Digital auxiliary repairing system for historical building

Abstract

The invention relates to a digital auxiliary repairing system for a historical building. The system comprises a building data acquisition module, a data primary processing module, an automatic damage evaluation module, a repairing auxiliary model construction module and a historical building repairing auxiliary module. According to the invention, original data is obtained through data acquisition; a primary processing method of multi-source data alignment, building value evaluation, point cloud feature extraction and damage pre-labeling is adopted; an improved clustering model is adopted to carry out automatic damage assessment, a self-adaptive clustering mechanism is introduced, and a scientific basis is provided for repair decision making in combination with the historical value of the building; a repair decision transformer model is adopted as a repair auxiliary model, cultural relic protection specifications are converted into prior constraints, polar coordinate position coding is combined, perception of a building symmetric structure is enhanced, and finally collaborative optimization of a three-dimensional target is achieved.

Description

Technical Field

[0001] The present invention relates to the technical field of digital assisted restoration of historical buildings, and in particular to a digital assisted restoration system for historical buildings. Background Art

[0002] Digitally assisted restoration of historic buildings uses modern digital technology to comprehensively record, analyze, and design restorations of historical buildings. This allows for a precise understanding of the building's current condition, assessment of structural and material damage, and the development of scientifically sound restoration plans. This technology not only preserves the building's historical value and cultural heritage, but also minimizes damage to the building caused by human intervention, improving the accuracy and efficiency of restoration efforts.

[0003] However, the traditional digital-assisted repair system for historical buildings has the problems of low efficiency of manual visual inspection and is constrained by subjective experience, making it difficult to systematically quantify the degree of damage. Conventional algorithms cannot effectively integrate multi-source heterogeneous data such as infrared thermal imaging and geometric point clouds, resulting in the lack of correlation analysis between internal defects and surface damage. In addition, the ability to conduct detailed assessments of high-value areas is insufficient, and it is easy to ignore hidden damage or over-intervene in low-value areas, affecting the accuracy of repair plans. The traditional digital-assisted repair system for historical buildings has the problem of weak relationship modeling of architectural spatial structures, ignoring the topological correlation of damaged areas, and often evaluating multiple objectives such as cost, technology, and value protection separately. It is impossible to quantify the impact of dynamic risks on decision variables, which can easily lead to resource allocation imbalance or excessive repairs. In addition, the traditional model lacks a mechanism for embedding domain knowledge, making it difficult to ensure that repair measures comply with cultural relics protection standards. Summary of the Invention

[0004] In view of the above situation, in order to overcome the defects of the prior art, the present invention provides a digital assisted repair system for historical buildings. The digital assisted repair system for traditional historical buildings has the problems of low efficiency of manual visual inspection and is restricted by subjective experience, making it difficult to systematically quantify the degree of damage. Conventional algorithms cannot effectively integrate multi-source heterogeneous data such as infrared thermal imaging and geometric point clouds, resulting in the lack of correlation analysis between internal defects and surface damage, and insufficient refined assessment capabilities for high-value areas, which easily leads to the neglect of hidden damage or excessive intervention in low-value areas, affecting the accuracy of repair plans. This solution creatively adopts an automatic damage assessment module, constructs an improved clustering model, and performs automatic damage assessment. It captures damage representations from micro texture to macro structure through multi-scale fusion, and dynamically adjusts recognition sensitivity in combination with the quantitative weight of the building's historical value. At the same time, it introduces an adaptive clustering mechanism to optimize regional segmentation based on the building's value density, which provides a basis for subsequent repair decisions. The digital assisted repair system for traditional historical buildings has a weak ability to model the relationship between the spatial structure of the building, ignores the topological correlation of the damaged area, and often evaluates multiple objectives such as cost, technology, and value protection separately. It is impossible to quantify the impact of dynamic risks on decision variables, which can easily lead to imbalanced resource allocation or excessive repairs. In addition, the traditional model lacks a domain knowledge embedding mechanism, and it is difficult to ensure that the repair measures meet the technical problems of cultural relics protection standards. This solution creatively adopts the repair decision transformer model as a repair auxiliary model. Through multimodal damage fusion and domain knowledge graph, the cultural relics protection standards are transformed into prior constraints to ensure that the decision complies with the regulations. Combined with polar coordinate position coding, the geometric topological relationship of the building is accurately modeled to enhance the perception of the symmetrical structure of the building. At the same time, the risk perception mechanism is used to dynamically associate the damage severity with the decision output to achieve the coordinated optimization of the three-dimensional goals of repair technology, cost control and value retention.

[0005] The technical solution adopted by the present invention is as follows: the present invention provides a digital auxiliary repair system for historical buildings, including a building data acquisition module, a preliminary data processing module, an automatic damage assessment module, a repair auxiliary model construction module and a historical building repair auxiliary module;

[0006] The building data acquisition module obtains a building repair original data set by performing data acquisition, and sends the building repair original data set to the data preliminary processing module;

[0007] The data preliminary processing module uses a data preliminary processing method of multi-source data alignment, building value assessment, point cloud feature extraction and damage pre-labeling to obtain a historical preliminary data set and a current preliminary data set, and sends the historical preliminary data set and the current preliminary data set to the automatic damage assessment module;

[0008] The automatic damage assessment module is used to automatically assess the damage status of historical buildings by constructing an improved clustering model to perform automatic damage assessment, obtain a dataset to be processed, an auxiliary repair training set, and an auxiliary repair test set, and send the dataset to be processed to the historical building repair auxiliary module, and send the auxiliary repair training set and the auxiliary repair test set to the repair auxiliary model construction module;

[0009] The repair auxiliary model construction module constructs a repair decision transformer model as a repair auxiliary model, and sends the repair auxiliary model to the repair auxiliary model construction module;

[0010] The historical building repair auxiliary module specifically assists in the repair of historical buildings by adopting the repair auxiliary model to obtain a reference result for building repair decision-making.

