A point-embedded micro-disturbance acupuncture-style protection and restoration method for old town historical sites

Abstract

This invention proposes a point-embedded micro-perturbation acupuncture-style protection and restoration method for old city historical sites, belonging to the field of old city restoration technology. The method includes: collecting full-element data to construct a spatiotemporal four-dimensional digital twin base and multi-source datasets; constructing a texture pattern knowledge graph using large language models and large visual models; training a cross-modal texture pattern recognition model using the base and knowledge graph to generate a pattern feature library; performing diagnosis and stress assessment through hyperspectral imaging, target detection, and ANSYS analysis to generate a diagnostic result set; optimizing a restoration strategy library through conflict detection; and completing the restoration using a point-embedded micro-perturbation construction process. This method achieves full-element data integration, solves the problem of deep texture pattern recognition, transforms traditional experience into a digital knowledge system, avoids over-restoration through cross-modal recognition, achieves minimal intervention through digital twin pre-simulation, and optimizes a reusable restoration strategy library through conflict detection, thereby improving restoration efficiency and quality.

Description

Technical Field

[0001] This invention relates to the field of old town restoration technology, and in particular to an acupuncture-style protection and restoration method for old town historical sites using point-embedded micro-disturbance. Background Technology

[0002] As an important carrier of urban cultural memory, the protection and restoration of old city historical areas has long been a focus of attention for academia and industry. At present, the protection and restoration of historical buildings in China mainly follows the traditional experience-driven restoration model. In terms of technical means, it mainly relies on traditional manual surveys, supplemented by digital detection methods such as three-dimensional laser scanning, close-range photogrammetry, and infrared thermal imaging to obtain information on the geometric form and some surface texture patterns of the buildings. The formulation of restoration plans depends heavily on expert experience. Restoration design documents are compiled after on-site surveys, photographic records, and manual mapping. Traditional techniques such as plastering, wooden component splicing, and traditional building reinforcement are used in the restoration process. In recent years, with the promotion of Building Information Modeling (BIM) technology, some research institutions have attempted to apply BIM to the information management of historical buildings. Some scholars have also explored the use of deep learning algorithms to automatically identify visible texture patterns such as cracks and peeling on building surfaces. Some cities have introduced finite element analysis software to perform structural verification of key wooden structures in historical district protection projects. However, most of these technologies remain at the level of single-point application in a single stage. There is a lack of systematic data integration and intelligent linkage between data collection, texture pattern diagnosis, scheme formulation, and construction execution. The inheritance of historical craft knowledge also mainly relies on oral records and paper drawings. There is no structured knowledge system that can be understood and reasoned by machines. Overall, it is still in a semi-automated stage of digital assistance for human decision-making. This results in long restoration cycles and high costs. Furthermore, due to the lack of deep integration of cross-modal data and intelligent reasoning, it is difficult to accurately grasp the original characteristics and evolution of the texture and pattern of historical buildings, which can easily lead to over-restoration or under-restoration. At the same time, the inheritance of traditional restoration techniques relies on human experience and is difficult to standardize and replicate, resulting in inconsistent restoration quality between different projects. Some restoration plans have caused unnecessary disturbance to the original spatial pattern and architectural texture of the old city historical area because they have not fully considered the compatibility of historical techniques. Summary of the Invention

[0003] This invention aims to at least solve the technical problems existing in the prior art, and innovatively proposes a point-embedded micro-disturbance acupuncture-style protection and restoration method for old town historical sites.

[0004] To achieve the above-mentioned objectives of this invention, this invention provides a point-embedded micro-disturbance acupuncture-style protection and restoration method for old town historical sites, the method comprising: S1. Collect all elements of historical site data, construct a spatiotemporal four-dimensional digital twin base including texture pattern, value carrier, social network, architectural style and historical evolution trajectory, and form a multi-source dataset. The multi-source dataset also includes environmental microclimate data, ecological environment data, and human activity data: By deploying IoT sensor arrays in key areas of historical sites, real-time parameters such as temperature, humidity, wind speed, and solar radiation intensity are collected and stored in a time series to generate a microclimate time series dataset; by using high-resolution remote sensing image interpretation and on-site soil and vegetation sampling analysis, ecological indicators such as vegetation coverage, soil moisture content, and groundwater depth are obtained to generate an ecological environment feature database; by using mobile terminal positioning data mining and digitization of resident behavior questionnaires, human information such as pedestrian flow distribution, business type, and frequency of traditional activities is extracted to generate a human activity association dataset; the above-mentioned microclimate time series dataset, ecological environment feature database, and human activity association dataset are fused with a spatiotemporal four-dimensional digital twin base through coordinate system unification and semantic association to supplement the dimension of interaction between building components and the surrounding environment, forming a complete multi-source dataset containing multi-dimensional information about the building itself and the external environment.

[0005] S2. Based on multi-source datasets, construct a texture pattern knowledge graph using large language models and large visual models; S3. Train a cross-modal recognition model using a spatiotemporal four-dimensional digital twin base and a texture pattern knowledge graph to generate a pattern feature library; S4. Based on hyperspectral imaging and target detection algorithms, perform multi-scale diagnosis of the pattern feature database, use ANSYS finite element analysis combined with large model reasoning to evaluate the distribution of value carriers, and generate a diagnostic result set containing pattern information and carrier location. S5. Generate a repair plan based on the diagnostic result set; S6. Conduct conflict detection and optimization of the restoration plan to form a restoration strategy library that includes restoration of texture and pattern, value carrier, social network, and architectural style. S7. Based on the adaptation scheme selected in the repair strategy library, the on-site repair work is completed using a point-embedded micro-disturbance construction process.

[0006] The beneficial effects of this invention are as follows: By constructing a spatiotemporal four-dimensional digital twin substrate containing details at the building component level, material spectral characteristics, and historical evolution trajectory, this invention achieves comprehensive data integration from geometric shape and material composition to historical changes. This overcomes the limitations of existing technologies that can only acquire surface geometric information and shallow textural patterns. Combined with hyperspectral imaging multi-scale diagnosis and ANSYS finite element analysis, it performs deep stress assessment on traditional buildings and other concealed structures, effectively solving the problem of difficulty in detecting deep repair areas such as internal wall hollowing and internal decay of wooden structures. Simultaneously, it utilizes large-scale imaging... The verbal and visual models transform historical craft knowledge from oral history materials, old photographs, and surveying archives into a structured knowledge graph containing multi-dimensional relationships such as craft-component, craft-texture pattern, material-texture pattern, and craft-time sequence. This upgrades traditional restoration experience from oral records relying on personal inheritance to a digital knowledge system that can be understood, reasoned about, and invoked by machines, fundamentally changing the semi-automated state of existing technologies that "digitally assist human decision-making." Based on this, the cross-modal texture pattern recognition model uses a five-dimensional joint feature vector encompassing visual, spectral, semantic, temporal, and structural safety dimensions. By achieving deep fusion and intelligent reasoning of multi-source data, the system accurately grasps the authentic characteristics and evolutionary patterns of historical buildings, effectively avoiding over-restoration or under-restoration. Furthermore, during the restoration scheme generation stage, digital twin simulations are used to quantitatively simulate construction paths, material usage, and disturbance ranges. The lowest disturbance index, highest technological feasibility, and satisfactory structural safety redundancy are used as selection criteria. Combined with point-embedded micro-disturbance construction techniques, the restoration effect is precisely controlled at the component level or even the point level. This achieves optimal results even in the constrained environment of densely built buildings and narrow alleyways in the old city's historical district. Minimal intervention and reversible repairs avoid unnecessary disturbance to the original spatial layout and architectural texture caused by traditional large-scale demolition and reconstruction. In addition, the scheme conflict detection and optimization mechanism automatically identifies, classifies, and iteratively resolves conflicts through multi-dimensional relationships of process-materials, process-time, and process-space. The optimized schemes are categorized by school of thought, region, and era to form an accumulative and iterative repair strategy library, which enables the standardized replication and cross-project reuse of traditional repair techniques. This significantly shortens the repair cycle, reduces repair costs, and improves the consistency of repair quality across different projects.

[0007] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0008] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, in which: Figure 1 This is a schematic diagram of the structure of a point-embedded micro-disturbance acupuncture-style protection and restoration method for old town historical sites according to the present invention. Detailed Implementation

[0009] Embodiments of the present invention are described in detail below. Examples of these embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and should not be construed as limiting the present invention.

[0010] like Figure 1 As shown, a point-embedded micro-disturbance acupuncture-style protection and restoration method for old town historical sites is described, the method comprising: S1. Collect all elements of historical site data, construct a spatiotemporal four-dimensional digital twin base including texture pattern, value carrier, social network, architectural style and historical evolution trajectory, and form a multi-source dataset. S2. Based on multi-source datasets, construct a texture pattern knowledge graph using large language models and large visual models; S3. Train a cross-modal recognition model using a spatiotemporal four-dimensional digital twin base and a texture pattern knowledge graph to generate a pattern feature library; S4. Based on hyperspectral imaging and target detection algorithms, perform multi-scale diagnosis of the pattern feature database, use ANSYS finite element analysis combined with large model reasoning to evaluate the distribution of value carriers, and generate a diagnostic result set containing pattern information and carrier location. S5. Generate a repair plan based on the diagnostic result set; S6. Conduct conflict detection and optimization of the restoration plan to form a restoration strategy library that includes restoration of texture and pattern, value carrier, social network, and architectural style. S7. Based on the adaptation scheme selected in the repair strategy library, the on-site repair work is completed using a point-embedded micro-disturbance construction process.

[0011] In step S7, firstly, based on the repair points and construction scope determined by the optimal repair scheme, the repair area is spatially positioned and laid out before on-site work. Combined with the coordinate information of the spatiotemporal four-dimensional digital twin base, precise registration of the on-site coordinates and the digital model is completed, marking the point-based construction area to be worked on, avoiding pollution or damage to surrounding intact historical components. Secondly, corresponding point-embedding processes are adopted for different types of repair areas: for hollow areas on the wall surface, a micro-drilling and adhesive injection process is used, injecting reversible repair material only through a few point-based injection holes at the edge of the hollow area, without large-area removal and re-plastering, preserving the original wall's complete texture; for loose traditional wooden structures, an invisible carbon fiber dowel point reinforcement process is used, inserting small... The dimensions of the reversibly reinforced components remain unchanged in terms of the original mortise and tenon joints and materials. For damaged decorative patterns on the surface, a point-based patching process is used, where only the missing or damaged patterns are repaired using the original techniques and materials, preserving the historical traces of the remaining original components. During the operation, on-site construction data is collected in real time and updated to a spatiotemporal four-dimensional digital twin base to dynamically monitor the degree of disturbance to surrounding components. If the disturbance range exceeds the pre-set threshold, the construction parameters and operation path are adjusted in a timely manner to ensure that the repair work is always controlled within the preset point-based operation range. After the repair work is completed, the three-dimensional morphological data, hyperspectral material data, and structural stress monitoring data of the repaired components are collected and stored in the repair strategy library and texture pattern knowledge graph to complete the knowledge accumulation for this round of repair.

[0012] As an optional embodiment of the present invention, optionally, forming a multi-source dataset includes: S101. Collect data on the building complex and components in the historical area, obtain the geometric shape and surface texture information of the building components, and generate the initial dataset of the three-dimensional point cloud model and orthophoto of the historical area. When collecting comprehensive data on the historical site, a combination of non-contact low-altitude laser scanning and close-range photogrammetry was prioritized to avoid disturbing the original environment of the old town's historical area with large-scale on-site operations. For concealed parts of the wooden beams and walls inside the buildings, handheld laser scanning combined with small endoscope imaging was used to supplement the data collection, ensuring coverage of all component-level details. At the same time, environmental elements such as ground paving, courtyard layout, and street texture were also collected to fully preserve the overall spatial characteristics of the historical site. Based on the collected multi-view images and point cloud data, existing incremental 3D reconstruction algorithms were used to complete model texture mapping and geometric repair, eliminating model holes and noise points caused by occlusion, preserving the original geometric deformation and surface unevenness features of the building components, and without smoothing or normalization processing, to restore the original state of the historical building components to the greatest extent possible.

