Intelligent evaluation method for historical building value based on multi-source data fusion

Abstract

The present application relates to the technical field of historical building digital evaluation, in particular to a historical building value intelligent evaluation method based on multi-source data fusion, comprising: preprocessing the multi-modal information of the historical building and generating a fusion feature vector, separating the spatial and time characteristic components, and respectively analyzing the spatial value contribution vector and the time value contribution vector through a value decoupling network. The two vectors are subjected to antagonistic interactive fusion and iterative optimization to obtain a stable value representation vector. The vector and the value anchor point set are subjected to graph structure similarity measurement, and a quantitative correlation matrix is output. The evaluation path is extracted to perform reverse tracing on the knowledge graph to correct the representation vector and generate a final value embedding vector. Finally, the evaluation value is calculated according to the projection coordinates of the vector in the value evaluation benchmark space. The method realizes fine-grained analysis and traceable dynamic optimization evaluation of the value of historical buildings.

Description

Technical Field

[0001] This invention relates to the field of digital assessment technology for historical buildings, and in particular to an intelligent assessment method for the value of historical buildings based on multi-source data fusion. Background Technology

[0002] Current historical building valuation techniques largely rely on expert experience and single-dimensional data, such as survey drawings or archival documents, resulting in a highly subjective and inefficient process. With the development of multi-source data acquisition technologies, methods have emerged that integrate multimodal information such as text, images, and point clouds. However, these methods typically involve simply splicing or weighting information from different dimensions, failing to effectively analyze and separate the independent contributions of the building's spatial attributes and its historical evolution to the overall value. This mixture of characteristics leads to poor interpretability of the valuation model, making it difficult to accurately quantify the individual impacts of spatial structural features and the value accumulated over time.

[0003] Existing intelligent evaluation models mostly employ an end-to-end, one-way reasoning process, directly mapping input data to evaluation results. This approach heavily relies on the quality and completeness of training data; when faced with sparse data or architectural cases with rare features, the stability and reliability of the evaluation results significantly decrease. Lacking a mechanism for dynamically verifying and correlating intermediate evaluation results with external structured knowledge systems, the models struggle to leverage existing value perception systems within the domain to retrospectively validate, refine, and optimize their judgments, thus limiting the robustness and authority of the evaluation system. Summary of the Invention

[0004] The purpose of this invention is to address the shortcomings of existing technologies by proposing an intelligent assessment method for the value of historical buildings based on multi-source data fusion.

[0005] To achieve the above objectives, the present invention adopts the following technical solution: an intelligent assessment method for the value of historical buildings based on multi-source data fusion, comprising:

[0006] Preprocessing is performed on the original multimodal information of historical buildings to separate structured value description information and unstructured perception information, and generate multi-source fusion feature vectors;

[0007] Spatial structure feature components and temporal evolution feature components are separated from the multi-source fusion feature vector. Using a pre-defined value decoupling network, the spatial structure feature components and temporal evolution feature components are decoupled and analyzed to obtain spatial value contribution vector and temporal value contribution vector.

[0008] The spatial value contribution vector and the temporal value contribution vector are fused adversarially, and the fusion result is converged through iterative optimization to obtain a stable value representation vector.

[0009] The stable value representation vector and the set of value anchors are measured by graph structure similarity, and a quantitative correlation matrix is ​​output.

[0010] The value assessment path that meets the confidence requirement is extracted from the quantitative correlation matrix. The knowledge graph is then traced back to its source based on the value assessment path to correct the stable value representation vector and generate the final value embedding vector.

[0011] The final intelligent assessment value of a historical building is calculated based on the projection coordinates of the value embedding vector in the preset value assessment benchmark space.

[0012] As a further aspect of the present invention, the preprocessing of the original multimodal information of historical buildings to separate structured value description information from unstructured perceived information specifically includes:

[0013] Receive raw multimodal information from different acquisition devices, wherein the raw multimodal information includes at least architectural drawing text, sensor monitoring time-series signals, historical document records, and image scan point cloud data;

[0014] A parsing pipeline is constructed for different modal information, which includes a text parser, a signal feature extractor, an optical character recognition engine, and a point cloud registration module;

[0015] Text and image information are processed by a text parser and an optical character recognition engine to extract field values ​​with clear field definitions and logical relationships, and the field values ​​and their relationships are integrated into structured value description information.

[0016] The signal feature extractor and point cloud registration module process the time-series signal and spatial point cloud data, extract the original signal waveform feature sequence and three-dimensional geometric feature point set, and integrate the signal waveform feature sequence and three-dimensional geometric feature point set into unstructured sensing information.

[0017] As a further aspect of the present invention, the generation of the multi-source fusion feature vector includes:

[0018] Information fingerprints are generated from structured value description information and unstructured perceived information, respectively.

[0019] Based on the information fingerprint of structured value description information, retrieve the set of associated value anchors in the knowledge graph;

[0020] Semantic space mapping is performed on unstructured perceptual information to generate a sequence of perceptual feature vectors with temporal correlation;

[0021] Based on the set of value anchors, temporal feature enhancement and dimension alignment operations are performed on the perceptual feature vector sequence to generate multi-source fusion feature vectors;

[0022] The process of generating information fingerprints from structured value description information and unstructured perceived information specifically includes:

[0023] A preset hash algorithm is used to process the integrated structured value description information to generate a structured information fingerprint that represents its global information content;

[0024] A pre-defined deep feature encoder is used to encode the integrated unstructured perceptual information, and the dimensionality reduction representation of the encoded feature vector is used as its perceptual information fingerprint.

[0025] As a further aspect of the present invention, the step of retrieving a set of associated value anchor points in a knowledge graph based on the information fingerprint of structured value description information includes:

[0026] Construct a domain knowledge graph that includes historical building entities, attributes, value dimensions, and historical event nodes. The edges in the knowledge graph represent the relationships or attribute affiliations between entities.

[0027] The similarity between the structured information fingerprint and the identifier fingerprint of all entities in the knowledge graph is calculated, and multiple initial associated entities with similarity exceeding the preset entity matching threshold are selected.

[0028] Starting from the multiple initial associated entities, a multi-hop traversal is performed along the relationship edges in the knowledge graph to explore all value dimension nodes and historical event nodes that have direct or indirect connection relationships with the multiple initial associated entities within a preset number of hops.

[0029] Collect all paths traversed during the traversal process and mark the endpoint of each path as a candidate value anchor.

[0030] Calculate the semantic relevance score between each candidate value anchor and the structured information fingerprint, and filter out candidate value anchors whose semantic relevance scores are lower than the preset anchor screening threshold.

[0031] The remaining candidate value anchors after filtering, along with their local topological structures in the knowledge graph, are used together as the retrieved value anchor set.

[0032] As a further aspect of the present invention, the step of performing semantic space mapping on unstructured perceptual information to generate a sequence of perceptual feature vectors with temporal correlation includes:

[0033] The signal waveform feature sequence in unstructured sensing information is divided into multiple signal segments by an equally spaced sliding window.

[0034] Spatial voxelization meshing is performed on the set of three-dimensional geometric feature points in unstructured perceptual information, and the statistical distribution characteristics of the internal points of each voxel mesh are calculated.