[0011] Furthermore, in the building data acquisition module, the building repair original data set specifically includes a building repair history original data set and a building repair current original data set. Both the building repair history original data set and the building repair current original data set include building optical image data, building infrared image data, building basic information data, building material data and building 3D point cloud data. The building repair history original data set also includes annotation data.

[0012] Furthermore, in the data preliminary processing module, the multi-source data alignment is used to unify the data space coordinate system, specifically by pixel-level alignment of the building optical image data, the building infrared image data and the building 3D point cloud data through feature point matching;

[0013] The building value assessment is used to quantify the historical value of a building. Specifically, the building construction year data, building cultural relic protection level data, and building process type data are converted into weights through preset rules combined with a multi-layer perceptron to construct a building value weight matrix;

[0014] The point cloud feature extraction is used to characterize the geometric characteristics of the building surface. Specifically, the building 3D point cloud data is processed through normal vector calculation and depth projection to generate point cloud geometric features that are aligned with the building optical image data and the building infrared image data;

[0015] The damage pre-labeling is used to pre-label the building damage level. Specifically, a binary mask of building damage is generated by analyzing the differences between the building optical image data and the building infrared image data. The binary mask of building damage is then manually corrected and the damage level is divided to obtain a building damage level label.

[0016] Through the multi-source data alignment, the building value assessment, the point cloud feature extraction and the damage pre-labeling, the historical original data set of the building repair and the current original data set of the building repair are preliminarily processed to obtain a historical preliminary data set and a current preliminary data set.

[0017] Furthermore, the automatic damage assessment module is used to automatically identify and assess damaged areas of historical buildings. Specifically, it extracts multi-scale damage features through analysis, integrates the building value weight matrix, and constructs an improved clustering model for processing to obtain a spatial heat map of damage levels.

[0018] The automatic damage assessment module specifically includes multi-scale feature extraction, feature enhancement, improved clustering model design, damage heat map generation and building damage assessment;

[0019] The multi-scale feature extraction is used to capture damage features of different sizes. Specifically, the multi-scale feature extraction is performed by analyzing the fusion information of architectural optical image data and architectural infrared image data through a convolutional neural network, extracting cross-scale feature maps from fine-grained textures to macroscopic structures, and obtaining cross-scale features, including:

[0020] Pyramid convolution is used to extract scale-sensitive features. Specifically, it fuses architectural optical features and architectural infrared features through channel splicing operations, and performs multi-granularity feature extraction on the fused features using convolution kernels of three different sizes to obtain multi-scale features.

[0021] Cross-scale feature fusion is used to integrate multi-granularity information. Specifically, small-scale features are fused with large-scale features through upsampling and weighted addition operations to obtain cross-scale features;

[0022] The feature enhancement is used to integrate the historical value and damage characteristics of the building. Specifically, the building value weight is expanded to the same dimension as the cross-scale feature through a tensor broadcast operation, and multiplied element-by-element with the cross-scale feature to obtain the value enhancement feature;

[0023] The improved clustering model design is used for damage area segmentation. Specifically, it realizes adaptive division of damage areas through superpixel segmentation and dynamic density clustering, and obtains pixel-level damage level labels, including:

[0024] Superpixel segmentation is used to reduce computational complexity. Specifically, a simple linear iterative clustering algorithm is used to segment the input image data into multiple homogeneous texture regions. Each homogeneous texture region is represented by a superpixel unit, and a superpixel unit set is obtained.

[0025] Regional feature aggregation is used to extract representative features of each homogeneous texture region. Specifically, the average value of the value enhancement feature in each superpixel unit is calculated through the average pooling operation to obtain the representative features of the superpixel unit.

[0026] Dynamic density clustering, for adaptive region division, specifically adjusting the search radius of DBSCAN clustering by building value weight, clustering superpixel units based on representative features, and obtaining damage region clustering labels;

[0027] Damage level mapping, for generating automatic damage assessment results, specifically mapping damage region clustering labels to five-level building damage level labels through a small number of labeled samples to obtain pixel-level damage level labels;

[0028] The damage heat map generation is used to visualize the automatic damage assessment results, specifically by color mapping pixel-level damage level labels and building value weight to the original input image data to obtain a damage level spatial heat map;

[0029] The building damage assessment, specifically, the historical preliminary data set and the current preliminary data set are used as the input of the automatic damage assessment module, and the multi-scale feature extraction, the feature enhancement, the improved clustering model design and the damage heat map generation are used to perform building damage automatic assessment, to obtain the to-be-processed data set and the historical evaluation data set, and to perform data set segmentation on the historical evaluation data set to obtain the auxiliary repair training set and the auxiliary repair test set.