[0013] S102. Collect data on the material spectral reflectance characteristics and elemental composition of the surfaces of building walls, timber frames and decorative components, and generate a material spectral feature library. Hyperspectral imaging equipment was used to sample different areas of different components point by point. Full-band reflectance data were collected for different materials such as traditional plaster, wood, brick, stone and stucco. Simultaneously, X-ray fluorescence spectroscopy analysis was combined to obtain the elemental composition information of the component materials. The reflectance characteristics and elemental composition information were bound to the component coordinates to complete the registration and fusion of material spectral data and geometric model. Finally, a standardized material spectral feature library was established with material type, age and region as classification dimensions.

[0014] S103. Based on the three-dimensional point cloud model of the historical site, use historical documents, old photos, surveying archives and oral history materials for digital processing and temporal semantic analysis to extract the time nodes and change trajectory information of the architectural spatial pattern evolution in various historical periods and generate a historical evolution trajectory dataset. For historical documents (such as local chronicles, repair records, and traditional architectural texts), after digitization using OCR technology, structured information is extracted using existing large language models (such as GPT-4 or BERT). Key data such as construction date, repair dates, and spatial layout adjustment events (such as the addition of auxiliary buildings or changes in door and window positions) are extracted and an event index is created by timestamp. For old photographs, image feature recognition is performed using a large visual model to extract visual information such as roof form, wall decoration patterns, and component materials. The photographs are then spatially aligned with corresponding areas in a 3D point cloud model using an image matching algorithm to determine their authenticity. The time frame of the photographs and the corresponding physical state of the buildings were determined. For paper-based survey archives (such as architectural floor plans and cross-sections from different periods), vectorization technology was used to convert the drawings into digital vector layers. Geometric parameters such as building dimensions and component layouts from each period were extracted and integrated with the coordinate system of the 3D point cloud model to form a multi-period architectural geometric evolution layer. For oral history materials (such as interview recordings of elderly residents and traditional craftsmen), the text was first converted using speech-to-text tools, and then semantic analysis was performed using a large language model to extract supplementary information such as the architectural history, details of traditional craftsmanship, and changes in spatial functions, filling the gaps in the documentary materials.

[0015] By integrating the aforementioned multi-source temporal information into a unified timeline, a three-dimensional historical evolution trajectory chain of "time-space-state" is constructed. Starting from the original construction date of the building, the time nodes of key events such as renovations, layout changes, and functional transformations are marked sequentially. The building state corresponding to each node (such as component additions or subtractions, material replacements, and spatial layout adjustments) is bound to the corresponding version of the 3D point cloud model, forming a dynamically traceable historical evolution trajectory dataset. Simultaneously, all temporal information is cross-validated. For example, the accuracy of time nodes is ensured by comparing the visual features of old photographs with the renovation times recorded in documents; the detailed dimensions of the evolution trajectory are improved by comparing the craftsmanship details in oral history materials with the component dimensions in survey archives. Finally, a temporal dataset containing a complete historical evolutionary timeline is generated.

[0016] S104. Based on the historical site 3D point cloud model, orthophoto initial dataset, material spectral feature library and historical evolution trajectory dataset, a unified coordinate framework is used to perform spatiotemporal alignment and semantic association fusion to generate a spatiotemporal registration fusion dataset. Using the 2000 National Geodetic Coordinate System of the city where the historical site is located as the global spatial reference, the coordinate system of the 3D point cloud model and the initial dataset of orthophotos is transformed to this reference to ensure that the positional accuracy error of all spatial data is controlled within ±5mm. At the same time, using the Gregorian calendar as the time reference, the time nodes (such as the construction year and the repair year) in the historical evolution trajectory dataset are converted into the standard timestamp format to establish a unified spatiotemporal framework of spatial coordinates + timestamps.

[0017] The reflectivity data and elemental composition information in the material spectral feature library are precisely bound to the corresponding component coordinates of the 3D point cloud model. The discrete spectral sampling point data are mapped to the continuous component surface through the existing spatial interpolation algorithm, realizing the spatial fusion of material information and geometric model. The building states of each period in the historical evolution trajectory dataset (such as component additions and subtractions, layout adjustments, and material replacements) are associated with the corresponding versions of the 3D point cloud model on the time axis. Each time node corresponds to a snapshot of the building state, forming a three-dimensional associated data structure of spatial location-material features-historical state.

[0018] Cross-validation was performed on data from different sources (such as differences in component dimensions between surveying archives and oral history materials, and geometric deviations between old photographs and laser scanning models): laser scanning data or official authoritative archive data were given priority as the benchmark, and conflicting data were corrected by assigning weights to multi-source data (such as laser scanning data weight 0.6, archive data weight 0.3, and oral history material weight 0.1); duplicate data from the same area were deduplicated to retain the most complete information dimensions (such as retaining both point cloud models and image data and establishing a correlation index when both exist).

[0019] The final spatiotemporal four-dimensional digital twin foundation is constructed by organizing the fused dataset into a hierarchical structure of time slices, spatial components, and semantic attributes. Each time slice corresponds to the complete state of a key period in a historical site, including the geometric shape, material spectral characteristics, historical event information, and spatial relationships of all components during that period. Spatiotemporal indexing technology enables rapid querying and status tracing of any component at any time point, supporting dynamic simulation of the evolution of a building from construction to its various renovations.

[0020] S105. Based on the spatiotemporal registration and fusion dataset, the digital twin modeling engine is used to perform four-dimensional coupling modeling of the geometric detail dimension of building components, the material spectral feature dimension, the historical evolution trajectory dimension and the time evolution dimension, to generate a spatiotemporal four-dimensional digital twin base containing building component-level details, material spectral features and historical evolution trajectory, forming a multi-source dataset.

[0021] By employing a digital twin modeling engine that supports real-time fusion of multi-dimensional data (such as a customized development platform based on the Unity engine), and using a component-level 3D geometric model as the spatial basis, material spectral feature layer, historical evolution trajectory layer and time evolution layer are superimposed in sequence to achieve deep coupling of four-dimensional data.

[0022] Specifically: The core construction of geometric detail dimension: convert the 3D point cloud model generated by S101 into parametric BIM components, and assign each component a unique identity (ID) containing geometric dimensions, spatial coordinates and component type (such as beam, column, wall, decorative component, etc.) information, forming an editable and queryable component-level spatial framework; Material spectral feature mapping and fusion: The material spectral data (reflectivity, elemental composition) with bound coordinates in S102 is associated with the surface mesh of the corresponding component through texture mapping technology to realize the visualization of material properties. For example, the spectral reflectance characteristics of traditional plaster components can be displayed through color gradient to show their compositional differences, and the elemental composition data of wooden components can be queried in real time through the tooltip floating window. Time binding of historical evolution trajectory: The historical evolution trajectory dataset generated by S103 is associated with the component ID by timestamp. The life cycle of each component (construction, repair, modification, replacement) is recorded in the model attributes in the form of time nodes. It supports quick switching to any historical period through the timeline slider to view the status changes of the component (such as the repair record of a wooden column in 1950 and the material replacement information in 2000). Dynamic simulation of time evolution: Based on time nodes of historical evolution trajectories, a dynamic correlation model of time-space-state is constructed, supporting continuous playback of the historical evolution process of building groups or individual components (such as from the original construction state to the current state after each renovation), intuitively displaying the changes in spatial pattern, component form, and material characteristics; the method for constructing the dynamic correlation model of time-space-state is as follows: Based on the time nodes in the historical evolution trajectory dataset, a sequence of state snapshots is generated for each component: each snapshot contains the component's geometric shape (such as existence, size changes, and positional offset), material characteristics (such as component proportion and surface texture state), and associated historical events (such as repair time, process type, and construction personnel information) at that time node, and establishes a one-to-one mapping relationship with the corresponding component in the spatiotemporal four-dimensional digital twin base through a unique component ID.

[0023] Design time-axis driven state transition rules: For gradual changes between adjacent time nodes (such as the gradual change in material spectral characteristics caused by the natural aging of wooden components, and the slow expansion of the hollow area on the surface of the wall), a linear interpolation algorithm is used to generate intermediate transition frames to ensure the continuity of state changes during dynamic simulation; for abrupt events (such as component replacement, spatial layout adjustment, and large-scale renovation), key frame markers are set to directly jump to the corresponding state during the simulation process, and the core information of the event (such as the cause of the event, the scope of components involved, and the traditional process used) is highlighted in a pop-up window.

[0024] Develop an interactive triggering and feedback mechanism: allow users to automatically retrieve complete status details of any component at a corresponding time point by clicking on it or dragging the timeline slider (including excerpts from historical documents, comparison images of old photos, material testing reports, structural safety assessment data, etc.); at the same time, allow users to manually set the simulation speed, pause, or rewind to a specific time point to achieve refined tracing and analysis of the component's life cycle.

[0025] The integrated multi-source data cross-validation module compares the current state snapshot with the original multi-source datasets (such as the geometric parameters of the laser scanning model, the composition data of the hyperspectral material feature library, and the event records in historical documents) in real time during the dynamic simulation. If a state deviation is found (such as the simulated component size not matching the survey archive), the deviation area is automatically marked and correction suggestions are pushed (such as prioritizing the adjustment of state parameters by referring to official archive data), ensuring the accuracy and reliability of the dynamically associated model.

[0026] Ultimately, the time-space-state dynamic correlation model constructed using the above methods can realize the dynamic simulation of the entire life cycle of the building complex and components in the old city's historical area from construction to the current state.

[0027] The data format is standardized using the IFC (Industry Foundation Classes) standard to ensure compatibility between the four-dimensional data and subsequent modules such as the texture and pattern knowledge graph and texture and pattern recognition model. The resulting spatiotemporal four-dimensional digital twin not only enables the visualization and historical tracing of all elements of the historical site, but also provides high-precision digital support for subsequent restoration diagnosis, restoration plan simulation, and construction process monitoring, ensuring the accuracy and reversibility of restoration work.

[0028] As an optional embodiment of the present invention, the construction of the texture pattern knowledge graph may include: S201. Use a large language model to identify textual knowledge entities in the historical evolution trajectory dataset in a multi-source dataset, and establish semantic relationships between textual knowledge entities. Using a core entity classification system of "construction-renovation-craftsmanship-materials-components," a large language model (such as GPT-4 and Llama2) fine-tuned for the historical building domain is employed to segment and semantically analyze multi-source text data, including historical documents, renovation archives, craftsmen's oral records, and local chronicles. This accurately identifies core knowledge entities such as "construction year," "renovation cycle," "traditional brick carving techniques," "glutinous rice mortar materials," and "timber structural components." Subsequently, leveraging the model's domain knowledge enhancement capabilities, diverse semantic relationships between entities are established: such as hierarchical relationships ("components" include "column bases" and "doors and windows"), temporal relationships ("renovation in 1950" is later than "construction in 1920"), associative relationships ("the walls of a certain courtyard house were constructed using glutinous rice mortar techniques"), and causal relationships ("decaying timber caused deformation of the beam frame"). Simultaneously, the identified entities and relationships are standardized and validated to ensure alignment with data terminology in the IFC standard.