[0035] Each signal segment is input into the temporal feature coding network, and the output is the temporal coding vector of each signal segment;

[0036] The statistical distribution characteristics of each voxel grid are input into the spatial feature encoding network, and the spatial encoding vector of each voxel grid is output.

[0037] According to the timestamp order of signal acquisition, the time-series coding vector is concatenated with the mean vector of all spatial coding vectors within the corresponding time window to form the preliminary fusion vector at each time point;

[0038] Arrange the preliminary fusion vectors from all time points in chronological order to form a preliminary perception feature vector sequence;

[0039] The initial perceptual feature vector sequence is input into a recurrent neural network to capture the dependencies between adjacent vectors in the sequence. The hidden state sequence output by the recurrent neural network is the perceptual feature vector sequence with temporal correlation.

[0040] The step of performing temporal feature enhancement and dimension alignment operations on the perceptual feature vector sequence based on the value anchor point set to generate a multi-source fusion feature vector specifically includes:

[0041] Extract the attribute description text of each value anchor node from the set of value anchors, and convert the attribute description text into an anchor feature vector through a text encoder;

[0042] Calculate the self-attention weight distribution of each temporal vector in the perceptual feature vector sequence, and perform a weighted summation of all vectors in the sequence based on the weight distribution to generate a global perceptual context vector;

[0043] The dot product similarity of each anchor feature vector with the global perception context vector is calculated to obtain the attention weight of each anchor to the global perception.

[0044] Anchor feature vectors are weighted and fused using attention weights to generate knowledge guidance vectors;

[0045] The knowledge-guided vector is multiplied element-wise with each temporal vector in the perceptual feature vector sequence to enhance the temporal features of the perceptual feature vector sequence, thus obtaining the knowledge-enhanced perceptual sequence.

[0046] All vectors in the knowledge-enhanced perception sequence are averaged over the time dimension to generate a multi-source fusion feature vector with a unified dimension.

[0047] As a further aspect of the present invention, the step of separating the spatial structure feature component and the temporal evolution feature component from the multi-source fusion feature vector, and using a preset value decoupling network to perform decoupling analysis on the spatial structure feature component and the temporal evolution feature component respectively, to obtain the spatial value contribution vector and the temporal value contribution vector, includes:

[0048] Construct a value decoupling network, which includes a shared encoder, a spatial decoupling branch, and a temporal decoupling branch;

[0049] The multi-source fusion feature vector is input into the shared encoder to obtain high-order fusion features;

[0050] High-order fusion features are simultaneously input into the spatial decoupling branch and the temporal decoupling branch;

[0051] The spatial decoupling branch contains multiple spatial convolutional layers, which are used to extract spatial structural feature components related to building structure, material and spatial layout from high-order fusion features, and map the spatial structural feature components into spatial value contribution vectors.

[0052] The time decoupling branch contains multiple recurrent connection layers, which are used to extract time evolution feature components related to building age, historical changes, and maintenance records from high-order fusion features, and map the time evolution feature components into time value contribution vectors.

[0053] During the training of the value decoupling network, the lower bound of mutual information between the spatial value contribution vector and the temporal value contribution vector is maximized, and the distance between the spatial structure feature component and the temporal evolution feature component in a common latent space is minimized, so as to enable the network to decouple its independent components.

[0054] As a further aspect of the present invention, the step of adversarially fusing the spatial value contribution vector and the temporal value contribution vector, and then iteratively optimizing the fusion result to obtain a stable value representation vector, includes:

[0055] Initialize a generator network and a discriminator network. Concatenate the spatial value contribution vector and the temporal value contribution vector and input them into the generator network to generate a preliminary fusion vector.

[0056] The initial fusion vector, the original spatial value contribution vector, and the original temporal value contribution vector are respectively input into the discriminator network. The discriminator network is trained to distinguish whether the initial fusion vector is a "generated sample" from the generator or a "real sample" from the real distribution.

[0057] Among them, the real samples were obtained by sampling the known value representation vectors of a large number of historical buildings;

[0058] The discriminator network is trained to determine whether the initial fused vector contains complete spatial structure information and temporal evolution information. Its loss function includes adversarial loss and reconstruction loss.

[0059] The generator network is trained to generate a fusion vector that can deceive the discriminator network. The goal is to make it difficult for the discriminator network to distinguish between generated samples and real samples, and to minimize the errors between the spatial and temporal components of the generated fusion vector reconstructed in the discriminator network and the original input vector.

[0060] Through multiple iterations of adversarial training between the generator network and the discriminator network, the training process stops when the discriminator network cannot effectively distinguish between generated samples and real samples, and the reconstruction error is lower than a preset threshold.

[0061] The outputs of the finally trained generator network on the spatial value contribution vector and the temporal value contribution vector are used as stable value representation vectors.

[0062] As a further aspect of the present invention, the step of measuring the graph structure similarity between the stable value representation vector and the set of value anchor points, and outputting a quantified correlation matrix, includes:

[0063] Extract the adjacency subgraph of each anchor node in the set of value anchors from the knowledge graph, and convert each adjacency subgraph into a corresponding graph embedding vector;

[0064] Calculate the cosine similarity between the stable value representation vector and each graph embedding vector to obtain a preliminary one-dimensional list of associations;

[0065] Pair all anchor nodes in the value anchor set together, calculate the shortest path length of each pair of anchor nodes in the knowledge graph, and form a path distance matrix between anchor nodes.

[0066] The path distance matrix between anchor points is converted into a relationship strength matrix between anchor points using a preset kernel function. The relationship strength matrix between anchor points characterizes the degree of prior association between anchor points.

[0067] The initial correlation list is expanded into a diagonal matrix, and matrix multiplication is performed between it and the relationship strength matrix between anchor points to spread the correlation between the stable value representation vector and a single anchor point to its associated neighbors in the knowledge graph.

[0068] The matrix obtained from the operation is normalized to obtain the final quantified correlation matrix. Each element in the matrix represents the correlation quantification value of each pair of anchor points jointly represented by the stable value representation vector and the knowledge graph relationship correction.

[0069] As a further aspect of the present invention, the steps of extracting value assessment paths that meet the confidence requirements from the quantified correlation matrix, performing reverse tracing of the knowledge graph based on the value assessment paths, correcting the stable value representation vector, and generating the final value embedding vector include:

[0070] Set a confidence threshold, and filter out all matrix elements in the quantized correlation matrix that are greater than the confidence threshold;

[0071] Each selected matrix element corresponds to a pair of anchor nodes in the knowledge graph. A pair of anchor nodes in the knowledge graph and their correlation values ​​in the quantized correlation matrix are recorded as a candidate correlation.

[0072] Based on the knowledge graph, all anchor nodes involved in the candidate associations are taken as necessary nodes. The optimal subgraph path connecting the necessary nodes is searched. The optimal subgraph path makes the comprehensive score of all associations on the path the highest. This optimal subgraph path is the value assessment path.

[0073] In the knowledge graph, the value assessment path is traced backward from the terminal node to the starting node to extract the attribute and relationship information of all nodes and edges on the path and generate a path description vector.

[0074] The path description vector and the stable value representation vector are input into an attention fusion module, which calculates the attention weight of each part of the path description vector to the stable value representation vector.