[0030] Further, in the repair assistance model construction module, a model required for digital auxiliary repair of historical buildings is constructed, specifically a repair decision transformer model is constructed as a repair assistance model;

[0031] The repair assistance model construction module includes multi-modal feature fusion, knowledge prior attention design, polar coordinate encoder design, building repair decision generation, and model construction and training;

[0032] The multi-modal feature fusion is used to integrate multi-source repair basis, specifically, the damage level spatial heat map, building material features and point cloud geometric features are aligned and fused through feature splicing and 1x1 convolution operation, and then compressed to obtain multi-modal fusion features;

[0033] The knowledge prior attention design is used to inject domain prior knowledge, specifically, the pre-defined ancient building knowledge graph and the multi-modal fusion features are used for attention interaction to obtain knowledge enhanced features, the pre-defined ancient building knowledge graph is used to convert known ancient building repair experience into machine understandable knowledge, specifically, a knowledge graph constructed based on ancient building related technical regulations and management methods, the knowledge prior attention design includes:

[0034] Knowledge graph embedding is used to introduce the experience of ancient building restoration. Specifically, the predefined ancient building knowledge graph is processed through a graph neural network to obtain ancient building knowledge embedding.

[0035] Attention interaction is used for knowledge-guided feature correction. Specifically, the correlation between the ancient building knowledge embedding and the multimodal fusion features is calculated through the attention mechanism to obtain knowledge-enhanced features.

[0036] The polar coordinate encoder design is used to model building spatial relationships. Specifically, it captures the topological associations of building damage areas through rotationally symmetric polar coordinate position encoding and a six-layer Transformer architecture to obtain deep building features, including:

[0037] Polar coordinate position encoding is used to enhance geometric perception. Specifically, the standard position encoding is replaced by a polar coordinate function to obtain a rotationally symmetric position embedding.

[0038] Deep feature extraction, used to extract deep historical building damage information, specifically by processing rotationally symmetric position embeddings through a six-layer Transformer architecture to obtain deep building features;

[0039] The building repair decision generation is used to generate quantifiable decisions. Specifically, it outputs building repair decision prediction results through damage risk quantification and multi-objective prediction, including:

[0040] Damage risk quantification is used to calculate the degree of damage risk. Specifically, the damage risk coefficient is obtained by weighting the proportion of pixels with different damage levels.

[0041] Multi-objective prediction is used for preliminary decision-making predictions. Specifically, it uses three independent output heads to predict the type of repair technology, the level of repair cost, and the degree of preservation of the value of architectural relics;

[0042] Risk-aware decision correction is used to integrate the damage risk coefficient. Specifically, it dynamically adjusts the output of three decision dimensions: the type of repair technology, the level of repair cost, and the degree of preservation of architectural cultural relic value through the damage risk coefficient to obtain the building repair decision prediction result;

[0043] The constructing and training of the model specifically involves constructing a repair decision transformer model through the multimodal feature fusion, the knowledge prior attention design, the polar coordinate encoder design, and the building repair decision generation. The model is trained based on the auxiliary repair training set and the auxiliary repair test set and the model performance is verified to obtain a repair decision transformer model as a repair auxiliary model.

[0044] Furthermore, in the historical building repair auxiliary module, the data set to be processed is used as the input of the repair auxiliary model to obtain a building repair decision reference result, and based on the building repair decision reference result, the digital repair of historical buildings is assisted.

[0045] The beneficial effects achieved by the present invention using the above scheme are as follows:

[0046] (1) The digital assisted repair system for traditional historical buildings has the problem that manual visual inspection is inefficient and restricted by subjective experience, making it difficult to systematically quantify the degree of damage. Conventional algorithms cannot effectively integrate multi-source heterogeneous data such as infrared thermal imaging and geometric point clouds, resulting in the lack of correlation analysis between internal defects and surface damage. In addition, the ability to conduct detailed assessment of high-value areas is insufficient, and it is easy to ignore hidden damage or over-intervene in low-value areas, affecting the accuracy of the repair plan. This solution creatively adopts an automatic damage assessment module, constructs an improved clustering model, and performs automatic damage assessment. It captures damage representations from micro-texture to macro-structure through multi-scale fusion, dynamically adjusts the recognition sensitivity based on the quantitative weight of the building's historical value, and introduces an adaptive clustering mechanism to optimize regional segmentation based on the building's value density, providing an objective and quantifiable scientific basis for subsequent repair decisions.

[0047] (2) The digital assisted repair system for traditional historical buildings has a weak ability to model the relationship between the spatial structure of the building and ignores the topological correlation of the damaged area. Multiple objectives such as cost, technology, and value protection are often evaluated separately, and the impact of dynamic risks on decision variables cannot be quantified, which easily leads to imbalanced resource allocation or excessive repairs. In addition, the traditional model lacks a domain knowledge embedding mechanism, making it difficult to ensure that the repair measures meet the technical problems of cultural relics protection standards. This solution creatively adopts the repair decision transformer model as a repair auxiliary model. Through multimodal damage fusion and domain knowledge graph, the cultural relics protection standards are converted into prior constraints to ensure that the decision complies with the regulations. Combined with polar coordinate position coding, the geometric topological relationship of the building is accurately modeled to enhance the perception of the symmetrical structure of the building. At the same time, the risk perception mechanism is used to dynamically associate the damage severity with the decision output to achieve the coordinated optimization of the three-dimensional goals of repair technology, cost control and value preservation. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 A schematic diagram of a module of a digital assisted restoration system for historical buildings provided by the present invention;

[0049] Figure 2 This is a flow chart of the data preliminary processing module;

[0050] Figure 3 This is a flowchart of the automatic damage assessment module;

[0051] Figure 4 Schematic diagram of the process of building modules for the repair assistance model.