[0029] S202. Using a large visual model, multi-source visual features are extracted from the 3D point cloud model of the historical site, the initial dataset of orthophotos, the component connection method, decorative component pattern features and material surface process traces in the material spectral feature library, to identify visual knowledge entities and establish spatial topological relationships and process logic associations between visual knowledge entities. The geometric topology analysis module based on a 3D point cloud model, and the visual large model (such as a customized version of CLIP or YOLOv8), extracts the features of the neighborhood point set and spatial coordinate relationships of the point cloud to identify typical component connection methods in traditional architecture, such as "dovetail tenon and mortise connection," "dougong tenon and mortise interlocking structure," and "timber frame through-beam connection." For each identified connection entity, its spatial topology relationship is further analyzed: such as "column components and beam components achieve a rigid vertical connection through dovetail tenon and mortise" and "dougong components are stacked on top of columns and form a horizontal support relationship with the lintel," and these relationships are mapped and aligned with the component connection type terminology in the IFC standard.

[0030] For areas such as building facades, doors, windows, and eaves in the initial orthophoto dataset, the visual big data model employs existing image semantic segmentation and feature matching algorithms to extract visual knowledge entities of decorative patterns, such as "brick carving with scrolling grass pattern," "wooden window lattice with cloud pattern," and "painted dragon and phoenix pattern." Simultaneously, by combining reflectance data from a material spectral feature library, the model distinguishes the material attributes of the patterns (e.g., the spectral differences between "brick carving patterns" and "painted patterns") and establishes the association between pattern entities and their respective components (e.g., "the brick carving pattern on the lintel of the main hall is a scrolling grass pattern, belonging to the decorative sub-entity of the door component"). Furthermore, by comparing with traditional pattern styles recorded in historical documents, the model verifies the cultural attributes of the pattern entities (e.g., "this scrolling grass pattern conforms to the decorative style of Suzhou-style architecture during the Ming and Qing dynasties").

[0031] By utilizing hyperspectral reflectance data from a material spectral feature library and detailed texture information from orthophotos, the visual big data model identifies surface artifacts using existing hyperspectral feature inversion models and texture feature extraction algorithms. Examples include "hand-carved wood grain marks," "the texture of traditional mortar plaster," "oxidation peeling marks on painted layers," and "rust marks on metal components." For each artifact, the model analyzes its technological logic: for instance, "the depth and direction of hand-carved marks indicate the use of traditional planing techniques," and "the distribution pattern of oxidation peeling on painted layers is related to the repair cycles recorded in historical documents." Simultaneously, the model binds the artifact's feature parameters (such as carving depth and peeling area) to the component's historical state data, forming a "craftsmanship mark - component state - time node" correlation chain.

[0032] This approach involves multimodal fusion of structural features from 3D point clouds, texture features from orthophotos, and compositional features from material spectra. By employing a cross-modal attention mechanism within a large visual model, the accuracy of entity recognition is improved (e.g., combining the 3D structure of the point cloud with the texture of the image allows for more precise identification of the layered structure of dougong brackets). Simultaneously, the identified visual knowledge entities and relationships are cross-validated: for example, the connection methods of components in the point cloud model are compared with construction methods recorded in historical documents to ensure the accuracy of the craft entities; the matching degree between material spectral features and craft traces verifies the rationality of the logical connections between craft processes. Finally, all visual knowledge entities and relationships are integrated into a unified knowledge graph framework, complementing the textual knowledge entities extracted by S201, to construct a texture pattern knowledge graph encompassing both textual and visual modalities.

[0033] S203. Cross-modal alignment and fusion of text knowledge entities, semantic relationships between text knowledge entities, spatial topological relationships between visual knowledge entities, and process logic associations are performed to establish cross-modal semantic mapping between visual knowledge entities and text knowledge entities, and generate a texture pattern knowledge graph.

[0034] A cross-modal alignment rule library is constructed, using IFC standard terminology as an intermediate mapping layer to establish correspondence rules between visual entities and text entities. For example, the visual entity "dovetail tenon joint structure" corresponds to the text entity "dovetail tenon craftsmanship," the visual entity "scroll pattern brick carving pattern" corresponds to the text entity "Ming and Qing Suzhou-style scroll pattern brick carving decoration," and the visual entity "hand-carved wood grain marks" corresponds to the text entity "traditional planing and chiseling process marks." The rule library covers core categories such as component connections, decorative patterns, and process marks, ensuring the accuracy and comprehensiveness of the mapping.

[0035] A cross-modal embedding model is employed for feature matching. Textual knowledge entities extracted in S201 (e.g., "glutinous rice mortar masonry technique") are converted into text vectors, and visual knowledge entities extracted in S202 (e.g., "mortar plastering texture features") are converted into visual feature vectors. A domain-adjusted CLIP model maps both to the same high-dimensional vector space, and cosine similarity is calculated to automatically match semantically related text and visual entities. Candidate pairs with similarity below 0.7 undergo manual review and correction to ensure alignment reliability.

[0036] Multimodal relationships are fused and linked using component IDs, merging textual semantic relationships (e.g., "the wall was constructed using glutinous rice mortar") with visual spatial / technical relationships (e.g., "the wall mortar texture matches the spectral characteristics of glutinous rice mortar") to form a cross-modal relationship chain of "component ID → glutinous rice mortar technique (text) → mortar finish texture (visual)". Simultaneously, visual entity time nodes (e.g., "the state of the brick carvings after the 1950 renovation") are bound to text timestamps (e.g., "the brick carving technique record of 1950") to ensure temporal consistency.

[0037] Implement cross-modal consistency verification and use domain rules (such as "dovetail tenon joints are only used for wooden frame connections") to verify the matching results: if the visual recognition is "dovetail tenon joint connection of stone components" but the text records it as stone tenon craftsmanship, then mark the conflict and trigger manual review; compare the material spectrum with the material composition in the text (such as whether the reflectance of glutinous rice mortar is consistent with the record) to verify the authenticity of the process.

[0038] A bimodal knowledge graph is formed by integrating and storing aligned entities and relationships as triples (e.g., <column base 123, craftsmanship: dovetail tenon and mortise, visual feature: dovetail tenon and mortise connection structure>) into the graph, constructing a multi-dimensional network covering textual semantics, visual features, spatial topology, and craftsmanship logic. Cross-modal queries are supported (e.g., entering "Ming and Qing dynasty scrollwork brick carving" returns textual craftsmanship, visual patterns, and component locations), and it is associated with a spatiotemporal digital twin base, enabling the interactive function of "clicking on the twin component → retrieving bimodal craftsmanship information".

[0039] As an optional embodiment of the present invention, the generation of the pattern feature library may include: S301. Utilize the process-component relationship, process-texture pattern relationship, material-texture pattern relationship, and temporal causal relationship in the texture pattern knowledge graph to perform multimodal texture pattern semantic guidance annotation on visual data, spectral data, and temporal data in the spatiotemporal four-dimensional digital twin base, and generate a multimodal texture pattern annotation dataset containing a visual texture pattern annotation dataset, a spectral texture pattern annotation dataset, and a temporal causal texture pattern annotation dataset. Guided by the "process-component-material-texture pattern" association network in the knowledge graph, semantic guidance annotation of multimodal data is completed in modules: Construction of a Visual Texture Pattern Annotation Dataset: Based on a spatiotemporal four-dimensional digital twin foundation, a 3D point cloud model and orthophoto are used. The mapping relationship between "component type - craftsmanship characteristics - susceptibility" in the knowledge graph is incorporated to accurately annotate visual features. For example, for components using the "mortise and tenon joint" technique, based on the knowledge graph's record that "mortise and tenon joints are prone to loosening due to wood shrinkage," visual features such as spatial gaps at the mortise and tenon joints and component displacement are identified in the point cloud model and labeled as "loose mortise and tenon joints." For "Ming and Qing Dynasty Suzhou-style brick carving components," combined with the knowledge that "brick carving is susceptible to weathering and peeling due to rainwater erosion," features such as cracks, missing areas, and color fading on the brick carving surface are marked in the orthophoto and labeled as "brick carving weathering and peeling." Simultaneously, referring to the severity grading standards in the knowledge graph (e.g., slight: crack length < 5cm; moderate: crack length 5-15cm with partial missing areas; severe: missing area > 10%), the annotated areas are classified into different levels.

[0040] Construction of a Spectral Texture Pattern Annotation Dataset: Based on the correspondence between "material composition-type-spectral features" in the knowledge graph, hyperspectral data in the material spectral feature library are annotated. For example, during the aging process of glutinous rice mortar, its spectral reflectance will significantly decrease in the 420nm-500nm band. Combining the association of "glutinous rice mortar aging → abnormal spectral features" in the knowledge graph, areas with reflectance in this band below the threshold (e.g., 80% of the baseline value) are labeled as "glutinous rice mortar aging". When wood decays, the reflectance in the near-infrared band (750nm-950nm) will increase by more than 20%, and the "decay" of wood components is labeled accordingly. In addition, for mixed material components (such as brick and stone mixed walls), the spectral features of different materials in the knowledge graph are used to distinguish and label the types of brick and stone respectively (e.g., brick weathering, stone alkali erosion).

[0041] Construction of a temporal causal texture pattern annotation dataset: Utilizing "temporal causal relationship chains" from the knowledge graph (such as "long-term humid environment → wood decay → beam deformation" and "frequent vibration → loosening of tenons and mortises → component displacement"), causal annotation is performed on the temporal data of the twin substrate. For example, temporal data of a certain wooden beam component shows: in 1995, the environmental humidity was consistently above 60% (in the knowledge graph, "wood humidity > 60% is prone to decay"); in 2005, spectral data showed wood decay characteristics; and in 2015, the point cloud model showed beam bending deformation. Combining this with the causal chain from the knowledge graph, the temporal causal relationship is labeled as "excessive humidity (cause) → wood decay (intermediate cause) → beam deformation (effect)". Simultaneously, the development cycle parameters in the knowledge graph (such as the 10-15 years required for wood decay from excessive humidity to deformation) are correlated to verify the rationality of the temporal causal chain.

[0042] Multimodal annotation fusion and verification: Visual, spectral, and temporal annotation results are bound together using a unique component ID to form a multimodal annotation record of "component ID - visual - spectral features - temporal causality". For example, the annotation record for column base component ID C101 includes: visual annotation "surface weathering cracks (moderate)", spectral annotation "stone calcium carbonate loss (reflectivity decrease of 15%)", and temporal annotation "weathering cycle from 2000 to 2020 caused crack propagation". Simultaneously, verification is performed using a knowledge graph rule base. If annotations such as "stone component annotation: mortise and tenon loosening" conflict with the process-component relationship, manual review and correction are automatically triggered to ensure the accuracy of the dataset.

[0043] The resulting multimodal texture pattern annotation dataset provides supervised information covering visual, spectral, and temporal dimensions for cross-modal texture pattern recognition models, achieving a deep correlation between features and process knowledge.

[0044] S302. Based on the multimodal texture pattern annotation dataset, a visual feature encoding stream is constructed using a large visual model. A semantic knowledge encoding stream is constructed by embedding process nodes, material nodes, process nodes and texture pattern nodes in the texture pattern knowledge graph using a graph neural network. A cross-modal fusion module based on the cross-attention mechanism and a multi-task output head including texture pattern classification, damage degree regression, position coordinate output and structural safety level assessment are constructed to generate a cross-modal texture pattern recognition model. The visual feature encoding stream employs a dual-branch parallel structure: For 3D point cloud data, the PointNet++ model is selected as the core encoder, extracting local geometric features (such as the curvature and normal direction of the point cloud neighborhood at mortise and tenon joints) and global structural features (such as the spatial topology of the timber frame) of the components through sampling and grouping layers, and mapping these features into a 256-dimensional fixed vector through a multilayer perceptron; For orthophotos and material spectral images, a ResNet-50 architecture combined with an attention mechanism is used to extract texture details (such as the direction and width of cracks), color features (such as the degree of color fading in peeling paint), and reflectance differences in hyperspectral bands of the repaired area, resulting in a 256-dimensional visual feature vector after feature compression. The two branches of features are integrated into a unified visual feature representation through a concatenation operation, providing a basic input for cross-modal interaction.

