

[0072] ;

[0073] where, represents the gradient vector of ; represents the square of the gradient modulus; represents the local area element of the surface (derived from the minimum surface theory); represents the minimum surface regular term;

[0074] This term is used to suppress sharp slope changes and ensure surface continuity and physical geometric stability, especially suitable for fitting slope roofs and dome structures in historical buildings;

[0075] Constructing the overall optimization objective function:

[0076] ;

[0077] where, represents the fitted ground height function, represents the balance coefficient, used to adjust the relative weight of fidelity and smoothness; represents the weighted coefficient of the area regular term, controlling the constraint strength on the slope change of the surface; represents the fitting function Solving the solution that makes the objective function minimum, used to construct the optimal terrain elevation function under the constraint of sparse ground points;

[0078] The ground height function constructed by the above formula realizes adaptive constraints on surface smoothness, geometric continuity, and slope change while maintaining fitting accuracy, and is suitable for fine processing of complex terrain and historical building point clouds. This function can be used in subsequent applications such as relative elevation feature calculation and building component defect positioning;

[0079] Figure 1 The figure shows the fitting results of the ground height function under different regular term constraints, where (a) only the L1 data fidelity term, the surface has more jitter and poor noise resistance; (b) adding a second-order smoothness term, the surface becomes continuous and smooth, but there are still local mutations; (c) adding the minimum surface area regular term significantly suppresses the sharp slope change, the fitted surface is continuous and natural, especially suitable for terrain reconstruction needs of complex structures such as domes and slope roofs in historical buildings; the left side represents the elevation value, and the lower side represents the position;

[0080] Step B3: For each point in the fused point cloud data, call the ground height function to calculate the terrain elevation at the location of the point, and obtain the relative elevation feature value of the point by subtracting the terrain elevation from the original height of the point. The relative elevation feature extraction of all points in the fused point cloud is completed, and the relative elevation feature is obtained.

[0081] In an embodiment, the method comprises the following steps: Figure 2 In this embodiment, based on the fourth embodiment, the process of generating high-fidelity reconstructed component data by using a three-loss joint loss guided GAN model to complete and reconstruct the three-dimensional morphology of the damaged area of the building component includes the following steps:

[0082] Step S41: Construct a complete component sample set according to the non-damaged area of the building component, and construct an initial pseudo sample set according to the damaged area of the building component;

[0083] Step S42: input the initial pseudo sample set into the generator, and the generator uses a residual structure composed of bilinear upsampling, 1x1 convolution and 3x3 group convolution to reconstruct the features of the input data, and generates a pseudo sample in the completion process; input the complete component sample set and the pseudo sample in the completion process into the discriminator, and the discriminator uses a shallow full convolution structure, and the output end is provided with a 4x4 deep convolution module and a linear mapping module, which are used for feature compression of the input sample, performing true and false discrimination operation, and realizing the distinguishing function of the generated sample and the real sample;

[0084] Step S43: construct a joint loss function to optimize the output of the generator and constrain the gradient change amplitude of the discriminator in the true and false sample space; 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:

[0085] Adversarial generation discrimination term:

[0086] ;

[0087] wherein, represents the parameters of the generator , represents the parameters of the discriminator , represents a random variable sampled from the latent space , represents the generated sample (pseudo sample) output by the generator according to the parameters and the random variable , represents a real sample sampled from the real data distribution , represents the output of the discriminator to the real sample , represents the output of the discriminator on the generated sample; represents the expected symbol; represents the adversarial generated discriminant term;

[0088] Adversarial generated discriminant term: based on the relative score between the generated sample and the real sample, guide the generator to output the completion result with higher authenticity;

[0089] Real sample gradient penalty term:

[0090] ;

[0091] wherein, represents the weight coefficient of the gradient penalty term, represents the gradient of the discriminator output on the real sample , measures the smoothness of the discriminant function; represents the regularization weight of the consistency constraint term, represents the reference mean value of the real sample set; represents the square of the Euclidean norm, used to quantify the deviation degree of the sample; represents the real sample gradient penalty term;

[0092] Generated sample gradient penalty term:

[0093] ;

[0094] wherein, represents the gradient of the discriminator output on the generated sample; represents the generated sample gradient penalty term;

[0095] ;

[0096] wherein, represents the joint loss function;

[0097] Step S44: based on the joint loss function, the generator and the discriminator are alternately updated and trained by using the Adam optimizer, to obtain the trained generator model; the optimization target of the discriminator is to maximize the joint loss function, and the target of the generator is to minimize the joint loss function;

[0098] The above steps S41 to S44 are the training phase; step S45 is the inference phase;

[0099] Figure 2The trend chart of the joint loss function training process is shown, the horizontal axis is the training round (epoch), and the vertical axis is the loss value; the black curve represents the joint loss, the blue curve represents the adversarial generation discrimination term, the green curve represents the real sample gradient penalty term, and the orange curve represents the generated sample gradient penalty term; with the increase of the training round, the loss gradually converges, and the generator and the discriminator are well optimized under the constraint of the three joint losses;

[0100] Step S45: entering the inference stage, voxelizing and coding the building component defect area and performing scale standardization to obtain sparse three-dimensional tensor data; inputting the sparse three-dimensional tensor data into the trained generator model to perform three-dimensional morphological completion reconstruction of the building component defect area and generate high-fidelity reconstructed component data.