[0052] The accompanying drawings are used to provide further understanding of the present invention and constitute a part of the specification. They are used to explain the present invention together with the embodiments of the present invention and do not constitute a limitation of the present invention. DETAILED DESCRIPTION

[0053] The technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments; based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0054] In the description of the present invention, it should be understood that terms such as "upper", "lower", "front", "back", "left", "right", "top", "bottom", "inside" and "outside" indicating directions or positional relationships are based on the directions or positional relationships shown in the accompanying drawings. They are only for the convenience of describing the present invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific direction, be constructed and operated in a specific direction. Therefore, they should not be understood as limiting the present invention.

[0055] Example 1, see Figure 1 The present invention provides a digital auxiliary repair system for historical buildings, which includes a building data acquisition module, a preliminary data processing module, an automatic damage assessment module, a repair auxiliary model construction module and a historical building repair auxiliary module;

[0056] The building data acquisition module obtains a building repair original data set by performing data acquisition, and sends the building repair original data set to the data preliminary processing module;

[0057] The data preliminary processing module uses a data preliminary processing method of multi-source data alignment, building value assessment, point cloud feature extraction and damage pre-labeling to obtain a historical preliminary data set and a current preliminary data set, and sends the historical preliminary data set and the current preliminary data set to the automatic damage assessment module;

[0058] The automatic damage assessment module is used to automatically assess the damage status of historical buildings by constructing an improved clustering model to perform automatic damage assessment, obtain a dataset to be processed, an auxiliary repair training set, and an auxiliary repair test set, and send the dataset to be processed to the historical building repair auxiliary module, and send the auxiliary repair training set and the auxiliary repair test set to the repair auxiliary model construction module;

[0059] The repair auxiliary model construction module constructs a repair decision transformer model as a repair auxiliary model, and sends the repair auxiliary model to the repair auxiliary model construction module;

[0060] The historical building repair auxiliary module specifically assists in the repair of historical buildings by adopting the repair auxiliary model to obtain a reference result for building repair decision-making.

[0061] Example 2, see Figure 1 This embodiment is based on the above embodiment. In the building data acquisition module, the building repair original data set specifically includes a building repair history original data set and a building repair current original data set. Both the building repair history original data set and the building repair current original data set include building optical image data, building infrared image data, building basic information data, building material data, and building 3D point cloud data. The building repair history original data set also includes annotation data.

[0062] The architectural optical image data are specifically RGB image data covering the building facade and structural details taken under different lighting conditions and at different angles; the architectural infrared image data are specifically architectural infrared image data that can reflect the overall temperature distribution differences of the building; the architectural basic information data specifically include the building construction year data, the building cultural relic protection level data and the type of building technology used data; the annotated data specifically include building damage level labels and building repair decision labels. The building damage level labels specifically include no damage, slight damage, moderate damage, severe damage and complete destruction. The building repair decision labels specifically include three dimensions, namely, the repair technology type label, the repair cost level label and the building cultural relic value preservation degree label.

[0063] Example 3, see Figure 1 and Figure 2 This embodiment is based on the above embodiment. In the data preliminary processing module, the multi-source data alignment is used to unify the data space coordinate system, specifically by pixel-level alignment of the building optical image data, the building infrared image data and the building 3D point cloud data through feature point matching;

[0064] The building value assessment is used to quantify the historical value of a building. Specifically, the building construction year data, building cultural relic protection level data, and building process type data are converted into weights through preset rules combined with a multi-layer perceptron to construct a building value weight matrix;

[0065] The point cloud feature extraction is used to characterize the geometric characteristics of the building surface. Specifically, the building 3D point cloud data is processed through normal vector calculation and depth projection to generate point cloud geometric features that are aligned with the building optical image data and the building infrared image data;

[0066] The damage pre-labeling is used to pre-label the building damage level. Specifically, a binary mask of building damage is generated by analyzing the differences between the building optical image data and the building infrared image data. The binary mask of building damage is then manually corrected and the damage level is divided to obtain a building damage level label. The damage pre-labeling includes fusion mask generation, morphological expansion processing, and manual correction and damage level division.

[0067] The fusion mask generation is used to detect abnormal areas in infrared images and RGB images. Specifically, a temperature difference mask and a color difference mask are generated and combined through a logical AND operation to obtain a fusion mask. The formula used is as follows:

[0068] ;

[0069] Where Mt represents the temperature difference mask, Mc represents the color difference mask, Represents the value of the building infrared image at coordinates (a, b), represents the pixel mean of building infrared image, represents the standard deviation of building infrared image pixels, Represents the value of the architectural optical image at coordinates (a, b), Represents the reference pixel value of the building health area, represents the preset color difference threshold, MF represents the fusion mask, Indicates the calculation of L2 norm, Represents logical AND operation;

[0070] The morphological dilation process is used to optimize the continuity of the damaged area. Specifically, a 3×3 circular structuring element is used to dilate the fused mask to obtain a binary mask of the building damage.

[0071] Through the multi-source data alignment, the building value assessment, the point cloud feature extraction and the damage pre-labeling, the historical original data set of the building repair and the current original data set of the building repair are preliminarily processed to obtain a historical preliminary data set and a current preliminary data set.