[0045] The semantic knowledge encoding stream is based on Graph Attention Network (GAT) to achieve node embedding learning. First, the process nodes (e.g., "mortise and tenon timber frame process"), material nodes (e.g., "glutinous rice mortar"), process nodes (e.g., "hand carving process"), and texture pattern nodes (e.g., "wood decay") in the texture pattern knowledge graph are transformed into 128-dimensional initial random vectors. Then, the multi-head attention mechanism of GAT is used to calculate the association weights between nodes (e.g., the causal association weight between "wood decay" and "excessive humidity"), and the node embeddings are updated to make nodes with strong semantic associations closer together in the vector space. The final output semantic knowledge vector contains deep association information between process logic, material properties, and texture pattern causality, providing knowledge guidance for visual features.

[0046] The cross-modal fusion module employs a bidirectional cross-attention mechanism: visual feature vectors are used as queries, and semantic knowledge feature vectors are used as keys and values. Attention weights are calculated to focus on semantic knowledge related to the current texture pattern (e.g., when identifying "brick carving weathering," nodes such as "brick carving craftsmanship" and "rainwater erosion texture pattern" are activated first). Simultaneously, semantic knowledge feature vectors are used as queries, and visual feature vectors are used as keys and values ​​to strengthen visual region features that match semantic logic (e.g., crack textures on the surface of brick carvings). The fused feature vectors retain both the visual details of the texture pattern and the inherent logic of the craftsmanship knowledge, enhancing the model's understanding of the essence of the texture pattern.

[0047] The multi-task output head comprises four collaborative subtasks: the texture pattern classification subtask uses a Softmax activation function to output the probability distribution of 12 common texture patterns (such as loose mortar joints, aging mortar, etc.); the damage severity regression subtask uses a linear activation function to output continuous values ​​of 0-1 (corresponding to three levels: minor, moderate, and severe); the location coordinate output subtask accurately locates the spatial position of texture patterns through bounding box regression (for images) and point cloud coordinate prediction (for 3D models); and the structural safety level assessment subtask outputs structural safety levels from level one to four based on fused features and safety rules in the knowledge graph (such as "if the beam deformation exceeds 5cm, the safety level is level three"). Each subtask is trained using a joint loss function (cross-entropy loss + mean squared error loss + L1 loss + weighted classification loss) to achieve collaborative optimization among multiple tasks, ensuring the model's comprehensive performance in repair area identification, localization, and risk assessment.

[0048] S303. Based on the multimodal texture pattern annotation dataset and the cross-modal texture pattern recognition model, a multi-stage joint training strategy is used to sequentially perform single-modal pre-training of visual feature encoding stream, pre-training of knowledge graph link prediction of semantic knowledge encoding stream, cross-modal alignment training of cross-modal fusion module, and end-to-end joint fine-tuning including knowledge regularization loss based on logical consistency constraints of texture pattern knowledge graph, to generate a trained cross-modal texture pattern recognition model. First, in the unimodal pre-training stage of the visual feature encoding stream: a self-supervised learning approach is adopted, utilizing unlabeled 3D point cloud, orthophoto, and hyperspectral image data from a spatiotemporal four-dimensional digital twin substrate for pre-training. For the PointNet++ branch, a point cloud local structure reconstruction task is designed, randomly occluding parts of the point cloud region to allow the model to predict the geometric features (such as curvature and normals) of the occluded region, thereby learning the spatial topology and local details of the component. For the ResNet-50 branch, the MoCov3 contrastive learning framework is used to augment the data of images of the same component from different viewpoints and at different time points, constructing positive sample pairs to allow the model to learn the consistency of visual features across viewpoints and time sequences. After pre-training, the parameters of the bottom convolutional layers of the encoder are fixed, while the trainability of the high-level feature extraction layers is preserved.

[0049] Secondly, in the pre-training stage of the semantic knowledge encoding stream knowledge graph link prediction: based on the triplet data of the texture pattern knowledge graph, a Graph Attention Network (GAT) combined with the TransE link prediction algorithm is used for pre-training. The node embeddings in the knowledge graph are initialized as random vectors. After learning the association weights between nodes through GAT, the score function of the triples is calculated using the existing TransE model. The node embeddings are optimized to minimize the scores of incorrect triples, enabling the model to accurately predict missing relationships in the knowledge graph (e.g., given a brick carving component, the model can output the correct prediction of "weathering and peeling"). After this training stage, the semantic knowledge encoding stream possesses strong reasoning ability related to craft knowledge.

[0050] Next, the cross-modal fusion module's cross-modal alignment training phase utilizes aligned text-visual entity pairs (such as "Ming and Qing Dynasty Suzhou-style scrollwork brick carving decoration" and "scrollwork brick carving pattern") from the bimodal texture pattern knowledge graph as training data. Text entities are transformed into semantic vectors via semantic encoding streams, and visual entities are transformed into visual vectors via visual encoding streams. The correlation between the two is calculated using a cross-attention mechanism. A contrastive loss function is employed to maximize the cosine similarity of matched text-visual vector pairs in high-dimensional space and minimize the similarity of unmatched pairs, thereby achieving effective alignment of cross-modal features.

[0051] Finally, in the end-to-end joint fine-tuning stage: the visual feature encoding stream, semantic knowledge encoding stream, and cross-modal fusion module are integrated into a complete model, and end-to-end training is performed using a multimodal texture pattern annotation dataset as supervised data. In this stage, a knowledge regularization loss based on the logical consistency constraint of the texture pattern knowledge graph is introduced: when the type of texture repair required predicted by the model conflicts with the association rule of "process-component-texture pattern" in the knowledge graph (e.g., predicting a loose mortise and tenon joint texture pattern in a stone component, while the knowledge graph clearly states that mortise and tenon joints are only used in wooden frames), the loss weight of that sample is increased. Simultaneously, the joint loss function of the multi-task output heads (cross-entropy loss for texture pattern classification, mean squared error loss for damage degree regression, L1 loss for location coordinate prediction, and weighted classification loss for structural safety level assessment) is used to jointly optimize all trainable parameters of the model. Through multiple rounds of iterative training, until the model's accuracy in texture pattern recognition, damage degree prediction error, and structural safety level assessment on the validation set all reach preset thresholds, a fully trained cross-modal texture pattern recognition model is finally generated.

[0052] S304. Based on the trained cross-modal texture pattern recognition model, a full texture pattern reasoning is performed using a spatiotemporal four-dimensional digital twin basis and cross-validated by combining a texture pattern knowledge graph. A structured texture pattern description is generated through large language model reasoning. The structured texture pattern description is encoded into a five-dimensional joint feature vector containing visual feature sub-vectors, spectral feature sub-vectors, semantic knowledge sub-vectors, temporal evolution sub-vectors, and structural safety sub-vectors. It is stored according to a three-level indexing system of building components-spatial location-temporal evolution to generate a pattern feature library.

[0053] The full-scale texture pattern reasoning process requires traversing all building component units in the spatiotemporal four-dimensional digital twin base. For each component, standardized preprocessing is performed on the 3D point cloud, orthophoto, hyperspectral data, and multi-period time-series monitoring data: 3D point clouds are filtered using statistical filtering to remove noise points and undergo voxelization downsampling; orthophotos are enhanced with histogram equalization to improve texture details; hyperspectral data undergoes band selection and normalization; and time-series data is sorted by timestamp and missing values ​​are filled in. The preprocessed data is then input into the trained cross-modal texture pattern recognition model. The model sequentially outputs the component type, continuous damage level values, spatial coordinates (3D coordinates in the point cloud model, bounding box coordinates in the orthophoto), and structural safety level.

[0054] The cross-validation process is based on a texture pattern knowledge graph: for each repair area result output by the model, the knowledge graph is automatically retrieved to identify the "prone texture patterns" associated with the process nodes and material nodes to which the component belongs, thus verifying the rationality of the repair type. For example, if the model predicts "loose mortise and tenon joints in stone column bases," while the knowledge graph clearly states "mortise and tenon structures only exist in wooden structural components," the result is marked as a "logical conflict," triggering a manual review process. If the repair type is consistent with the association rules in the knowledge graph, the degree of damage output by the model is further compared with the degree of matching in the repair grading standards in the knowledge graph to ensure that the result conforms to the logic of process knowledge.

[0055] The generation of structured texture patterns is achieved using existing large language models: the structured data output by the model (component ID, type, damage level, spatial location, temporal evolution characteristics, and safety level) is used as input. Prompt words guide the large language model to generate a standardized natural language description, which must cover basic component information, details, temporal changes, and safety assessment. For example: "Component ID: C203 (Ming and Qing Dynasty Suzhou-style brick carving screen wall) exhibits weathering and peeling of brick carvings, with a damage level of moderate (0.52). Its spatial location is in the left-central area of ​​the screen wall (orthophoto bounding box coordinates: x1=5.2, y1=3.1, x2=7.8, y2=5.6; 3D point cloud coordinates: X=18.5, Y=22.3, Z=0.8). The temporal evolution shows that from 2015 to 2025, rainwater erosion has caused the peeling area to expand from 5% to 8%, and the structural safety level is level three." The five-dimensional joint feature vector encoding needs to map the structured description into a multi-dimensional feature representation: the visual feature sub-vector is extracted from the 256-dimensional features output by the model's visual encoding stream, preserving the geometric and texture information of the repaired area; the spectral feature sub-vector comes from the 64-dimensional features encoded by the ResNet-50 branch of hyperspectral data, reflecting changes in material composition; the semantic knowledge sub-vector is a 128-dimensional vector embedded by GAT of the relevant nodes (process, material) of the component in the knowledge graph, containing semantic association logic; the temporal evolution sub-vector obtains 64-dimensional features through LSTM encoding of temporal data, characterizing the development trend; the structural safety sub-vector transforms the safety level into a 32-dimensional numerical vector (such as the one-hot encoding extension corresponding to [1,0,0,0] for level one). Finally, the five sub-vectors are concatenated into a 544-dimensional five-dimensional joint feature vector, realizing a unified representation of multimodal information.

[0056] The three-level index system is constructed following the logic of "building components - spatial location - temporal evolution": the first-level index is divided by component type (e.g., timber frame, brick and stone, painted), and each type is associated with the feature vectors and descriptions of all components in that type; the second-level index is grouped by building spatial region under component type (e.g., east courtyard, west courtyard, main hall, wing rooms, etc. of a building complex), facilitating retrieval by spatial range; the third-level index is divided by time sequence under spatial region (e.g., every 5 years is a time window), recording the feature changes at different times. The index data is stored in a distributed database, supporting fast querying and multi-dimensional filtering.