[0075] Based on attention weights, the path description vector and the stable value representation vector are weighted and fused, and the stable value representation vector is supplemented and corrected to obtain the final value embedding vector.

[0076] As a further aspect of the present invention, the step of calculating the final intelligent assessment value of a historical building based on the projection coordinates of the value embedding vector in a preset value assessment benchmark space includes:

[0077] A value assessment benchmark space is pre-constructed, which is spanned by multiple orthogonal value benchmark axes. Each value benchmark axis represents a core value dimension that has been defined and standardized by experts.

[0078] The final value embedding vector is projected onto each value reference axis, the projection length is calculated, and the projection coordinate components of the value embedding vector on each value reference axis are obtained.

[0079] Each value benchmark axis is associated with a preset weight coefficient, which reflects the importance of the core value dimension in the overall evaluation;

[0080] The weighted value score of the core value dimension is obtained by multiplying the projected coordinate components of the value embedding vector on each value reference axis with the weight coefficient corresponding to the value reference axis.

[0081] The weighted value scores corresponding to all value benchmark axes are summed to obtain the preliminary comprehensive assessment score of the historical building;

[0082] Retrieve from the knowledge graph several reference buildings that are most similar to the current historical building in terms of type, region, and age, and obtain the existing authoritative assessment value of the reference buildings;

[0083] The relative deviation between the preliminary comprehensive assessment score of the current historical building and the average value of the authoritative assessment of the reference building is calculated. The preliminary comprehensive assessment score is calibrated based on the relative deviation. The calibrated result is the final intelligent assessment value of the historical building.

[0084] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0085] By independently decoupling the separated spatial structural feature components and temporal evolution feature components, value contribution vectors specific to the spatial and temporal dimensions are obtained. This method solves the problem of mutual interference between value signals of different attributes in traditional multi-source feature mixing processing, and achieves a clear separation and quantitative expression of the material composition value and historical layered value of the building itself. This process enables the evaluation model to identify and track the specific impact of static attributes such as spatial form and construction techniques, as well as dynamic processes such as historical events and functional evolution, on the comprehensive value, thereby improving the fineness and interpretability of value causal analysis.

[0086] This method employs graph structure similarity measurement between intermediate value representations and external value anchor sets, and uses the measurement results to drive reverse tracing of the knowledge graph, thereby revising the value representations. It constructs a dynamic feedback optimization loop that incorporates domain prior knowledge. Leveraging the structured and networked domain cognition within the knowledge graph, it performs similarity association, logical path tracing, and verification on the value judgments generated by the model itself. This process enhances the model's robustness in reasoning when facing sparse data or outlier cases, ensuring that the evaluation conclusions are consistent with the existing value cognition system, and improving the credibility and authority of the evaluation results. Attached Figure Description

[0087] Figure 1 This is a flowchart of the intelligent assessment method for the value of historical buildings based on multi-source data fusion as described in this invention;

[0088] Figure 2 A flowchart for preprocessing raw multimodal information;

[0089] Figure 3A flowchart for retrieving a set of value anchor points;

[0090] Figure 4 A heat map showing the strength of the relationship between anchor points;

[0091] Figure 5 A radar chart for assessing the value of historical buildings. Detailed Implementation

[0092] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0093] In the description of this invention, it should be understood that the terms "length," "width," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "top," "bottom," "inner," and "outer," etc., indicating orientation or positional relationships, are based on the orientation or positional relationships shown in the accompanying drawings and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, in the description of this invention, "a plurality of" means two or more, unless otherwise explicitly specified.

[0094] See Figure 1 The system preprocesses the original multimodal information of historical buildings, separating structured value description information from unstructured perceptual information to generate multi-source fusion feature vectors. From these feature vectors, spatial structural feature components and temporal evolution feature components are separated. Using a pre-defined value decoupling network, these components are decoupled and analyzed to obtain spatial value contribution vectors and temporal value contribution vectors. These vectors are then adversarially fused, and iterative optimization is used to converge the fusion result, yielding stable value representation vectors. A graph structure similarity measurement is performed between the stable value representation vectors and the set of value anchor points, outputting a quantified correlation matrix. Value assessment paths meeting confidence requirements are extracted from the quantified correlation matrix. These paths are then used to reverse-engineer the knowledge graph, correcting the stable value representation vectors and generating the final value embedding vector. Finally, the final intelligent evaluation value of the historical building is calculated based on the projection coordinates of the value embedding vectors in a pre-defined value assessment benchmark space.

[0095] See Figure 2In one embodiment of the present invention, the raw multimodal information is received from different acquisition devices. This raw multimodal information includes at least architectural drawing text, sensor monitoring time-series signals, historical document records, and image scan point cloud data. The architectural drawing text is a digital design drawing file containing annotations and dimensional information of building components. The sensor monitoring time-series signals are continuous monitoring data streams from vibration, temperature, and humidity sensors installed on the building structure. The historical document records are scanned copies of archival documents recording building changes, repair records, and important events. The image scan point cloud data is a dense point cloud set obtained by three-dimensional scanning of the building's exterior facade and internal structure using LiDAR.

[0096] A parsing pipeline for different modalities of information was constructed, comprising a text parser, a signal feature extractor, an optical character recognition engine, and a point cloud registration module. The text parser parses structured character information from architectural drawings and historical documents; the signal feature extractor extracts raw waveform features from sensor-monitored time-series signals; the optical character recognition engine identifies and transcribes text content from historical document images; and the point cloud registration module aligns image scan point cloud data from different scanning sites to generate a complete 3D point cloud model. By processing text and image information through the text parser and optical character recognition engine, field values ​​with clear definitions and logical relationships are extracted, such as building name, construction year, structural type, and conservation level. These field values ​​and their predefined relationships are integrated into structured value description information. The signal feature extractor and point cloud registration module process the time-series signal and spatial point cloud data to extract the original signal waveform feature sequence and three-dimensional geometric feature point set. The signal waveform feature sequence is the spectrum amplitude sequence of the sensor-monitored time-series signal after Fourier transform. The three-dimensional geometric feature point set is the three-dimensional coordinates and normal vector of each point in the registered image scanning point cloud data. The signal waveform feature sequence and the three-dimensional geometric feature point set are integrated together into unstructured sensing information.