[0072] Example 4, see Figure 1 and Figure 3 This embodiment, based on the above embodiment, is used in the automatic damage assessment module to automatically identify and assess damaged areas of historical buildings. Specifically, it extracts multi-scale damage features through analysis, integrates the building value weight matrix, and constructs an improved clustering model for processing to obtain a damage level spatial heat map;

[0073] The automatic damage assessment module specifically includes multi-scale feature extraction, feature enhancement, improved clustering model design, damage heat map generation and building damage assessment;

[0074] The multi-scale feature extraction is used to capture damage features of different sizes. Specifically, the multi-scale feature extraction is performed by analyzing the fusion information of architectural optical image data and architectural infrared image data through a convolutional neural network, extracting cross-scale feature maps from fine-grained textures to macroscopic structures, and obtaining cross-scale features, including:

[0075] Pyramid convolution is used to extract scale-sensitive features. Specifically, it fuses architectural optical features and architectural infrared features through channel splicing operations. It then uses three convolution kernels of different sizes to perform multi-granularity feature extraction on the fused features to obtain multi-scale features. The formula used is as follows:

[0076] ;

[0077] Where, represents the fusion feature, Represents the splicing operation function, Indicates the infrared characteristics of the building, Represents the optical characteristics of the building, Represents multi-scale features of size sc×sc, represents the two-dimensional convolution function;

[0078] Cross-scale feature fusion is used to integrate multi-granularity information. Specifically, small-scale features are fused with large-scale features through upsampling and weighted addition operations to obtain cross-scale features. The formula used is as follows:

[0079] ;

[0080] Where, represents the upsampled features of multi-scale features with a size of 64×64, represents the upsampled features of multi-scale features with a size of 128×128, represents the upsampling function, Represents multi-scale features of size 64×64, Represents multi-scale features of size 128×128, Represents cross-scale features, represents the ReLU activation function, represents the first cross-scale fusion weight, represents the second cross-scale fusion weight, represents the third cross-scale fusion weight;

[0081] The feature enhancement is used to integrate the historical value and damage characteristics of the building. Specifically, the building value weight is expanded to the same dimension as the cross-scale feature through a tensor broadcast operation, and multiplied element-by-element with the cross-scale feature to obtain the value enhancement feature. The formula used is as follows:

[0082] ;

[0083] Where, Indicates value-enhancing characteristics, Represents the dimension expansion operation function, represents the building value weight, Represents element-wise multiplication operation;

[0084] The improved clustering model design is used for damage area segmentation. Specifically, it realizes adaptive division of damage areas through superpixel segmentation and dynamic density clustering, and obtains pixel-level damage level labels, including:

[0085] Superpixel segmentation is used to reduce computational complexity. Specifically, a simple linear iterative clustering algorithm is used to segment the input image data into multiple homogeneous texture regions. Each homogeneous texture region is represented by a superpixel unit, and a superpixel unit set is obtained.

[0086] Regional feature aggregation is used to extract representative features of each homogeneous texture region. Specifically, the average value of the value enhancement feature in each superpixel unit is calculated through the average pooling operation to obtain the representative features of the superpixel unit.

[0087] Dynamic density clustering is used for adaptive region segmentation. Specifically, the search radius of DBSCAN clustering is dynamically adjusted by the building value weight. Superpixel units are clustered based on their representative features to obtain cluster labels for damaged areas. The search radius is dynamically adjusted using the following formula:

[0088] ;

[0089] Where, represents the search radius of the kth superpixel unit, represents the basic search radius, represents the weighted mean of the building value calculated by the average pooling operation within the k-th superpixel unit;

[0090] Damage level mapping is used to generate automatic damage assessment results. Specifically, it maps the damaged area cluster labels to five-level building damage level labels using a small number of annotated samples, obtaining pixel-level damage level labels.

[0091] The damage heat map is generated to visualize the automatic damage assessment results. Specifically, pixel-level damage level labels and building value weights are superimposed on the original input image data through color mapping to obtain a damage level spatial heat map;

[0092] The building damage assessment specifically uses the historical preliminary data set and the current preliminary data set as inputs of the automatic damage assessment module, performs automatic building damage assessment through the multi-scale feature extraction, the feature enhancement, the improved clustering model design and the damage heat map generation, obtains a data set to be processed and a historical assessment data set, and performs data set segmentation on the historical assessment data set to obtain an auxiliary repair training set and an auxiliary repair test set.

[0093] By performing the above operations, the digital assisted repair system of traditional historical buildings has the problems of low efficiency of manual visual inspection and is restricted by subjective experience, making it difficult to systematically quantify the degree of damage. Conventional algorithms cannot effectively integrate multi-source heterogeneous data such as infrared thermal imaging and geometric point clouds, resulting in the lack of correlation analysis between internal defects and surface damage, and insufficient refined assessment capabilities for high-value areas. It is easy to ignore hidden damage or over-intervene in low-value areas, affecting the accuracy of repair plans. This solution creatively adopts an automatic damage assessment module, constructs an improved clustering model, and performs automatic damage assessment. It captures damage representations from micro-texture to macro-structure through multi-scale fusion, and dynamically adjusts the recognition sensitivity based on the quantitative weight of the building's historical value. At the same time, it introduces an adaptive clustering mechanism to optimize regional segmentation according to the building's value density, providing an objective and quantifiable scientific basis for subsequent repair decisions.