[0057] As an optional embodiment of the present invention, optionally, generating a diagnostic result set including the degree of damage, location, and structural safety level includes: S401. Based on the spectral feature sub-vectors and position coordinate information in the pattern feature library, the surface of building components is scanned in multiple bands using a hyperspectral imaging device. Combined with the texture pattern sensitive bands in the pattern feature library, multi-scale pyramid decomposition and damage degree weighting are performed to generate a multi-scale hyperspectral diagnostic image set. For the identified restoration areas, sensitive bands corresponding to the restoration areas in the pattern feature library are extracted first, such as cellulose degradation feature bands corresponding to wood decay, mineral composition change feature bands corresponding to brick and stone weathering, and pigment molecule feature bands corresponding to paint fading. Irrelevant bands are skipped to reduce computational redundancy. Then, a Gaussian pyramid is constructed on the hyperspectral scan data, and multi-scale decomposition is performed sequentially from the whole image to local pixels. The bottom layer retains the macroscopic distribution information of the whole component, the middle layer retains the contour information of the restoration area, and the top layer retains the microscopic detail information of the edge of the restoration area. Then, weights are applied to sub-images of different scales according to the degree of damage recorded in the pattern feature library. The top detailed sub-image corresponding to severe restoration is given higher weight, and the bottom macroscopic sub-image corresponding to mild restoration is given higher weight. Finally, a multi-scale hyperspectral diagnostic image set that takes into account both global and local details is obtained.

[0058] S402. Based on the visual feature sub-vectors in the multi-scale hyperspectral diagnostic image set and the pattern feature library, the target detection algorithm (YOLOv7) finely tuned by transfer learning of the pattern feature library is used to perform four-level scale texture pattern localization and recognition and damage degree regression quantitative assessment, and generate texture pattern localization diagnostic results. First, the transfer learning fine-tuning stage: Based on the pre-trained YOLOv7 model, transfer learning is performed on the model using multi-scale hyperspectral images, visual feature sub-vectors, and corresponding repair location and damage degree labels already labeled in the pattern feature library. Specifically, the parameters of the YOLOv7 backbone network are frozen, and only the feature fusion layer of the neck and the detection and regression branches of the head are trained; to meet the requirements of four-level scale detection, a cross-scale feature fusion module is added to the neck to concatenate hyperspectral sub-image features and visual feature sub-vectors at different scales, enhancing the correlation of multi-scale features; the head is designed with a dual-branch structure, one of which is a texture pattern classification and localization branch, outputting the bounding box coordinates and repair type probability at four scales; the other is a damage degree regression branch, which takes the fused feature vector as input and outputs a continuous damage value of 0-1 through a fully connected layer.

[0059] Secondly, the four-level scale repair location and recognition process is as follows: Level 1 (building cluster scale): Input the bottom macro sub-image of the multi-scale hyperspectral diagnostic image, the model detects all individual buildings within the cluster that are undergoing repair, and outputs building-level bounding boxes and preliminary repair types; Level 2 (individual building scale): Based on the location results of Level 1, the middle-level sub-image of the corresponding individual building is cropped and input into the model to locate specific components (such as timber frames and brick walls), and outputs component-level bounding boxes; Level 3 (component scale): The top-level detail sub-image of the component is cropped to accurately outline the repair area on the surface of the component (such as weathering spots on brick carvings and cracks in wooden beams), and outputs the bounding box of the repair area; Level 4 (local micro scale): Combine the spectral feature sub-vectors in the pattern feature library to quantify and extract the micro-texture of the repair area (such as crack width and the proportion of peeling area) and supplement it to the location results.

[0060] Finally, the damage severity is quantitatively assessed through regression: the model's regression branch takes the fused features (visual texture + spectral composition) at four scales as input and outputs continuous damage severity values ​​through a multilayer perceptron (MLP). During training, the mean squared error (MSE) loss function is used to compare and optimize the predicted values ​​with the labeled damage severity (e.g., 0.52 corresponds to moderate). Simultaneously, historical process knowledge constraints from the pattern feature library are introduced. For example, when the model predicts the damage severity of brick and stone components, the regression results are calibrated by referring to the "association rules between the degree of weathering damage and the loss rate of mineral components in the knowledge graph," ensuring that the quantified values ​​conform to the process logic. After training, the model's localization mAP on the validation set must reach over 90%, and the mean absolute error (MAE) of the damage severity regression must be controlled within 0.05. Finally, a localization diagnosis result containing four-scale bounding boxes, type, and damage severity is generated.

[0061] S403. Based on the location coordinates and damage information of the traditional wooden building in the texture pattern location diagnosis results, the traditional building mechanical model is constructed using ANSYS finite element analysis. The stress distribution is solved by applying material properties and damage degradation parameters from the material spectral feature library and combined load conditions. The structural safety assessment is carried out by combining the large language model reasoning and the process-component relationship in the texture pattern knowledge graph, and the traditional building structural safety diagnosis results are generated. Traditional building mechanics model construction relies on precisely extracted 3D geometric information from a digital twin base: the target traditional building is located using positional coordinates from a pattern feature library, and the original 3D point cloud data of its tenons and mortises are exported from the spatiotemporal 4D digital twin model. A 1:1 scale refined geometric model is then reconstructed using reverse engineering software. Local corrections are made to the model based on damage level information (such as the gap between loose tenons and mortises, and the extent of wood decay). If loose tenons and mortises are found, gap units (the gap size matches the quantified value of the damage level) are added to the interface between the tenon and mortise. If wood decay is detected, the decayed area is removed using Boolean operations, and this area is marked as a zone of mechanical performance degradation.

[0062] The application of material properties and damage degradation parameters requires association with the material spectral feature library: key indicators such as wood moisture content, cellulose degradation rate, and lignin loss rate are extracted from the spectral feature subvectors of the texture feature library. Combined with the "wood physicochemical indicators-mechanical properties" association rules in the texture pattern knowledge graph (e.g., for every 10% increase in moisture content, the elastic modulus decreases by 15%; when the cellulose degradation rate reaches 30%, the shear strength decreases by 25%), the spectral indicators are converted into material parameters (elastic modulus, Poisson's ratio, shear strength, etc.) that can be recognized by ANSYS, and the parameters of the decayed area are subjected to corresponding degradation treatment.

[0063] The combined load case settings need to cover the actual stress scenarios of the old city buildings: referring to the technical specifications for the maintenance and reinforcement of old city building timber structures, apply dead load (component self-weight, calculated according to timber density and geometric volume), live load (a uniformly distributed load of 0.5 kN / m2 applied to the upper components), wind load (calculated according to the basic wind pressure of the building area and the building shape coefficient, applied in the horizontal direction), and seismic load (calculated using the response spectrum method, inputting the peak acceleration corresponding to the regional seismic intensity); at the same time, combined with the load change characteristics in the time series monitoring data, simulate the timber shrinkage and expansion stress caused by temperature changes in different seasons (apply temperature load through the thermal analysis module).

[0064] In the stress distribution solution stage, the statics solver of ANSYS Workbench was used for calculation: contact pairs were set (the tenon and mortise are in frictional contact, and the friction coefficient is referenced from the knowledge graph "the range of friction coefficient of wooden tenon and mortise"). When dividing the mesh, the contact area of ​​the tenon and mortise and the decayed area were densified (the mesh size was controlled within 5mm). The equivalent stress cloud map of the node, the maximum stress value and the stress concentration area were obtained by solving.

[0065] In the structural safety assessment phase, finite element results are integrated with multi-source knowledge: data such as maximum stress value, stress concentration location, and damage degree are input into the large language model along with the "allowable stress threshold for traditional buildings" (e.g., the allowable shear stress of cedar tenon joints is 1.2MPa) from the texture pattern knowledge graph. The model is guided to reason through prompts, such as: "Given that the target traditional building has a maximum stress of 1.5MPa, an allowable stress of 1.2MPa, and a damage degree of moderate (loosening gap of 2mm), please assess the safety level and provide recommendations based on the 'correlation rules between tenon stress exceeding limits and structural failure' in the knowledge graph." The model output needs to refer to historical cases in the knowledge graph (e.g., the safety status of similar traditional buildings with stress exceeding limits after reinforcement), and finally generate a traditional building structural safety diagnosis result that includes the safety level (e.g., Level 1: safe, Level 2: requires monitoring, Level 3: requires reinforcement), risk points (e.g., the stress concentration area is the root of the tenon), and targeted recommendations (e.g., using traditional patching techniques to reinforce the tenon).

[0066] S404. Based on the location diagnosis results of the texture pattern and the traditional building structure safety diagnosis results, a unified coordinate framework is used to perform spatial correlation fusion and to combine the texture pattern knowledge graph to perform multi-source diagnostic information conflict detection and weighted evidence fusion and resolution. The structural safety level is divided according to the preset safety level standard and the temporal evolution sub-vectors in the pattern feature library are combined to generate the texture pattern evolution trend prediction, and a diagnostic result set including the degree of damage, location and structural safety level is obtained.

[0067] First, a unified coordinate framework for spatial association and fusion is implemented: based on the transformation matrix between the WGS84 geographic coordinate system and the local building component coordinate system of the spatiotemporal four-dimensional digital twin model, the orthophoto bounding box coordinates and 3D point cloud coordinates in the repair and positioning diagnostic results are uniformly mapped with the local coordinates of the traditional building mechanics model, ensuring that all diagnostic information is accurately aligned in the spatial dimension. For example, the 3D point cloud coordinates (X=18.5, Y=22.3, Z=0.8) of the brick carving screen wall repair area of ​​component ID: C203 are converted into the global coordinates of the digital twin model, and simultaneously associated with the spatial region index of the building to which the component belongs, achieving a one-to-one correspondence between the location and the structural model.

[0068] Secondly, multi-source diagnostic information conflict detection: The component-texture association rules in the texture and pattern knowledge graph are invoked to cross-validate the repair and positioning results with traditional building safety diagnostic results. If a conflict is detected (e.g., the diagnosis of loose mortise and tenon joints in masonry components, or the stress value of a slightly damaged component exceeding the allowable threshold of 150%), the conflict item is marked and the conflict type (logical conflict / numerical conflict) is recorded, triggering a manual review process; if no conflict is found, the process proceeds to the weighted evidence fusion stage.

[0069] Next, weighted evidence fusion and resolution are performed: weights are assigned based on the credibility of the diagnostic sources, with the damage degree result from the combined diagnosis of hyperspectral imaging and YOLOv7 having a weight of 0.6, and the stress safety assessment result from ANSYS finite element analysis having a weight of 0.4. Using Dempster-Shafer evidence theory, the probability distributions of each source are combined. For example, the probability of the location diagnosis indicating moderate damage (0.52) is 0.8, the probability of finite element analysis supporting this damage degree is 0.7, and the combined probability after fusion is 0.94, thus reducing information redundancy and enhancing the credibility of the results.

[0070] Then, the structural safety level is classified: based on the preset safety level standards (Level 1: stress ≤ 80% of the allowable threshold and damage degree < 0.3; Level 2: stress 80%-100% or damage degree 0.3-0.6; Level 3: stress > 100% or damage degree > 0.6), combined with the fused damage degree and stress data, the structural safety level is automatically determined. For example, if the stress of a certain traditional wooden building after fusion is 110% of the allowable threshold and the damage degree is 0.55, it is determined to be a Level 3 safety building.

[0071] Finally, the evolution trend is predicted: Temporal evolution sub-vectors (64-dimensional features encoded by LSTM) are extracted from the pattern feature library and input into a Transformer-based time series prediction model. Combined with historical environmental monitoring data (such as annual rainfall and average temperature), the model predicts the development trend of the texture pattern over the next 5-10 years. For example, regarding the weathering and flaking pattern of the brick carving in component ID: C203, the model predicts that the flaking area will expand to 12% by 2030, the safety level will drop to level three, and the model outputs a trend curve and early warning information for key time nodes.

[0072] By integrating all the above steps, a standardized diagnostic result set is finally generated, which includes component ID, type, damage level, spatial location, safety level, and evolution trend. This set is stored in a three-level index system of a distributed database.