[0097] Information fingerprints are generated for both structured value description information and unstructured perceptual information. A pre-defined SHA-256 hash algorithm is used to process the integrated structured value description information. The input to the hash algorithm is a structured string of key-value pairs, and the output is a fixed-length 256-bit hash value. This hash value is used as the structured information fingerprint representing the global information content of the structured value description information. A pre-defined deep feature encoder is used to encode the integrated unstructured perceptual information. The deep feature encoder is a pre-trained convolutional autoencoder network. The encoding process involves inputting a signal waveform feature sequence and a set of three-dimensional geometric feature points into the network. The output of the intermediate layers of the network is a 512-dimensional feature vector. Principal component analysis is performed on this 512-dimensional feature vector to reduce its dimensionality to 128 dimensions. This 128-dimensional vector is used as its perceptual information fingerprint. Based on the information fingerprint of the structured value description information, a set of associated value anchor points is retrieved from the knowledge graph. Semantic space mapping is performed on the unstructured perceptual information to generate a sequence of perceptual feature vectors with temporal correlation. Based on the set of value anchors, temporal feature enhancement and dimensional alignment operations are performed on the perceived feature vector sequence to generate a multi-source fused feature vector. In one example, for a brick-and-wood structure built during the Republic of China period, the structured value description information might include "construction year: 1920", "structure type: brick-and-wood", and "protection level: municipal-level cultural relic protection unit". The unstructured perceived information might include the floor vibration spectrum sequence recorded by vibration sensors and a set of point cloud geometric features reflecting the degree of wall tilt. The structured information fingerprint is used to locate relevant entity nodes such as "brick-and-wood structure" and "Republic of China building" in the knowledge graph. The perceived information fingerprint is mapped to a temporal feature sequence reflecting the structural health status of the building. Based on the set of value anchors retrieved from the knowledge graph, such as nodes like "structural stability" and "typical techniques of historical periods", the perceived feature vector sequence is enhanced and adjusted, ultimately generating a unified feature vector that integrates multi-dimensional information.

[0098] See Figure 3In one embodiment of the present invention, a domain knowledge graph is pre-constructed. The knowledge graph includes historical building entities, building attributes, value dimensions, and historical event nodes. Edges in the knowledge graph represent relationships or attribute attribution relationships between entities. Historical building entity nodes represent specific individual buildings, building attribute nodes represent features such as "construction style" and "structural materials," value dimension nodes represent abstract dimensions such as "historical value," "scientific value," and "artistic value," and historical event nodes represent temporal events such as "renovation records" and "functional changes." The structured information fingerprint is compared with the identifier fingerprints of all entities in the knowledge graph for similarity calculation. The identifier fingerprint is a hash value generated by each entity node in the knowledge graph based on its attribute text. Jaccard similarity is used for the calculation, and multiple initial associated entities with similarity exceeding a preset entity matching threshold are selected. In a specific implementation, for a building whose structured information fingerprint originates from the description "brick and wood structure, hip roof, Republican era style," the calculation finds that it has the highest similarity to the identifier fingerprints of the three entity nodes "brick and wood structure building," "hip roof architectural style," and "Republican era architectural style" in the knowledge graph. These three entities become the initial associated entities.

[0099] Starting with multiple initially associated entities, the system performs a multi-hop traversal along the relationship edges in the knowledge graph, exploring all value dimension nodes and historical event nodes that have direct or indirect connections with these entities within a preset number of hops. The preset number of hops can be set to 3. Starting from the "brick and wood structure building" entity, a one-hop relationship leads to the "timber frame craftsmanship" attribute node, and a two-hop relationship further connects to the "scientific value (structural research)" value dimension node. All paths traversed during the traversal are collected, and the endpoint node of each path is marked as a candidate value anchor. For example, the endpoint node of the aforementioned path, "scientific value (structural research)," is marked as a candidate value anchor. The semantic relevance score between each candidate value anchor and the structured information fingerprint is calculated. The calculation process vectorizes the description text of the candidate value anchor and the source description text of the structured information fingerprint, then calculates the cosine similarity between the vectors as the score. Candidate value anchors with semantic relevance scores below a preset anchor selection threshold are filtered out. For example, the "artistic value (carved decoration)" node may have low semantic relevance to the structural description of the current building and is therefore filtered out. The remaining candidate value anchors after filtering, along with their local topological structures in the knowledge graph, are used together as the retrieved value anchor set. The value anchor set includes not only the node itself, but also the node's connection relationships with neighboring nodes in the knowledge graph.

[0100] The signal waveform feature sequence in the unstructured sensing information is segmented into equally spaced sliding windows, each with a length of 256 sampling points and an overlap of 50 sampling points between adjacent windows, resulting in multiple signal segments. The 3D geometric feature point set in the unstructured sensing information is then spatially divided into voxel grids of 0.1m × 0.1m × 0.1m size. Statistical distribution characteristics of the points within each voxel grid are calculated, including the mean number of points, the standard deviation of the coordinates, and the average direction of the normal vector. Each signal segment is input into a temporal feature encoding network (a one-dimensional convolutional neural network), which outputs a 128-dimensional temporal encoding vector for each segment. The statistical distribution characteristics of each voxel grid are then input into a spatial feature encoding network (a fully connected multilayer perceptron), which outputs a 64-dimensional spatial encoding vector for each voxel grid. Following the timestamp order of signal acquisition, the temporal encoding vector is concatenated with the mean vector of all spatial encoding vectors within the corresponding time window (the same time period from which the signal segments were acquired), forming a preliminary 192-dimensional fusion vector for each time point. The initial fused vectors from all time points are arranged in chronological order to form an initial perceptual feature vector sequence. This initial perceptual feature vector sequence is then input into a recurrent neural network (RNN). The RNN employs a long short-term memory (LSTM) network structure to capture the dependencies between adjacent vectors in the sequence. The hidden state sequence output by the LSM network is the perceptual feature vector sequence with temporal correlation.

[0101] The attribute description text for each value anchor node is extracted from the set of value anchor nodes. For example, the attribute description text for the value anchor node "Scientific Value (Structural Research)" might be "Reflecting the typical practices and technological level of the timber frame stress system during a specific historical period." This attribute description text is then converted into anchor feature vectors using a pre-trained BERT model via a text encoder. The self-attention weight distribution for each temporal vector in the perceived feature vector sequence is calculated. The self-attention mechanism calculates attention weights within the sequence of temporal vectors and performs a weighted summation of all vectors within the sequence based on the weight distribution, generating a global perceived context vector. This global perceived context vector is a single vector that summarizes the global information of the entire temporal sequence. The dot product similarity between each anchor feature vector and the global perceived context vector is calculated to obtain the attention weight of each anchor point to the global perception. Finally, the anchor feature vectors are weighted and fused using these attention weights to generate a knowledge-guided vector. The knowledge-guided vector is element-wise multiplied with each temporal vector in the perceptual feature vector sequence to enhance the temporal features of the perceptual feature vector sequence, resulting in a knowledge-enhanced perceptual sequence. Each vector in the knowledge-enhanced perceptual sequence incorporates prior information from the value dimension of the knowledge graph. In essence, all vectors in the knowledge-enhanced perceptual sequence undergo average pooling in the temporal dimension to generate a unified-dimensional multi-source fusion feature vector. The average pooling operation takes the average of all temporal vectors in the temporal dimension, resulting in a fixed-length vector as the multi-source fusion feature vector.

[0102] In some embodiments, the attention weight of the anchor point to global perception The calculation formula is:

[0103]

[0104] in: Indicates the first Attention weight of each value anchor point Indicates the first Anchor feature vector of each value anchor point Represents the global awareness context vector. This represents the vector dot product operation. This represents the natural exponential function. This represents the total number of anchors in the set of value anchors. Knowledge-guided vector. It is obtained by weighted summation of the eigenvectors of all anchor points: Knowledge-guided vectors Used for feature modulation of the perceived feature vector sequence.