[0094] Example 5, see Figure 1 and Figure 4 This embodiment is based on the above embodiment. In the repair auxiliary model construction module, it is used to construct the model required for the digital auxiliary repair of historical buildings, specifically to construct a repair decision transformer model as a repair auxiliary model;

[0095] The repair auxiliary model construction module specifically includes multimodal feature fusion, knowledge prior attention design, polar coordinate encoder design, building repair decision generation, and model construction and training;

[0096] The multimodal feature fusion is used to integrate multi-source repair evidence. Specifically, the damage level spatial heat map, building material features, and point cloud geometric features are aligned and compressed through feature splicing and 1×1 convolution operations to obtain multimodal fusion features.

[0097] The prior knowledge attention design is used to inject domain prior knowledge. Specifically, it interacts with the predefined ancient building knowledge graph and multimodal fusion features to obtain knowledge reinforcement features. The predefined ancient building knowledge graph is used to convert known ancient building repair experience into machine-understandable knowledge. Specifically, it is a knowledge graph constructed based on relevant technical regulations and management methods for ancient buildings. The prior knowledge attention design includes:

[0098] Knowledge graph embedding is used to introduce the experience of ancient building restoration. Specifically, the predefined ancient building knowledge graph is processed through a graph neural network to obtain ancient building knowledge embedding.

[0099] Attention interaction is used for knowledge-guided feature correction. Specifically, the correlation between the ancient building knowledge embedding and the multimodal fusion features is calculated through the attention mechanism to obtain the knowledge enhancement features. The formula used is as follows:

[0100] ;

[0101] Where, represents interactive attention, represents the softmax activation function, represents the attention query transformation matrix, represents the attention key transformation matrix, represents the attention value transformation matrix, represents multimodal fusion features, Ar represents the embedding of ancient building knowledge, represents the dimension of the attention key, represents the knowledge enhancement feature, Representation layer normalization function;

[0102] The polar coordinate encoder design is used to model building spatial relationships. Specifically, it captures the topological associations of building damage areas through rotationally symmetric polar coordinate position encoding and a six-layer Transformer architecture to obtain deep building features, including:

[0103] Polar coordinate position encoding is used to enhance geometric perception. Specifically, the standard position encoding is replaced by a polar coordinate function to obtain a rotationally symmetric position embedding. The formula used is as follows:

[0104] ;

[0105] Where r represents the distance from the pixel point of the input image data to the center of the image, a represents the column coordinate of the pixel point of the input image data, b represents the row coordinate of the pixel point of the input image data, A represents the height of the input image data, and B represents the width of the input image data. Represents the angle between the pixel point of the input image data and the center of the image. Represents the value of the position encoding in dimension 2c, Indicates the value of the position code in dimension 2c+1, C indicates the number of position code dimensions, represents rotationally symmetric position embedding, and Pe represents position encoding;

[0106] Deep feature extraction, used to extract deep historical building damage information, specifically by processing rotationally symmetric position embeddings through a six-layer Transformer architecture to obtain deep building features;

[0107] The building repair decision generation is used to generate quantifiable decisions. Specifically, it outputs building repair decision prediction results through damage risk quantification and multi-objective prediction, including:

[0108] Damage risk quantification is used to calculate the degree of damage risk. Specifically, the damage risk coefficient is obtained by weighting the proportion of pixels with different damage levels. The formula used is as follows:

[0109] ;

[0110] Where, Indicates the proportion of pixels with damage level m, The element representing the pixel-level damage level label map is the building damage level label of the input image data at coordinates (a, b). Represents an indicator function. When the building damage level label of the input image data at the coordinate (a, b) is m, its value is 1, otherwise, its value is 0. Ri represents the damage risk coefficient. represents the risk weight of the damage level m;

[0111] Multi-objective prediction is used for preliminary decision-making prediction. Specifically, it uses three independent output heads to predict the type of repair technology, the level of repair cost, and the degree of preservation of the value of architectural relics. The formula used is as follows:

[0112] ;

[0113] Where, represents the prediction result of the type of repair technology, It represents the repair cost level prediction result, It represents the predicted result of the preservation degree of architectural cultural relics value. represents the weight of the second fully connected layer of the output head of the repair technology type, Represents the weight of the first fully connected layer of the output head of the repair technology type, Represents the bias term of the first fully connected layer of the output head of the repair technology type, Represents the bias term of the second fully connected layer of the output head of the repair technology type, represents the weight of the second fully connected layer of the repair cost level output head, represents the weight of the first fully connected layer of the repair cost level output head, represents the bias term of the first fully connected layer of the repair cost level output head, represents the bias term of the second fully connected layer of the repair cost level output head, The weight of the second fully connected layer of the output head representing the degree of preservation of the architectural heritage value, The weight of the first fully connected layer of the output head representing the degree of preservation of the architectural heritage value, The first fully connected layer bias term of the output head representing the degree of preservation of architectural heritage value, The second fully connected layer bias term of the output head representing the degree of preservation of architectural heritage value, represents the sigmoid activation function, represents the GELU activation function, represents the global average pooling function, Indicates the deep features of the building;

[0114] Risk-aware decision correction is used to integrate the damage risk coefficient. Specifically, the damage risk coefficient is used to dynamically adjust the output of the three decision dimensions of repair technology type, repair cost level, and preservation degree of architectural cultural relic value to obtain the building repair decision prediction result. The formula used is as follows:

[0115] ;

[0116] Where, represents the building renovation decision prediction result, represents the compensation coefficient of the type of repair technology, represents the compensation coefficient of the repair cost level, Indicates the compensation coefficient for the preservation of architectural cultural relics value, Indicates the weight of repair technology that can be learned;

[0117] The constructing and training of the model specifically involves constructing a repair decision transformer model through the multimodal feature fusion, the knowledge prior attention design, the polar coordinate encoder design, and the building repair decision generation. The model is trained based on the auxiliary repair training set and the auxiliary repair test set and the model performance is verified to obtain a repair decision transformer model as a repair auxiliary model.