[0101] In this embodiment, the process of performing three-dimensional morphological completion reconstruction on the building component defect area to generate high-fidelity reconstructed component data includes the following steps:

[0102] Step R1: constructing a complete component sample set according to the non-defect area of the building component; constructing an initial pseudo sample set according to the defect area of the building component;

[0103] Step R2: inputting the initial pseudo sample set into the generator to generate a pseudo sample in the completion process; inputting the complete component sample set and the pseudo sample in the completion process into the discriminator to perform a true-false discrimination operation and realize the differentiation function of the generated sample and the real sample;

[0104] Step R3: constructing a standard adversarial training target function of the generator and the discriminator, which includes the following two parts:

[0105] The loss function of the discriminator is used to maximize the accuracy of discriminating the real sample as "true" and the generated sample as "false", which is a cross-entropy loss function;

[0106] The loss function of the generator is used to minimize the probability of the generated sample being discriminated as "false" to guide the generator to output more realistic completion process results;

[0107] Step R4: based on the joint loss function, the generator and the discriminator are alternately updated and trained using the Adam optimizer to obtain a trained generator model;

[0108] The above steps R1 to R4 are the training stage; and step R5 is the inference stage;

[0109] Step R5: entering the reasoning stage, the building component defect area is voxelized and coded and scaled, to obtain sparse three-dimensional tensor data; the sparse three-dimensional tensor data is input into the trained generator model, and three-dimensional morphological completion reconstruction of the building component defect area is performed, to generate high-fidelity reconstructed component data.

[0110] The above describes the present application and its embodiments, which are not restrictive, and the drawings only show one of the embodiments of the present application, and the actual structure is not limited thereto; in general, if a person skilled in the art is inspired thereby, without departing from the purpose of the present application, without creative design, similar structure and embodiments to the technical solution, which should belong to the protection scope of the present application.

Claims

1. An AI-based 3D reconstruction method for historical buildings, characterized by: The following steps are involved: Step S1: Acquire multi-source point cloud data and texture image data of historical buildings; Step S2: performing fusion processing on multi-source point cloud data to generate fused point cloud data; Step S3: Using a historical building structure recognition model to identify component areas on the fused point cloud data, locating damaged and non-damaged building component areas; the historical building structure recognition model includes a RandLA-Net network and a fully connected network; Step S4: Using the three-way joint loss to guide the GAN model to perform 3D morphological completion and reconstruction of the defective area of ​​the building component to generate high-fidelity reconstructed component data; the three-way joint loss guided GAN model includes a generator and a discriminator; Step S5: Map and align the fused point cloud data, high-fidelity reconstructed component data, and texture image data to generate a parametric BIM model. Based on the output of the parametric BIM model, a complete digital building archive file is obtained to achieve refined digital management of historical buildings.

2. The AI-based 3D reconstruction method for historical buildings according to claim 1, characterized in that: The process of locating the damaged areas of building components using the historical building structure identification model includes the following steps: Step S31: extracting basic feature data and enhanced feature data from the fused point cloud data, combining them to form extended feature data, and inputting them into a fully connected network to obtain 8-dimensional fused feature data; Step S32: Input the 8-dimensional fused feature data into the RandLA-Net network to perform feature extraction and point-level classification to obtain the probability distribution of building component categories; Step S33: extracting the maximum category probability label of each point in the building component category probability distribution to generate a point cloud data set of the same category; performing spatial clustering processing on the point cloud data set of the same category to obtain a spatial clustering result; Step S34: Generate a structural component label map based on the probability distribution of building component categories, divide the component area based on the spatial clustering results, perform defect detection, and locate the building component defect area.

3. The AI-based 3D reconstruction method for historical buildings according to claim 2, characterized in that: The enhanced feature data includes relative elevation features, point density features, height variance features, flatness features and normal vertical component features.

4. The AI-based 3D reconstruction method for historical buildings according to claim 3, characterized in that: The process of extracting relative elevation features from fused point cloud data includes the following steps: Step B1: Project the fused point cloud data horizontally 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 the ground height function; Step B3: For each point in the fused point cloud data, call the ground height function to obtain the relative elevation feature value of the point, complete the relative elevation feature value extraction of all points in the fused point cloud data, and obtain the relative elevation feature.

5. The AI-based 3D reconstruction method for historical buildings according to claim 1, characterized in that: The process of generating high-fidelity reconstruction component data includes the following steps: Step S41: constructing a complete component sample set based on the non-defective areas of the building components; constructing an initial pseudo sample set based on the defective areas of the building components; Step S42: Input the initial pseudo sample set into the generator to generate the pseudo samples of the completion process; input the complete component sample set and the pseudo samples of the completion process into the discriminator to perform the authenticity discrimination operation; Step S43: Construct a joint loss function to optimize the output of the generator and constrain the gradient change amplitude of the discriminator; Step S44: Based on the joint loss function, the Adam optimizer is used to alternately update the generator and the discriminator to obtain a trained generator model; Step S45: Voxel encoding and scale normalization are performed on the defective area of ​​the building component to obtain sparse three-dimensional tensor data; the sparse three-dimensional tensor data is input into the trained generator model to perform three-dimensional morphological completion and reconstruction of the defective area of ​​the building component to generate high-fidelity reconstructed component data.

6. The AI-based 3D reconstruction method for historical buildings according to claim 5, characterized in that: The joint loss function includes the adversarial generation discriminant term, the real sample gradient penalty term and the generated sample gradient penalty term.

Images