[0105] In one embodiment of the invention, a value decoupling network is constructed, comprising a shared encoder, a spatial decoupling branch, and a temporal decoupling branch. The shared encoder is a multi-layer fully connected neural network with an input dimension consistent with the dimension of the multi-source fused feature vector. The shared encoder's role is to map the multi-source fused feature vector to a high-dimensional shared feature space. The spatial decoupling branch is a network structure containing multiple two-dimensional convolutional layers, and the temporal decoupling branch is a network structure containing multiple long short-term memory (LSM) layers. The multi-source fused feature vector is input to the shared encoder, and the forward propagation process of the shared encoder performs a nonlinear transformation on the input vector, ultimately obtaining a high-dimensional, high-order fused feature.

[0106] The high-order fusion features are simultaneously input into both the spatial and temporal decoupling branches. The spatial decoupling branch contains multiple spatial convolutional layers. These layers use two-dimensional convolutional kernels to extract features from the input high-order fusion features, specifically spatial structural feature components related to building structure, materials, and spatial layout. These spatial structural feature components are then mapped to a spatial value contribution vector. In practice, the high-order fusion features are reshaped into a two-dimensional feature map. Multiple two-dimensional convolutional layers perform convolution and downsampling operations on this feature map to capture local spatial correlations. Finally, a fully connected layer maps the extracted spatial structural feature components to a fixed-dimensional vector, which is the spatial value contribution vector. The temporal decoupling branch contains multiple recurrent connected layers. These recurrent connected layers employ Long Short-Term Memory (LSTM) network units or gated recurrent units to extract temporal evolution feature components related to building age, historical changes, and maintenance records from the high-order fusion features. These temporal evolution feature components are then mapped to a temporal value contribution vector. It is understandable that high-order fused features are treated as a sequence, with multiple recurrent connected layers processing the input step-by-step to capture temporal dependencies. The final output is mapped to a time value contribution vector through a fully connected layer. In one example, for a multi-source fused feature vector that integrates structural vibration monitoring sequences and three-dimensional geometric features, the spatial decoupling branch of the value decoupling network may extract feature components reflecting the stiffness of beam-column joint connections and the properties of wall materials, while the temporal decoupling branch may extract feature components reflecting the evolution trend of the building's structural response during multiple earthquake events.

[0107] During the training of the value decoupling network, optimizing the objective function encourages the network to decouple its independent spatial and temporal components. The training objective comprises two parts: maximizing the lower bound of mutual information between the spatial and temporal value contribution vectors, and minimizing the distance between the spatial structural feature component and the temporal evolution feature component within a common latent space. The lower bound of mutual information measures the correlation between the spatial and temporal value contribution vectors; maximizing this lower bound encourages the two vectors to share common information related to the overall value of the building. The distance between the spatial structural feature component and the temporal evolution feature component within a common latent space is measured by calculating the Euclidean distance between the vectors projected onto the common latent space. Minimizing this distance forces the spatial and temporal components extracted by the network to be as close as possible within the common latent space, ensuring they originate from the same building entity and serve a unified value assessment task, while maintaining their respective representational focuses through the decoupling mechanism. For example, for a century-old brick church, the spatial value contribution vector generated after training the value decoupling network may be strongly correlated with the structural mechanical characteristics of its Gothic arches, while the temporal value contribution vector may be strongly correlated with the traces of its history of war and multiple repairs. High mutual information between the two indicates that they jointly define the unique value of the church, and close proximity in the public potential space ensures that the decoupled components are consistent. In some embodiments, this is achieved by minimizing the following loss function... To train a value decoupling network:

[0108]

[0109] in: This represents the total loss of the value decoupling network. Represents the spatial value contribution vector With time value contribution vector The lower bound estimate of mutual information between them It is a hyperparameter that controls the weights of mutual information terms. and These are the spatial structural feature components. and time evolution characteristic components The projection function mapped to the common latent space. The L2 norm of a vector is the square of its value. It is a hyperparameter that controls the weight of the distance term.

[0110] In one embodiment of the invention, a generator network and a discriminator network are initialized. The generator network is a multi-layer fully connected neural network, and the discriminator network is a neural network containing multiple fully connected layers and an output head. The spatial value contribution vector and the temporal value contribution vector are concatenated along the feature dimension. The concatenated vector is fed into the generator network as input. The forward propagation process of the generator network performs a nonlinear transformation and dimensionality reduction on the input vector, ultimately generating a preliminary fusion vector. The preliminary fusion vector, the original spatial value contribution vector, and the original temporal value contribution vector are respectively input into the discriminator network, which contains three processing sub-branches. The discriminator network is trained to distinguish whether the preliminary fusion vector comes from "generated samples" from the generator network or from "real samples" from a real distribution. Real samples are obtained by sampling a large number of known value representation vectors of historical buildings. These known value representation vectors are derived from a sample database of historical buildings that have been evaluated by experts and have completed feature encoding. The discriminator network is trained to determine whether the initial fused vector contains complete spatial structure information and temporal evolution information. The loss function of the discriminator network includes adversarial loss and reconstruction loss. The adversarial loss is used to determine the authenticity of the sample, and the reconstruction loss is used to measure the difference between the spatial and temporal components decoded from the initial fused vector and the original input.

[0111] The generator network is trained to generate fusion vectors that can deceive the discriminator network. The generator network aims to make it difficult for the discriminator network to distinguish between generated and real samples, and to minimize the errors of the spatial and temporal components reconstructed by the discriminator network from the original input vector. Through multiple iterative adversarial training sessions between the generator and discriminator networks, in each iteration, the generator network updates its parameters based on feedback from the discriminator network to generate more realistic fusion vectors, while the discriminator network updates its parameters based on real and generated samples to improve its discrimination and reconstruction capabilities. The training process stops when the discriminator network can no longer effectively distinguish between generated and real samples, and the reconstruction error is below a preset threshold. This preset threshold is set based on the error convergence on the validation set. The outputs of the finally trained generator network in the spatial and temporal value contribution vectors are used as stable value representation vectors. These stable value representation vectors integrate value information from both spatial and temporal dimensions and possess statistical distribution characteristics similar to those of the real value representation vectors.

[0112] The adjacency subgraph of each anchor node in the value anchor set is extracted from the knowledge graph. The adjacency subgraph, centered on the anchor node, contains all adjacent nodes within one or two hops and their connecting edges. Each adjacency subgraph is converted into a corresponding graph embedding vector, which is obtained by processing the node features and adjacency relationships of the adjacency subgraph using a graph neural network. The cosine similarity between the stable value representation vector and each graph embedding vector is calculated. Cosine similarity calculates the proximity of two vectors in a direction, with values ​​ranging from -1 to 1, resulting in a preliminary one-dimensional list of associations, where each element corresponds to a value anchor node. All anchor nodes in the value anchor set are paired, and the shortest path length in the knowledge graph for each pair of anchor nodes is calculated. The shortest path length is calculated using a graph traversal algorithm such as Dijkstra's algorithm, forming an anchor-to-anchor path distance matrix. The anchor-to-anchor path distance matrix is ​​a symmetric matrix, where the matrix element d_{ij} represents the number of shortest path edges between anchor node i and anchor node j.