[0073] As an optional embodiment of the present invention, the generation of the repair solution may include: S501. Based on the texture pattern type, damage degree quantification value and structural safety level in the diagnostic result set, use the process-texture pattern relationship, process-component relationship and process-time relationship in the texture pattern knowledge graph to perform three-level traditional building-process adaptation matching and historical repair compatibility verification, and generate a traditional building-process adaptation mapping table. The first level is a rough matching of traditional building and craftsmanship: based on the traditional building type and component type in the diagnostic result set, it directly matches the pre-associated basic craftsmanship set in the knowledge graph. For example, for traditional buildings with weathered brick and stone walls, it initially matches three types of craftsmanship: "surface cleaning, brick repair, and sealing and reinforcement", and excludes irrelevant crafts such as "wooden mortise and tenon reinforcement" that do not match the traditional building type. The second level is damage adaptation screening: combining the quantitative value of damage degree with the structural safety level, screening options that match the applicable scope of the process. For example, for mild brick carving weathering with a damage degree of less than 0.3, only "desalination cleaning + water-repellent sealing" is retained, while the "partial reconstruction" process that requires chiseling away the remaining parts is excluded. The third level is time sequence compatibility matching: extract the time sequence information of historical repair records in the digital twin model, and verify the compatibility between the proposed process and the existing repair. For example, if the target component has been surface sealed 10 years ago, the sealing material to be selected this time must match the compatibility of the original material. If there is a material compatibility conflict, the corresponding process option will be replaced.

[0074] After completing the three-level matching, all matched processes are mapped one-to-one with their corresponding traditional building components, generating a standardized traditional building-process compatibility mapping table. The historical restoration compatibility check specifically involves: traversing the materials and processes used in existing restoration records, querying the "material-material compatibility rules" through a knowledge graph; if the materials of the newly selected process pose a chemical reaction risk with the original restoration materials (e.g., acidic sealant incompatible with the original alkaline mortar), process replacement is triggered, prioritizing alternative processes that meet compatibility requirements to ensure that the overall restoration does not cause secondary damage to historical relics.

[0075] S502. Based on the traditional building-process adaptation mapping table, candidate solutions are retrieved by utilizing the process-repair scheme relationship in the texture pattern knowledge graph. Material compatibility, mechanical requirements and spatial adaptation are sorted by combining the material spectral characteristics, mechanical requirement parameters and stress concentration area coordinates in the diagnostic result set. Multi-process combination schemes are combined according to process priority to generate a set of candidate repair schemes. The candidate solution retrieval relies on the triplet relation database of the texture and pattern knowledge graph: using the process item in the traditional architecture-process adaptation mapping table as the search key, the edge relationship of "process → repair solution instance" in the knowledge graph is associated to extract all historical repair solutions matching the process (including solution ID, material list, construction process, effect verification data, etc.). For example, for the process combination of "weathering of brick and stone walls - desalination cleaning + water-repellent sealing", similar repair solutions for a brick-carved screen wall in an old city in 2018 (solution ID: F023) and the repair solution for a brick and stone wall of a guild hall in 2020 (solution ID: F047) are retrieved from the knowledge graph to form an initial candidate solution pool.

[0076] Material compatibility ranking needs to be combined with the material spectral characteristics in the diagnostic results: extract the spectral feature sub-vectors of the target component from the diagnostic result set (such as the calcium carbonate content and porosity of bricks and stones, the cellulose residue rate of wood, etc.), query the "compatibility rules of repair materials-original component materials" through the knowledge graph, calculate the compatibility score of the material used in each candidate solution with the original component (maximum score of 100 points, such as the compatibility score of natural tung oil with old fir wood is 92 points, and that of synthetic resin is 68 points), and rank the candidate solutions from high to low scores.

[0077] The ranking of mechanical requirements needs to match the mechanical parameters of the finite element analysis: extract data such as the coordinates of stress concentration areas, maximum stress values, and structural safety levels from the diagnostic results, associate them with the quantitative relationship of "repair process - mechanical enhancement effect" in the knowledge graph (e.g., the mortise and tenon joint filling process can improve shear strength by 30%, and the brick and stone grouting process can improve compressive strength by 25%), calculate the degree of satisfaction of each candidate scheme with mechanical requirements (satisfaction = actual enhancement effect / required enhancement effect × 100%), and sort them in descending order of satisfaction.

[0078] Spatial adaptation ranking should be based on the spatial characteristics of the repair area: using the three-dimensional coordinates and dimensions of the repair area in the unified coordinate framework, combined with the "repair process-construction space requirements" in the knowledge graph (such as the minimum operating space of 0.5m×0.5m for partial reconstruction process and 0.3m×0.3m for surface sealing process), the construction feasibility of each candidate scheme is evaluated and ranked in descending order according to the spatial adaptation score (maximum score of 100 points).

[0079] The process priority combination of multi-process combination schemes: Referring to the "process priority rules" in the knowledge graph (e.g., pretreatment process takes precedence over main body repair process, structural reinforcement process takes precedence over decorative repair process), the sorted single process schemes are combined in a logical order. For example, for a certain traditional wooden building (damage level 0.55, safety level 3), the process order determined after sorting is "cleaning of the decayed area of ​​the wood → tenon patching and reinforcement → water-repellent sealing", which is combined into a complete candidate scheme, and the construction parameters of each process are marked (e.g., the ratio of patching material, curing time of 28 days), material usage (e.g., the size and quantity of patching wood chips), and quality acceptance standards (e.g., the tenon gap after patching ≤ 0.5mm). Finally, a set of 3-5 candidate schemes is generated, each scheme with a detailed process flow, material list and expected effect.

[0080] S503. Based on the candidate repair scheme set, the repair action area is mapped using the building component-level detailed dimension data in the spatiotemporal four-dimensional digital twin base and the positioning coordinate information in the diagnostic result set. A digital twin pre-simulation scenario for the repair operation is constructed. The construction path, material usage, and disturbance range of the candidate schemes are quantitatively simulated, and the disturbance index and process feasibility score of each candidate scheme are output. The large language model is used to perform semantic reasoning and scheme adjustment suggestions on the process conflict points, material adaptation deviations, and structural safety redundancy in the pre-simulation results, and the candidate schemes are iteratively optimized. Based on the screening criteria of the lowest disturbance index, the highest process feasibility, and the achievement of structural safety redundancy, the optimal repair scheme is determined and the repair scheme is output.

[0081] First, the mapping of the repair area needs to be based on the component-level fine modeling data of the spatiotemporal four-dimensional digital twin model: extract the unique ID of the target component, the three-dimensional coordinates and size parameters of the repair area from the diagnostic results set, locate the corresponding three-dimensional model instance of the component through the spatial indexing system of the digital twin model, and then automatically delineate the precise spatial range of the repair operation in the model according to the rules of the scope of each process in the traditional building-process adaptation mapping table (such as the mortise and tenon patching process needs to cover the root of the tenon and the adjacent 5cm area, and the brick and stone sealing process needs to extend to 10cm outside the repair boundary). The material properties (such as wood density, brick and stone porosity) and structural connection relationship (such as the traditional building position with the adjacent beam frame) of the area are associated to form a visual repair area layer.

[0082] Secondly, the construction of a digital twin pre-construction scenario requires the integration of multi-source data: importing the historical repair records of the repair area layer and components, real-time environmental monitoring data (such as current temperature, humidity, and vibration values), and 3D models of construction equipment (such as miniature electric drills and patching tools) into the pre-construction engine to build a 1:1 virtual construction scenario. The scenario must include a dynamic timeline to simulate the entire process from construction preparation to acceptance, while also supporting the step-by-step breakdown and detailed magnification of key procedures.

[0083] The quantitative simulation process specifically includes: Construction path simulation: Based on the operating radius of construction equipment and the space requirements for worker activities, the optimal construction path is automatically planned. For example, in narrow gaps between columns, the side of the component is preferred for patching to avoid obstructing adjacent painted components; Material usage simulation: Based on the volume of the repair area, the material loss rate of the process (e.g., the loss rate of patching wood chips is 15%), and the material density, the precise usage is calculated. For example, if the volume of a decayed tenon area is 0.002m3, 0.0023m3 of patching wood chips is required; Disturbance range simulation: Through the finite element dynamic analysis module, the stress transmission of tool vibration and material installation to surrounding components during construction is simulated, and the radius of disturbance influence (e.g., the vibration disturbance radius of electric drilling is 0.3m) and the stress change value of adjacent components are output. If there are fragile brick carving components within the disturbance range, they are marked as high-risk points.

[0084] The calculation rules for the disturbance index and process feasibility score are as follows: The disturbance index adopts a weighted comprehensive method, with vibration impact (weight 40%), the degree of change of the original structure due to material replacement (weight 30%), and stress changes of surrounding components (weight 30%). The index ranges from 0 to 10, with lower values ​​indicating smaller disturbances. The process feasibility score covers the satisfaction of construction space (weight 30%), material supply cycle (weight 20%), worker skill matching degree (weight 25%), and the compliance rate of quality acceptance standards (weight 25%). The full score is 100 points, and the higher the score, the stronger the feasibility.

[0085] The semantic reasoning and optimization steps of the large language model are as follows: Conflict points from the pre-simulation results (e.g., the distance between the construction path and adjacent components is less than 0.2m), material compatibility deviations (e.g., the moisture content of the proposed patch wood differs from the original wood by more than 5%), and insufficient safety redundancy (e.g., the stress redundancy after reinforcement is only 10%, lower than the required 20%) are input into the model. This is combined with component repair construction specifications from the knowledge graph, historical component material compatibility manuals, and similar repair cases to generate targeted adjustment suggestions. For example, "It is recommended to adjust the construction path to enter from the back of the component, replace the patch wood with old fir wood with a moisture content ≤12%, and increase the patch layer thickness to 1.5cm to improve the stress redundancy to 25%." The model also needs to verify whether the adjusted solution solves the original problem and output the iteratively optimized pre-simulation parameters.

[0086] Optimal Solution Selection Phase: All candidate solutions are comprehensively ranked based on their disturbance index, process feasibility score, and safety redundancy. Solutions with a disturbance index ≤ 3, a process feasibility score ≥ 85, and a safety redundancy ≥ 20% are prioritized. If multiple solutions meet these criteria, their construction period and cost are further compared to determine the single optimal repair solution. The output solution must include a complete construction flowchart, a materials list (including specifications, quantities, and compatibility information), a pre-simulation verification report (including disturbance simulation results and safety redundancy calculations), and quality acceptance standards (e.g., mortise and tenon gap ≤ 0.5mm after patching, and sealing layer adhesion ≥ 0.5MPa).

[0087] As an optional embodiment of the present invention, the formation of the repair strategy library may include: S601. Based on the material ratio parameters, construction sequence, and spatial distribution of disturbance points of each process in the repair plan, material compatibility conflict detection, process sequence constraint verification, and space occupancy analysis are performed using the process-material relationship, process-sequence relationship, and process-space relationship in the texture pattern knowledge graph to generate a conflict detection result set. Material compatibility conflict detection relies on a predefined material-material interaction rule base within the texture pattern knowledge graph. For each step in the repair plan, the chemical properties (such as pH value and composition) and physical characteristics (such as moisture content and expansion coefficient) of the materials used in different steps are extracted. Through a triplet query in the knowledge graph (e.g., "acidic desalination agent → and → alkaline sealing material → presence → salting-out reaction risk"), the system verifies whether adjacent or simultaneously used material combinations exhibit chemical reactions or physical incompatibility. For example, if a repair plan first uses an acidic desalination agent with a pH of 2 to treat masonry components, and then directly applies an alkaline silicone sealing agent with a pH of 10, the system will automatically match the conflict rules in the knowledge graph, marking the combination as a "material compatibility conflict," and recording the conflict type (chemical reaction risk), the involved steps (desalination treatment and sealing reinforcement), and the suggested solution (replacing with a pH-neutral sealing material).