[0118] By performing the above operations, the digital assisted repair system of traditional historical buildings has a weak ability to model the relationship between building spatial structures, ignores the topological correlation of damaged areas, and often evaluates multiple objectives such as cost, technology, and value protection separately. It is impossible to quantify the impact of dynamic risks on decision variables, which can easily lead to resource allocation imbalance or excessive repairs. In addition, the traditional model lacks a domain knowledge embedding mechanism, and it is difficult to ensure that repair measures meet the technical problems of cultural relics protection standards. This solution creatively adopts the repair decision transformer model as a repair auxiliary model. Through multimodal damage fusion and domain knowledge graph, the cultural relics protection standards are converted into prior constraints to ensure that decisions comply with regulations. Combined with polar coordinate position coding, the geometric topological relationship of the building is accurately modeled to enhance the perception of the symmetrical structure of the building. At the same time, the risk perception mechanism is used to dynamically associate the damage severity with the decision output to achieve the coordinated optimization of the three-dimensional goals of repair technology, cost control and value retention.

[0119] Example 6, see Figure 1 This embodiment is based on the above embodiment. In the historical building repair auxiliary module, the data set to be processed is used as the input of the repair auxiliary model to obtain a building repair decision reference result. Based on the building repair decision reference result, the digital repair of historical buildings is assisted.

[0120] It should be noted that, in this document, relational terms such as first and second, etc., are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "comprises," "comprising," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not explicitly listed, or elements inherent to such process, method, article, or apparatus.

[0121] While the embodiments of the present invention have been shown and described, it will be apparent to those skilled in the art that various changes, modifications, substitutions, and alterations can be made to the embodiments without departing from the principles and spirit of the invention.

[0122] The present invention and its embodiments are described above. This description is not restrictive. The drawings show 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 and, without departing from the purpose of the present invention, designs structures and embodiments similar to this technical solution without inventiveness, they shall fall within the scope of protection of the present invention.

Claims

1. A digital assisted restoration system for historical buildings, characterized by: The system includes a building data acquisition module, a data preliminary processing module, an automatic damage assessment module, a repair auxiliary model construction module and a historical building repair auxiliary module; The building data acquisition module acquires a building repair original data set by performing data acquisition, wherein the building repair original data set specifically includes a building repair history original data set and a building repair current original data set; The data preliminary processing module uses a data preliminary processing method of multi-source data alignment, building value assessment, point cloud feature extraction and damage pre-labeling to obtain a historical preliminary data set and a current preliminary data set; The automatic damage assessment module is used to automatically identify and assess damaged areas of historical buildings. Specifically, it extracts multi-scale damage features through analysis, integrates the building value weight matrix, and constructs an improved clustering model for processing to obtain a spatial heat map of damage levels. The repair auxiliary model construction module is used to construct the model required for the digital auxiliary repair of historical buildings, specifically to construct a repair decision transformer model as a repair auxiliary model; The historical building repair auxiliary module specifically uses the data set to be processed as the input of the repair auxiliary model to obtain a reference result for building repair decision-making, and assists in the digital repair of historical buildings based on the reference result for building repair decision-making.

2. The digital assisted restoration system for historical buildings according to claim 1, characterized in that: The automatic damage assessment module specifically includes multi-scale feature extraction, feature enhancement, improved clustering model design, damage heat map generation and building damage assessment.

3. The digital assisted restoration system for historical buildings according to claim 2, characterized in that: The multi-scale feature extraction is used to capture damage features of different sizes. Specifically, the multi-scale feature extraction is performed by analyzing the fusion information of architectural optical image data and architectural infrared image data through a convolutional neural network, extracting cross-scale feature maps from fine-grained textures to macroscopic structures, and obtaining cross-scale features, including: Pyramid convolution is used to extract scale-sensitive features. Specifically, it fuses architectural optical features and architectural infrared features through channel splicing operations, and performs multi-granularity feature extraction on the fused features using convolution kernels of three different sizes to obtain multi-scale features. Cross-scale feature fusion is used to integrate multi-granularity information. Specifically, small-scale features are fused with large-scale features through upsampling and weighted addition operations to obtain cross-scale features; The feature enhancement is used to integrate the historical value and damage characteristics of the building. Specifically, the building value weight is expanded to the same dimension as the cross-scale feature through a tensor broadcast operation, and multiplied element-by-element with the cross-scale feature to obtain the value enhancement feature; The improved clustering model design is used for damage area segmentation. Specifically, it realizes adaptive division of damage areas through superpixel segmentation and dynamic density clustering, and obtains pixel-level damage level labels, including: Superpixel segmentation is used to reduce computational complexity. Specifically, a simple linear iterative clustering algorithm is used to segment the input image data into multiple homogeneous texture regions. Each homogeneous texture region is represented by a superpixel unit, and a superpixel unit set is obtained. Regional feature aggregation is used to extract representative features of each homogeneous texture region. Specifically, the average value of the value enhancement feature in each superpixel unit is calculated through the average pooling operation to obtain the representative features of the superpixel unit. Dynamic density clustering is used for adaptive area segmentation. Specifically, the search radius of DBSCAN clustering is dynamically adjusted by the building value weight. Superpixel units are clustered based on their representative features to obtain cluster labels for damaged areas. Damage level mapping is used to generate automatic damage assessment results. Specifically, it maps the damaged area cluster labels to five-level building damage level labels using a small number of annotated samples, obtaining pixel-level damage level labels. The damage heat map is generated to visualize the automatic damage assessment results. Specifically, pixel-level damage level labels and building value weights are superimposed on the original input image data through color mapping to obtain a damage level spatial heat map; The building damage assessment specifically uses the historical preliminary data set and the current preliminary data set as inputs of the automatic damage assessment module, performs automatic building damage assessment through the multi-scale feature extraction, the feature enhancement, the improved clustering model design and the damage heat map generation, obtains a data set to be processed and a historical assessment data set, and performs data set segmentation on the historical assessment data set to obtain an auxiliary repair training set and an auxiliary repair test set.