[0113] An anchor-point path distance matrix is ​​transformed into an anchor-point relationship strength matrix using a predefined kernel function. The kernel function maps path distance to relationship strength; shorter distances indicate higher relationship strength. The anchor-point relationship strength matrix represents the prior association strength between anchors. The initial association list is expanded into a diagonal matrix, where the diagonal elements are the values ​​from the initial association list, and the off-diagonal elements are zero. A matrix multiplication operation is performed between the expanded diagonal matrix and the anchor-point relationship strength matrix. This multiplication spreads the association between the stable value representation vector and a single anchor to its associated neighbors in the knowledge graph. An anchor highly correlated with the stable value representation vector will increase the association score of its neighboring anchors. The resulting matrix is ​​then normalized using row normalization to ensure that the sum of each row's elements is 1, yielding the final quantified association matrix. Each element in the matrix represents the quantified association metric of the stable value representation vector and each pair of anchors, corrected for knowledge graph relationships. In an example scenario, the stable value representation vector has a high initial correlation with the "Gothic Flying Buttress" anchor point. The "Gothic Flying Buttress" anchor point is 1 distance from the "Structural Innovation" value dimension node and 2 distance from the "Medieval Architectural Technology" historical event node in the knowledge graph. Through matrix operations, the correlation between the "Structural Innovation" and "Medieval Architectural Technology" nodes is enhanced due to their strong correlation with the "Gothic Flying Buttress." The quantified correlation matrix reflects this graph-based diffusion effect (see Table 1).

[0114] Table 1: Example of conversion between path distance matrix and relationship strength matrix between anchor points

[0115]

[0116] Table 1 presents a simplified path distance matrix between anchor points, where the values ​​represent the number of hops in the shortest path of the knowledge graph. This is achieved using a Gaussian kernel function. Convert it into a relation strength matrix, where This represents the path distance between anchor point i and anchor point j. These are the kernel function parameters that control the decay rate. This indicates the strength of the relationship after transformation. For example, when The relationship strength is relatively high when The strength of the relationship with time decreased significantly.

[0117] See Figure 4 This is a heatmap showing the strength of relationships between anchor points, used to illustrate the degree of association between value anchor points in the knowledge graph during the valuation of historical buildings. The graph displays the strength of relationships between five value anchor points. The color gradient from blue to red represents the degree of association between anchor points in the knowledge graph. The cells on the diagonal are dark red, with a relationship strength of 1.0, because each anchor point has a path distance of 0 from itself, indicating the highest relationship strength. The relationship strength between "brick and wood structure" and "spatial layout" is relatively high (approximately 0.8), indicating a close association between them in the knowledge graph; the relationship strength between "carving techniques" and "Republican style" is moderate (approximately 0.7), reflecting the association between craftsmanship and style in Republican-era architecture; the relationship strength between "historical changes" and "brick and wood structure" is relatively low (approximately 0.1), indicating a relatively large path distance between them in the knowledge graph.

[0118] In one embodiment of the invention, a confidence threshold is set, which is a scalar value between 0 and 1, to filter significant associations. All matrix elements greater than the confidence threshold are selected from the quantified association matrix. Each selected matrix element corresponds to a pair of anchor nodes in the knowledge graph. In the quantified association matrix, the row index and column index each represent a value anchor node, and the value of the matrix element represents the association degree between the stable value representation vector and the joint representation of this pair of anchor nodes.

[0119] Based on a knowledge graph, all anchor nodes involved in candidate associations are considered mandatory nodes, and the optimal subgraph path connecting these mandatory nodes is searched. The search process is performed on the topological structure of the knowledge graph, and the goal of the optimal subgraph path is to maximize the comprehensive score of the candidate association degrees corresponding to all edges traversed along the path. The comprehensive score can be calculated by multiplying or weighted summing all association degree values ​​along the path. The optimal subgraph path is the value assessment path, a connected subgraph in the knowledge graph that connects multiple value anchors highly correlated with stable value representation vectors and reveals the semantic relationships between these anchors. In the knowledge graph, the value assessment path is traced backward from the terminal node to the starting node, with the starting and terminal nodes being the two endpoints of the path. The attributes and relationship information of all nodes and edges on the path are extracted. Node attributes include node name and descriptive text, and edge attributes include relationship type, generating a path description vector. The path description vector is obtained by encoding and aggregating the attribute text of all nodes and edges.

[0120] The path description vector and the stable value representation vector are input into an attention fusion module, which is a fully connected neural network based on an attention mechanism. The attention fusion module calculates the attention weight of each part of the path description vector to the stable value representation vector. The path description vector is divided into multiple parts, each corresponding to a node or edge on the path. Based on the attention weights, the path description vector and the stable value representation vector are weighted and fused, supplementing and correcting the stable value representation vector with information. Weighted fusion is achieved by adding or concatenating the weighted sum of the path description vector with the stable value representation vector element-wise through a fully connected layer. The result of this information supplementation and correction is the final value embedding vector, which incorporates deep semantic information from specific value evaluation paths in the knowledge graph.

[0121] A pre-constructed value assessment benchmark space is established, spanned by multiple orthogonal value benchmark axes. Each value benchmark axis represents a core value dimension defined and standardized by experts, such as "historical value," "artistic value," "scientific value," and "social value." Orthogonality is guaranteed through principal component analysis or Schmidt orthogonalization, ensuring that each value dimension is mathematically independent. The final value embedding vector is projected onto each value benchmark axis by calculating the dot product of the value embedding vector and the unit vector of each value benchmark axis. The projection length is calculated to obtain the projected coordinate components of the value embedding vector on each value benchmark axis. Each value benchmark axis is associated with a pre-defined weight coefficient, reflecting the importance of the core value dimension in the overall assessment. The weight coefficient is determined using expert scoring or the analytic hierarchy process. The projected coordinate components of the value embedding vector on each value benchmark axis are multiplied by the corresponding weight coefficient to obtain the weighted value score of the core value dimension. The weighted value scores corresponding to all value benchmark axes are summed to obtain the preliminary comprehensive assessment score for the historical building. The preliminary comprehensive assessment score is a comprehensive scalar value.

[0122] The system retrieves several reference buildings from the knowledge graph that are most similar to the current historical building in terms of type, region, and era. The search criteria are based on matching the attribute tags of the building entities in the knowledge graph. Existing authoritative assessment values ​​for the reference buildings are obtained, derived from evaluation records by cultural heritage departments or published academic assessment reports. The relative deviation between the current historical building's preliminary comprehensive assessment score and the average of the authoritative assessment values ​​of the reference buildings is calculated. The formula for calculating the relative deviation is:

[0123]

[0124] in: Indicates relative deviation. This indicates the initial comprehensive assessment score for the current historical building. Indicates the first The authoritative assessment value of a reference building, This indicates the total number of reference buildings. The preliminary comprehensive assessment score is calibrated based on relative deviation; the calibration process involves multiplying the preliminary comprehensive assessment score by an adjustment factor based on relative deviation. ,in It is a smoothing coefficient less than 1. The calibrated result is the final intelligent assessment value of the historical building, which is a comprehensive assessment score calibrated with reference to similar cases. In some embodiments, for a modern industrial building with a high initial assessment score, the retrieved reference buildings may be other industrial heritage sites from the same period and region. If the authoritative assessment values ​​of these reference buildings are generally low, and the calculated relative deviation is positive, the initial comprehensive assessment score will be calibrated downwards so that the final intelligent assessment value is more consistent with the overall value distribution of buildings of this category.