[0088] The process sequence constraint verification needs to be based on the "process process dependency model" in the knowledge graph. This model includes predefined process priority rules (such as "pretreatment process → must precede → main body repair process" and "structural reinforcement process → must precede → decorative repair process") and causal relationships between processes (such as "wood decay cleaning → must be completed before → tenon patching"). The system will traverse the construction sequence of the repair plan and compare it with the constraint rules in the knowledge graph. If a process is found to violate the dependency relationship (such as arranging "water-repellent sealing" before "residual brick repair"), it is determined to be a "process sequence conflict". The system will output the conflict point (process order reversed), the violated constraint rule ("residual brick repair process must be completed before surface sealing") and adjustment suggestions (adjust the residual brick repair process before the water-repellent sealing process).

[0089] Space occupancy analysis requires the use of a spatial collision detection module within a spatiotemporal four-dimensional digital twin model. First, the three-dimensional coordinates of the construction area for each process, the required operating space (such as the minimum operating radius of construction equipment and worker activity space), and the spatial distribution of disturbance points are extracted. Then, the spatial indexing system of the digital twin model is used to analyze whether there is overlap between the construction areas of different processes and whether there is a risk of collision between construction equipment and surrounding components. For example, if a repair plan involves simultaneous mortise and tenon jointing of two adjacent wooden components, and their operating spaces (each requiring a radius of 0.4m) overlap by 0.12m², the system will mark this as a "space occupancy conflict" and output the three-dimensional coordinates of the conflict area, the overlapping area, and optimization suggestions (such as staggering the construction of these two processes or using miniaturized construction tools to reduce the operating space).

[0090] After completing the above three tests, the system classifies all conflict points according to their severity (high, medium, low) and generates a conflict detection result set that includes conflict type, involved process, conflict location, rule violation, and preliminary suggestions.

[0091] S602. Based on the conflict detection result set, use the process-compatibility relationship, process-substitutability relationship and process-temporal flexibility relationship in the texture pattern knowledge graph to assess the severity of conflict and classify the conflict type, and generate conflict classification results. The assessment of conflict severity needs to be conducted in conjunction with a predefined conflict risk weight model in the fabric pattern knowledge graph. This model uses the impact weight of conflict on construction authenticity (40%), structural safety risk weight (35%), and construction feasibility obstacle weight (25%) as core dimensions. By querying the knowledge graph, it retrieves the process-material risk level, process-structural safety correlation, and process-construction constraint strength involved in the conflict, and calculates a comprehensive risk score (out of 100). A score ≥80 indicates high severity, 50-79 indicates medium severity, and <50 indicates low severity. For example, in the material compatibility conflict, "the acidic desalination agent and the alkaline sealing material cause a salt precipitation reaction, resulting in the crumbling of the brick and stone components." Its impact on authenticity scores 90 points, structural safety risk scores 85 points, and construction obstruction scores 80 points, with a total score of 85 points, which is judged as high severity. In the process sequence conflict, "the hydrophobic sealing process precedes the brick repair process." Its impact on authenticity scores 40 points, structural safety risk scores 30 points, and construction obstruction scores 60 points, with a total score of 45 points, which is judged as low severity.

[0092] Conflict type classification needs to be based on the core dimensions of conflict detection and the process relationship types in the knowledge graph: conflicts are divided into three main categories: material compatibility conflicts, process sequence conflicts, and space occupation conflicts, and each category is further subdivided according to severity. Material compatibility conflicts are subdivided into high severity (e.g., chemical reactions causing irreversible damage to components), medium severity (e.g., material performance degradation affecting repair results), and low severity (e.g., slight appearance color difference does not affect structure and authenticity); process sequence conflicts are subdivided into high severity (e.g., delayed structural reinforcement processes leading to safety hazards), medium severity (e.g., unreasonable process sequence extending the construction period), and low severity (e.g., reversed auxiliary process sequence does not affect the core effect); space occupation conflicts are subdivided into high severity (e.g., construction equipment colliding with valuable components), medium severity (e.g., overlapping operating spaces leading to reduced construction efficiency), and low severity (e.g., slight spatial interference can be resolved through tool adjustments).

[0093] When generating conflict classification results, it is necessary to associate the process-substitutability relationship and the process-temporal flexibility relationship in the knowledge graph. For example, for high-severity material compatibility conflicts, if there are compatible alternative materials (such as neutral sealing materials) in the knowledge graph, the result should be marked with "Recommended Replacement Material"; for medium-severity space occupation conflicts, if the process-temporal flexibility relationship allows for staggered construction, the result should be marked with "Temporal Adjustment Recommendation". The final output conflict classification results should include conflict ID, conflict type, severity level, comprehensive risk score, associated process nodes, and optimization direction suggestions.

[0094] S603. Based on the conflict classification results, the process-substitution relationship, process-process reorganization relationship and process-parameter optimization relationship in the texture pattern knowledge graph are used to perform alternative material retrieval and replacement, conflict process time sequence reorganization and disturbance point parameter optimization and adjustment. After global consistency verification and iterative resolution, an optimized repair scheme set is generated. The search and replacement of alternative materials relies on the process-substitution relationship database of the texture pattern knowledge graph for precise execution: For material compatibility conflicts in the conflict classification results, the system extracts the core performance indicators (such as moisture content, pH value, strength grade, and reversibility requirements) and process adaptation conditions (such as material hardness in patching and air permeability in sealing) of the original conflicting materials. Through the triple query of the knowledge graph (such as "acidic desalination agent → alternative material → neutral desalination agent → compliant with → masonry component repair specifications"), a list of alternative materials that meet the requirements of authenticity preservation, structural safety, and process adaptation is obtained. The materials in the list are sorted according to "homogeneity priority (such as old fir wood replacing old pine wood), richness of historical cases, and cost controllability", and the top 3 candidate materials are selected. The construction effect (such as mortise and tenon gap after patching and adhesion of the sealing layer) and long-term stability (such as 5-year aging simulation) of the alternative materials are simulated through digital twin models to verify their feasibility. If the verification is successful, the conflicting materials in the original scheme are replaced, and the material list and process parameters are updated (such as adjusting the thickness of the patch wood chips to match the strength of the alternative materials).

[0095] The reordering of conflicting work processes requires flexible adjustments using a process-work reordering relationship model. For conflicts in process sequence and space occupancy, the system extracts the dependencies between conflicting processes (e.g., "wood preservation treatment → must follow → decay removal"), the flexible time interval (e.g., a process can be delayed by 1-3 days), and spatial resource constraints (e.g., the usage time of construction equipment). Existing heuristic algorithms are used to reorder the process sequence. For example, adjacent component repair processes with overlapping construction can be staggered (e.g., completing the east corridor column patching first, followed by a 24-hour interval to allow the material to cure before proceeding with the west corridor column work), or the process order can be adjusted to meet structural safety priorities (e.g., moving "foundation reinforcement" before "surface decoration repair"). The reordered sequence needs to be imported into a digital twin pre-simulation scenario to simulate the space occupancy and process connection efficiency of the entire construction process, ensuring no new conflict points arise.

[0096] Optimization and adjustment of disturbance point parameters need to be based on precise measures from the process-parameter optimization relationship database: For conflict points where the disturbance index exceeds the standard (such as excessive vibration disturbance radius of construction equipment), the system searches the knowledge graph for parameter adjustment schemes to reduce disturbance (such as replacing the mini electric drill, adjusting the tool vibration frequency to below 500r / min, and optimizing the construction path to enter from the back of the component); the existing finite element dynamic analysis module is used to simulate the disturbance range after parameter adjustment (such as reducing the vibration disturbance radius to within 0.2m) and the stress changes of adjacent components (such as controlling the stress value of brick carving components within the safety threshold) to ensure that the disturbance index is ≤3; at the same time, the construction operation parameters are adjusted (such as controlling the pressure of the patching tool to 0.3MPa) to minimize the impact on surrounding components while ensuring construction quality.

[0097] Global consistency verification needs to cover all dimensions of conflict resolution in the solution: the solution after material replacement, process sequence reorganization and parameter optimization is imported into the conflict detection module for secondary conflict detection to check for new material compatibility, process sequence or space occupation conflicts; if new conflicts exist, the above optimization steps are repeated until all conflicts are resolved; in addition, the overall consistency of the solution also needs to be verified, such as the matching of the bill of materials and process parameters, the rationality of construction sequence and resource allocation (such as the matching of the number of workers and process progress), and the compliance of disturbance index and safety redundancy (such as safety redundancy ≥20%).

[0098] The generation of the iterative resolution and optimization scheme set needs to undergo multiple rounds of verification: After 3-5 rounds of iterative optimization, all repair schemes that meet the criteria of "disturbance index ≤ 3, process feasibility score ≥ 85 points, safety redundancy ≥ 20% and no conflict" are selected to form the optimized repair scheme set; each scheme needs to include an updated construction flow chart (marking the adjusted sequence of procedures), a material list (including compatibility descriptions and usage of alternative materials), a timing adjustment table (clearly defining the execution time of each procedure), a disturbance optimization report (including simulation results), and pre-simulation verification results (including spatial conflict resolution).

[0099] S604. Based on the optimized set of repair schemes, the schemes are classified into different schools of thought and labeled with multi-dimensional attribute tags using the process-school relationship, process-region relationship and process-age relationship in the texture pattern knowledge graph. The schemes are combined with traditional gray plastic repair, reversible reinforcement material repair and structural reinforcement schemes as basic templates to build a strategy index and store it according to a four-level index system to generate a repair strategy library.

[0100] The classification of styles needs to be based on the craft-style relationship model in the texture and pattern knowledge graph: the system extracts the core craft features of each scheme in the optimized scheme set (such as the mortar-piling technique of gray plaster and the mortise and tenon jointing technique of wooden structure reinforcement), regional cultural elements (such as the horse-head wall and the courtyard house style), and the craft style of the era (such as the official practices of the Ming and Qing Dynasties and local characteristics), and matches them with the style tags (Huizhou style, Suzhou style, Shanxi style, Beijing style, etc.) in the knowledge graph. For example, if a scheme uses the "shallow relief mortar-piling" technique of traditional Huizhou gray plaster to repair the damaged part of the horse-head wall, and the process steps conform to the traditional "three mortars and three plasters" process of Huizhou folk gray plaster, then it is classified as "Huizhou gray plaster restoration style"; if the scheme combines the "hidden tenon reinforcement" technique of Shanxi wooden structure to strengthen the beam frame, then it is classified as "Shanxi wooden structure reinforcement style".

[0101] Multi-dimensional attribute labeling needs to integrate process-region, process-era, and core characteristics of the solution: The labeling system includes regional labels (e.g., Huizhou), era labels (e.g., Ming and Qing dynasties), material labels (e.g., traditional tung oil putty, reversible epoxy resin, old cedar wood), type labels (e.g., decayed wooden components, crumbling bricks and stones, peeling plaster, wall cracks), repair type labels (structural reinforcement, decorative repair, moisture-proofing), and reversibility labels (e.g., fully reversible, semi-reversible). For example, for a repair solution for decayed wooden pillars in a Qing dynasty residence on Pingjiang Road in Suzhou, the labeling would be "Suzhou region, Qing dynasty era, old cedar wood material, decay of wooden components, structural reinforcement type, fully reversible".