4. The digital assisted restoration system for historical buildings according to claim 1, characterized in that: The repair auxiliary model construction module specifically includes multimodal feature fusion, knowledge prior attention design, polar coordinate encoder design, building repair decision generation, and model construction and training.

5. The digital assisted restoration system for historical buildings according to claim 4, characterized in that: The multimodal feature fusion is used to integrate multi-source repair evidence. Specifically, the damage level spatial heat map, building material features, and point cloud geometric features are aligned and compressed through feature splicing and 1×1 convolution operations to obtain multimodal fusion features. The prior knowledge attention design is used to inject domain prior knowledge. Specifically, it interacts with the predefined ancient building knowledge graph and multimodal fusion features to obtain knowledge reinforcement features. The predefined ancient building knowledge graph is used to convert known ancient building repair experience into machine-understandable knowledge. Specifically, it is a knowledge graph constructed based on relevant technical regulations and management methods for ancient buildings. The prior knowledge attention design includes: Knowledge graph embedding is used to introduce the experience of ancient building restoration. Specifically, the predefined ancient building knowledge graph is processed through a graph neural network to obtain ancient building knowledge embedding. Attention interaction is used for knowledge-guided feature correction. Specifically, the correlation between the ancient building knowledge embedding and the multimodal fusion features is calculated through the attention mechanism to obtain knowledge-enhanced features. The polar coordinate encoder design is used to model building spatial relationships. Specifically, it captures the topological associations of building damage areas through rotationally symmetric polar coordinate position encoding and a six-layer Transformer architecture to obtain deep building features, including: Polar coordinate position encoding is used to enhance geometric perception. Specifically, the standard position encoding is replaced by a polar coordinate function to obtain a rotationally symmetric position embedding. Deep feature extraction, used to extract deep historical building damage information, specifically by processing rotationally symmetric position embeddings through a six-layer Transformer architecture to obtain deep building features; The building repair decision generation is used to generate quantifiable decisions. Specifically, it outputs building repair decision prediction results through damage risk quantification and multi-objective prediction, including: Damage risk quantification is used to calculate the degree of damage risk. Specifically, the damage risk coefficient is obtained by weighting the proportion of pixels with different damage levels. Multi-objective prediction is used for preliminary decision-making predictions. Specifically, it uses three independent output heads to predict the type of repair technology, the level of repair cost, and the degree of preservation of the value of architectural relics; Risk-aware decision correction is used to integrate the damage risk coefficient. Specifically, it dynamically adjusts the output of three decision dimensions: the type of repair technology, the level of repair cost, and the degree of preservation of architectural cultural relic value through the damage risk coefficient to obtain the building repair decision prediction result; The constructing and training of the model specifically involves constructing a repair decision transformer model through the multimodal feature fusion, the knowledge prior attention design, the polar coordinate encoder design, and the building repair decision generation. The model is trained based on an auxiliary repair training set and an auxiliary repair test set and the model performance is verified to obtain a repair decision transformer model as a repair auxiliary model.

6. The digital assisted restoration system for historical buildings according to claim 1, characterized in that: The original data set of building renovation history and the current original data set of building renovation both include building optical image data, building infrared image data, building basic information data, building material data and building 3D point cloud data. The original data set of building renovation history also includes annotation data.

7. The digital assisted restoration system for historical buildings according to claim 1, characterized in that: In the data preliminary processing module, the multi-source data alignment is used to unify the data space coordinate system, specifically by pixel-level alignment of the building optical image data, building infrared image data and building 3D point cloud data through feature point matching; The building value assessment is used to quantify the historical value of a building. Specifically, the building construction year data, building cultural relic protection level data, and building process type data are converted into weights through preset rules combined with a multi-layer perceptron to construct a building value weight matrix; The point cloud feature extraction is used to characterize the geometric characteristics of the building surface. Specifically, the building 3D point cloud data is processed through normal vector calculation and depth projection to generate point cloud geometric features that are aligned with the building optical image data and the building infrared image data; The damage pre-labeling is used to pre-label the building damage level. Specifically, a binary mask of building damage is generated by analyzing the differences between the building optical image data and the building infrared image data. The binary mask of building damage is then manually corrected and the damage level is divided to obtain a building damage level label. Through the multi-source data alignment, the building value assessment, the point cloud feature extraction and the damage pre-labeling, the historical original data set of the building repair and the current original data set of the building repair are preliminarily processed to obtain a historical preliminary data set and a current preliminary data set.

Images