[0125] See Figure 5 This is a radar chart of the historical building's value assessment, showing the composition of the final assessment value across six core value dimensions. As the chart shows, the assessment values ​​for each dimension generally fall between 75 and 85 points. Historical and scientific value scores are relatively high, approaching 85 points, representing the building's outstanding strengths. Cultural value also scores at a high level. Practical and social value scores are relatively low, representing weaknesses in the building's value composition. The building's value distribution is relatively balanced, without extreme weaknesses, but it doesn't reach the top level in all dimensions either, exhibiting an overall comprehensive value characteristic centered on historical and scientific value. The outstanding historical and scientific value indicates that the building possesses high research and conservation value in terms of its age, connection to historical events, construction technology, and materials. The relatively weak practical and social value suggests that in subsequent protection and utilization, its social influence and practical value can be enhanced through reasonable functional renovations, community participation, and the planning of cultural activities.

[0126] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments that can be applied to other fields. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A method for intelligent assessment of the value of historical buildings based on multi-source data fusion, characterized in that: include: Preprocessing is performed on the original multimodal information of historical buildings to separate structured value description information and unstructured perception information, and generate multi-source fusion feature vectors; Spatial structure feature components and temporal evolution feature components are separated from the multi-source fusion feature vector. Using a pre-defined value decoupling network, the spatial structure feature components and temporal evolution feature components are decoupled and analyzed to obtain spatial value contribution vector and temporal value contribution vector. The spatial value contribution vector and the temporal value contribution vector are fused adversarially, and the fusion result is converged through iterative optimization to obtain a stable value representation vector. The stable value representation vector and the set of value anchors are measured by graph structure similarity, and a quantitative correlation matrix is ​​output. The value assessment path that meets the confidence requirement is extracted from the quantitative correlation matrix. The knowledge graph is then traced back to its source based on the value assessment path to correct the stable value representation vector and generate the final value embedding vector. The final intelligent assessment value of a historical building is calculated based on the projection coordinates of the value embedding vector in the preset value assessment benchmark space.

2. The intelligent assessment method for the value of historical buildings based on multi-source data fusion according to claim 1, characterized in that, The preprocessing of the raw multimodal information of historical buildings to separate structured value description information from unstructured perceived information specifically includes: Receive raw multimodal information from different acquisition devices, wherein the raw multimodal information includes at least architectural drawing text, sensor monitoring time-series signals, historical document records, and image scan point cloud data; A parsing pipeline is constructed for different modal information, which includes a text parser, a signal feature extractor, an optical character recognition engine, and a point cloud registration module; Text and image information are processed by a text parser and an optical character recognition engine to extract field values ​​with clear field definitions and logical relationships, and the field values ​​and their relationships are integrated into structured value description information. The signal feature extractor and point cloud registration module process the time-series signal and spatial point cloud data, extract the original signal waveform feature sequence and three-dimensional geometric feature point set, and integrate the signal waveform feature sequence and three-dimensional geometric feature point set into unstructured sensing information.

3. The intelligent assessment method for the value of historical buildings based on multi-source data fusion according to claim 2, characterized in that, The generation of the multi-source fusion feature vector includes: Information fingerprints are generated from structured value description information and unstructured perceived information, respectively. Based on the information fingerprint of structured value description information, retrieve the set of associated value anchors in the knowledge graph; Semantic space mapping is performed on unstructured perceptual information to generate a sequence of perceptual feature vectors with temporal correlation; Based on the set of value anchors, temporal feature enhancement and dimension alignment operations are performed on the perceptual feature vector sequence to generate multi-source fusion feature vectors; The process of generating information fingerprints from structured value description information and unstructured perceived information specifically includes: A preset hash algorithm is used to process the integrated structured value description information to generate a structured information fingerprint that represents its global information content; A pre-defined deep feature encoder is used to encode the integrated unstructured perceptual information, and the dimensionality reduction representation of the encoded feature vector is used as its perceptual information fingerprint.

4. The intelligent assessment method for the value of historical buildings based on multi-source data fusion according to claim 3, characterized in that, The step of retrieving a set of associated value anchors in a knowledge graph based on the information fingerprint of structured value description information includes: Construct a domain knowledge graph that includes historical building entities, attributes, value dimensions, and historical event nodes. The edges in the knowledge graph represent the relationships or attribute affiliations between entities. The similarity between the structured information fingerprint and the identifier fingerprint of all entities in the knowledge graph is calculated, and multiple initial associated entities with similarity exceeding the preset entity matching threshold are selected. Starting from the multiple initial associated entities, a multi-hop traversal is performed along the relationship edges in the knowledge graph to explore all value dimension nodes and historical event nodes that have direct or indirect connection relationships with the multiple initial associated entities within a preset number of hops. Collect all paths traversed during the traversal process and mark the endpoint of each path as a candidate value anchor. Calculate the semantic relevance score between each candidate value anchor and the structured information fingerprint, and filter out candidate value anchors whose semantic relevance scores are lower than the preset anchor screening threshold. The remaining candidate value anchors after filtering, along with their local topological structures in the knowledge graph, are used together as the retrieved value anchor set.

5. The intelligent assessment method for the value of historical buildings based on multi-source data fusion according to claim 4, characterized in that, The step of semantically mapping unstructured perceptual information to generate a sequence of perceptual feature vectors with temporal correlation includes: The signal waveform feature sequence in unstructured sensing information is divided into multiple signal segments by an equally spaced sliding window. Spatial voxelization meshing is performed on the set of three-dimensional geometric feature points in unstructured perceptual information, and the statistical distribution characteristics of the internal points of each voxel mesh are calculated. Each signal segment is input into the temporal feature coding network, and the output is the temporal coding vector of each signal segment; The statistical distribution characteristics of each voxel grid are input into the spatial feature encoding network, and the spatial encoding vector of each voxel grid is output. According to the timestamp order of signal acquisition, the time-series coding vector is concatenated with the mean vector of all spatial coding vectors within the corresponding time window to form the preliminary fusion vector at each time point; Arrange the preliminary fusion vectors from all time points in chronological order to form a preliminary perception feature vector sequence; The initial perceptual feature vector sequence is input into a recurrent neural network to capture the dependencies between adjacent vectors in the sequence. The hidden state sequence output by the recurrent neural network is the perceptual feature vector sequence with temporal correlation. The step of performing temporal feature enhancement and dimension alignment operations on the perceptual feature vector sequence based on the value anchor point set to generate a multi-source fusion feature vector specifically includes: Extract the attribute description text of each value anchor node from the set of value anchors, and convert the attribute description text into an anchor feature vector through a text encoder; Calculate the self-attention weight distribution of each temporal vector in the perceptual feature vector sequence, and perform a weighted summation of all vectors in the sequence based on the weight distribution to generate a global perceptual context vector; The dot product similarity of each anchor feature vector with the global perception context vector is calculated to obtain the attention weight of each anchor to the global perception. Anchor feature vectors are weighted and fused using attention weights to generate knowledge guidance vectors; The knowledge-guided vector is multiplied element-wise with each temporal vector in the perceptual feature vector sequence to enhance the temporal features of the perceptual feature vector sequence, thus obtaining the knowledge-enhanced perceptual sequence. All vectors in the knowledge-enhanced perception sequence are averaged over the time dimension to generate a multi-source fusion feature vector with a unified dimension.