[0102] The strategy index is built around three basic templates and employs a four-level indexing system: Level 1 is the type of repair (structural reinforcement, decorative repair, comprehensive repair); Level 2 is the regional style (Huizhou style, Suzhou style, Shanxi style, etc.); Level 3 is the material type (traditional materials, reversible modern materials, mixed materials); and Level 4 is the type of damage (sub-types such as wood decay and brick crumbling). For example, the index path for the Huizhou-style traditional plasterwork repair scheme is "decorative repair → Huizhou style → traditional materials → plasterwork detachment"; the index path for the Shanxi-style reversible wood structure reinforcement scheme is "structural reinforcement → Shanxi style → reversible modern materials → wood component decay".

[0103] The repair strategy library is stored using a combination of graph and relational databases: the graph database stores semantic relationships of style classification, tag associations, and index paths, facilitating cross-dimensional retrieval; the relational database stores specific files of optimization schemes (construction flowcharts, material lists, pre-performance reports, etc.) and metadata (scheme ID, applicable scenarios, case source). Simultaneously, the strategy library supports a dynamic update mechanism. When new repair cases or process knowledge are added, the system automatically extracts features and supplements them to the corresponding style, tags, and index paths, ensuring the timeliness and coverage of the strategy library. For example, when a new case of swallowtail ridge plaster restoration of a traditional Minnan house is added, the system automatically classifies it as "Minnan style plaster restoration," labels it with tags such as "Minnan region, Qing Dynasty era," and supplements it under the index path "Decorative Restoration → Minnan → Traditional Materials → Plaster Detachment," enriching the regional coverage of the strategy library.

[0104] Although embodiments of the invention have been shown and described, those skilled in the art will understand that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the claims and their equivalents.

Claims

1. A point-embedded micro-disturbance acupuncture-style protection and restoration method for old town historical sites, characterized in that the method... include: Collect all elements of historical site data, construct a spatiotemporal four-dimensional digital twin base including texture pattern, value carrier, social network, architectural style and historical evolution trajectory, and form a multi-source dataset; Based on multi-source datasets, a texture pattern knowledge graph is constructed using large language models and large visual models. A cross-modal recognition model is trained using a spatiotemporal four-dimensional digital twin base and a texture pattern knowledge graph to generate a pattern feature library; Based on hyperspectral imaging and target detection algorithms, multi-scale diagnosis of the pattern feature database is performed. ANSYS finite element analysis is combined with large model reasoning to evaluate the distribution of value carriers and generate a diagnostic result set containing pattern information and carrier location. A repair plan is generated based on the diagnostic results set; The restoration plans were tested for conflicts and optimized to form a restoration strategy library that includes restoration of the texture and pattern, value carriers, social networks, and architectural style. Based on the selected adaptation scheme in the repair strategy library, the on-site repair work was completed using a point-embedded micro-disturbance construction process.

2. The point-embedded micro-disturbance acupuncture-style protection and restoration method for old town historical sites as described in claim 1, characterized in that, The formation of multi-source datasets includes: Data was collected on the building complexes and components in the historical area to obtain geometric shape and surface texture information at the building component level, and to generate a three-dimensional point cloud model and an initial dataset of orthophotos of the historical area. Data on the material spectral reflectance characteristics and elemental composition of building walls, timber frames and decorative components are collected to generate a material spectral feature library; Based on the 3D point cloud model of historical sites, the data is digitally processed and temporal semantically analyzed using historical documents, old photos, surveying archives and oral history materials to extract the time nodes and change trajectory information of the architectural spatial pattern evolution in various historical periods, and generate a historical evolution trajectory dataset. Based on the 3D point cloud model of historical sites, the initial dataset of orthophotos, the material spectral feature library and the historical evolution trajectory dataset, a unified coordinate framework is used to perform spatiotemporal alignment and semantic association fusion to generate a spatiotemporal registration fusion dataset. Based on the spatiotemporal registration and fusion dataset, a digital twin modeling engine is used to couple the dimensions of building component-level geometric details, material spectral features, historical evolution trajectory, and temporal evolution in four dimensions to generate a spatiotemporal four-dimensional digital twin base containing building component-level details, material spectral features, and historical evolution trajectory, thus forming a multi-source dataset.

3. The point-embedded micro-disturbance acupuncture-style protection and restoration method for old town historical sites as described in claim 1, characterized in that, Constructing a knowledge graph of textural patterns includes: Using a large language model, we can identify textual knowledge entities in historical evolution trajectory datasets from multiple sources and establish semantic relationships between these entities. Using a large visual model, multi-source visual features are extracted from the 3D point cloud model of the historical site, the initial dataset of orthophotos, the component connection method, decorative component pattern features, and material surface process traces in the material spectral feature library. Visual knowledge entities are identified, and spatial topological relationships and process logic associations between visual knowledge entities are established. Cross-modal alignment and fusion of textual knowledge entities, semantic relationships between textual knowledge entities, spatial topological relationships between visual knowledge entities, and process logic associations are performed to establish cross-modal semantic mapping between visual knowledge entities and textual knowledge entities, thereby generating a texture pattern knowledge graph.

4. The point-embedded micro-disturbance acupuncture-style protection and restoration method for old town historical sites as described in claim 1, characterized in that, The generated pattern feature library includes: By utilizing the process-component relationship, process-texture pattern relationship, material-texture pattern relationship, and temporal causal relationship in the texture pattern knowledge graph, multimodal texture pattern semantic guidance annotation is performed on visual data, spectral data, and temporal data in the spatiotemporal four-dimensional digital twin base. This generates a multimodal texture pattern annotation dataset that includes a visual texture pattern annotation dataset, a spectral texture pattern annotation dataset, and a temporal causal texture pattern annotation dataset. Based on a multimodal texture pattern annotation dataset, a visual feature encoding stream is constructed using a large visual model. A semantic knowledge encoding stream is constructed by embedding process nodes, material nodes, process nodes, and texture pattern nodes in the texture pattern knowledge graph using a graph neural network. A cross-modal fusion module based on a cross-attention mechanism and a multi-task output head including texture pattern classification, damage degree regression, position coordinate output, and structural safety level assessment are constructed to generate a cross-modal texture pattern recognition model. Based on a multimodal texture pattern annotation dataset and a cross-modal texture pattern recognition model, a multi-stage joint training strategy is used to sequentially perform single-modal pre-training of visual feature encoding stream, pre-training of knowledge graph link prediction of semantic knowledge encoding stream, cross-modal alignment training of cross-modal fusion module, and end-to-end joint fine-tuning including knowledge regularization loss based on logical consistency constraints of texture pattern knowledge graph, thereby generating a trained cross-modal texture pattern recognition model. Based on the trained cross-modal texture pattern recognition model, a full texture pattern reasoning is performed using a spatiotemporal four-dimensional digital twin basis and cross-validated by combining a texture pattern knowledge graph. A structured texture pattern description is generated through large language model reasoning. The structured texture pattern description is encoded into a five-dimensional joint feature vector containing visual feature sub-vectors, spectral feature sub-vectors, semantic knowledge sub-vectors, temporal evolution sub-vectors, and structural safety sub-vectors. It is stored according to a three-level indexing system of building components-spatial location-temporal evolution to generate a pattern feature library.

5. A point-embedded micro-disturbance acupuncture-style protection and restoration method for old town historical sites as described in claim 1, characterized in that, The generated diagnostic result set includes information on the extent, location, and structural safety level of the damage. Based on the spectral feature sub-vectors and location coordinate information in the pattern feature library, the surface of building components is scanned in multiple bands using a hyperspectral imaging device. Combined with the texture pattern sensitive bands in the pattern feature library, multi-scale pyramid decomposition and damage degree weighting are performed to generate a multi-scale hyperspectral diagnostic image set. Based on visual feature subvectors in a multi-scale hyperspectral diagnostic image set and a pattern feature library, a target detection algorithm finely tuned by transfer learning from the pattern feature library is used to perform four-level scale texture pattern localization and recognition and damage degree regression quantitative assessment, generating texture pattern localization diagnostic results. Based on the location coordinates and damage information of the traditional wooden building in the texture pattern location diagnosis results, the ANSYS finite element analysis is used to construct the mechanical model of the traditional building and apply the material properties and damage degradation parameters in the material spectral feature library and the combined load conditions to solve the stress distribution. The structural safety assessment is carried out by combining the large language model reasoning and the process-component relationship in the texture pattern knowledge graph, and the structural safety diagnosis results of the traditional building are generated. Based on the results of the texture pattern location diagnosis and the traditional building structure safety diagnosis, a unified coordinate framework is used to perform spatial correlation fusion and to combine the texture pattern knowledge graph to perform multi-source diagnostic information conflict detection and weighted evidence fusion and resolution. The structural safety level is divided according to the preset safety level standard and the texture pattern evolution trend prediction is generated by combining the temporal evolution sub-vectors in the pattern feature library. A diagnostic result set including damage degree, location and structural safety level is obtained.

6. The point-embedded micro-disturbance acupuncture-style protection and restoration method for old town historical sites as described in claim 1, characterized in that, The generated repair plan includes: Based on the texture pattern type, damage degree quantification value and structural safety level in the diagnostic result set, the three-level traditional building-process adaptation matching and historical repair compatibility verification are carried out by using the process-texture pattern relationship, process-component relationship and process-time relationship in the texture pattern knowledge graph, and a traditional building-process adaptation mapping table is generated. Based on the traditional building-process adaptation mapping table, candidate solutions are retrieved by utilizing the process-repair scheme relationship in the texture pattern knowledge graph. The material spectral characteristics, mechanical requirement parameters and stress concentration area coordinates in the diagnostic results set are combined to sort the material compatibility, mechanical requirements and spatial adaptation. Multi-process combination schemes are combined according to the priority of the process to generate a set of candidate repair schemes. Based on the candidate repair scheme set, the repair action area is mapped using the building component-level detailed dimension data in the spatiotemporal four-dimensional digital twin base and the positioning coordinate information in the diagnostic result set. A digital twin pre-simulation scenario for the repair operation is constructed, and the construction path, material usage, and disturbance range of the candidate schemes are quantitatively simulated. The disturbance index and process feasibility score of each candidate scheme are output. A large language model is used to perform semantic reasoning and generate scheme adjustment suggestions for process conflict points, material adaptation deviations, and structural safety redundancy in the pre-simulation results, and the candidate schemes are iteratively optimized. Based on the screening criteria of the lowest disturbance index, the highest process feasibility, and the achievement of structural safety redundancy, the optimal repair scheme is determined and output.

7. The point-embedded micro-disturbance acupuncture-style protection and restoration method for old town historical sites as described in claim 1, characterized in that, A repair strategy library has been developed, encompassing traditional plaster repair, reversible reinforcement material repair, and structural strengthening solutions. Based on the material ratio parameters, construction sequence, and spatial distribution of disturbance points of each process in the repair plan, material compatibility conflict detection, process sequence constraint verification, and space occupancy analysis are performed using the process-material relationship, process-time relationship, and process-space relationship in the texture pattern knowledge graph, generating a conflict detection result set. Based on the conflict detection result set, the conflict severity is assessed and the conflict type is classified by utilizing the process-compatibility relationship, process-substitutability relationship and process-temporal flexibility relationship in the texture pattern knowledge graph, and a conflict classification result is generated. Based on the conflict classification results, the process-substitution relationship, process-process reorganization relationship and process-parameter optimization relationship in the texture pattern knowledge graph are used to perform alternative material retrieval and replacement, conflict process time sequence reorganization and disturbance point parameter optimization and adjustment. After global consistency verification and iterative resolution, an optimized set of repair schemes is generated. Based on the optimized set of repair schemes, the schemes are categorized by style, region, and age using the texture pattern knowledge graph. The schemes are labeled with multi-dimensional attribute tags and combined with traditional gray plastic repair, reversible reinforcement material repair, and structural reinforcement schemes as basic templates to build a strategy index and store it according to a four-level index system, thus generating a repair strategy library.

Images