6. The intelligent assessment method for the value of historical buildings based on multi-source data fusion according to claim 5, characterized in that, The steps of separating spatial structure feature components and temporal evolution feature components from multi-source fusion feature vectors, and using a pre-defined value decoupling network to perform decoupling analysis on the spatial structure feature components and temporal evolution feature components respectively, to obtain spatial value contribution vectors and temporal value contribution vectors, include: Construct a value decoupling network, which includes a shared encoder, a spatial decoupling branch, and a temporal decoupling branch; The multi-source fusion feature vector is input into the shared encoder to obtain high-order fusion features; High-order fusion features are simultaneously input into the spatial decoupling branch and the temporal decoupling branch; The spatial decoupling branch contains multiple spatial convolutional layers, which are used to extract spatial structural feature components related to building structure, material and spatial layout from high-order fusion features, and map the spatial structural feature components into spatial value contribution vectors. The time decoupling branch contains multiple recurrent connection layers, which are used to extract time evolution feature components related to building age, historical changes, and maintenance records from high-order fusion features, and map the time evolution feature components into time value contribution vectors. During the training of the value decoupling network, the lower bound of mutual information between the spatial value contribution vector and the temporal value contribution vector is maximized, and the distance between the spatial structure feature component and the temporal evolution feature component in a common latent space is minimized, so as to enable the network to decouple its independent components.

7. The intelligent assessment method for the value of historical buildings based on multi-source data fusion according to claim 6, characterized in that, The step of adversarially fusing the spatial value contribution vector and the temporal value contribution vector, and then iteratively optimizing the fusion result to obtain a stable value representation vector, includes: Initialize a generator network and a discriminator network. Concatenate the spatial value contribution vector and the temporal value contribution vector and input them into the generator network to generate a preliminary fusion vector. The initial fusion vector, the original spatial value contribution vector, and the original temporal value contribution vector are respectively input into the discriminator network. The discriminator network is trained to distinguish whether the initial fusion vector is a "generated sample" from the generator or a "real sample" from the real distribution. Among them, the real samples were obtained by sampling the known value representation vectors of a large number of historical buildings; The discriminator network is trained to determine whether the initial fused vector contains complete spatial structure information and temporal evolution information. Its loss function includes adversarial loss and reconstruction loss. The generator network is trained to generate a fusion vector that can deceive the discriminator network. The goal is to make it difficult for the discriminator network to distinguish between generated samples and real samples, and to minimize the errors between the spatial and temporal components of the generated fusion vector reconstructed in the discriminator network and the original input vector. Through multiple iterations of adversarial training between the generator network and the discriminator network, the training process stops when the discriminator network cannot effectively distinguish between generated samples and real samples, and the reconstruction error is lower than a preset threshold. The outputs of the finally trained generator network on the spatial value contribution vector and the temporal value contribution vector are used as stable value representation vectors.

8. The intelligent assessment method for the value of historical buildings based on multi-source data fusion according to claim 7, characterized in that, The step of measuring the graph structure similarity between the stable value representation vector and the set of value anchors, and outputting a quantified correlation matrix, includes: Extract the adjacency subgraph of each anchor node in the set of value anchors from the knowledge graph, and convert each adjacency subgraph into a corresponding graph embedding vector; Calculate the cosine similarity between the stable value representation vector and each graph embedding vector to obtain a preliminary one-dimensional list of associations; Pair all anchor nodes in the value anchor set together, calculate the shortest path length of each pair of anchor nodes in the knowledge graph, and form a path distance matrix between anchor nodes. The path distance matrix between anchor points is converted into a relationship strength matrix between anchor points using a preset kernel function. The relationship strength matrix between anchor points characterizes the degree of prior association between anchor points. The initial correlation list is expanded into a diagonal matrix, and matrix multiplication is performed between it and the relationship strength matrix between anchor points to spread the correlation between the stable value representation vector and a single anchor point to its associated neighbors in the knowledge graph. The matrix obtained from the operation is normalized to obtain the final quantified correlation matrix. Each element in the matrix represents the correlation quantification value of each pair of anchor points jointly represented by the stable value representation vector and the knowledge graph relationship correction.

9. The intelligent assessment method for the value of historical buildings based on multi-source data fusion according to claim 8, characterized in that, The steps of extracting value assessment paths that meet confidence requirements from the quantified correlation matrix, performing reverse tracing of the knowledge graph based on the value assessment paths, correcting stable value representation vectors, and generating the final value embedding vector include: Set a confidence threshold, and filter out all matrix elements in the quantized correlation matrix that are greater than the confidence threshold; Each selected matrix element corresponds to a pair of anchor nodes in the knowledge graph. A pair of anchor nodes in the knowledge graph and their correlation values ​​in the quantized correlation matrix are recorded as a candidate correlation. Based on the knowledge graph, all anchor nodes involved in the candidate associations are taken as necessary nodes. The optimal subgraph path connecting the necessary nodes is searched. The optimal subgraph path makes the comprehensive score of all associations on the path the highest. This optimal subgraph path is the value assessment path. In the knowledge graph, the value assessment path is traced backward from the terminal node to the starting node to extract the attribute and relationship information of all nodes and edges on the path and generate a path description vector. The path description vector and the stable value representation vector are input into an attention fusion module, which calculates the attention weight of each part of the path description vector to the stable value representation vector. Based on attention weights, the path description vector and the stable value representation vector are weighted and fused, and the stable value representation vector is supplemented and corrected to obtain the final value embedding vector.

10. The intelligent assessment method for the value of historical buildings based on multi-source data fusion according to claim 9, characterized in that, The step of calculating the final intelligent assessment value of a historical building based on the projection coordinates of the value embedding vector in a preset value assessment benchmark space includes: A value assessment benchmark space is pre-constructed, which is spanned by multiple orthogonal value benchmark axes. Each value benchmark axis represents a core value dimension that has been defined and standardized by experts. The final value embedding vector is projected onto each value reference axis, the projection length is calculated, and the projection coordinate components of the value embedding vector on each value reference axis are obtained. Each value benchmark axis is associated with a preset weight coefficient, which reflects the importance of the core value dimension in the overall evaluation; The weighted value score of the core value dimension is obtained by multiplying the projected coordinate components of the value embedding vector on each value reference axis with the weight coefficient corresponding to the value reference axis. The weighted value scores corresponding to all value benchmark axes are summed to obtain the preliminary comprehensive assessment score of the historical building; Retrieve from the knowledge graph several reference buildings that are most similar to the current historical building in terms of type, region, and age, and obtain the existing authoritative assessment value of the reference buildings; The relative deviation between the preliminary comprehensive assessment score of the current historical building and the average value of the authoritative assessment of the reference building is calculated. The preliminary comprehensive assessment score is calibrated based on the relative deviation. The calibrated result is the final intelligent assessment value of the historical building.

Images