Urban construction archive abnormity early warning method based on artificial intelligence

Abstract

The invention discloses an urban construction archive abnormity early warning method based on artificial intelligence, and the method comprises the steps: S1, collecting and preprocessing multi-source archive data, and constructing an archive feature library; s2, extracting file text features, drawing component features and approval label features, and generating a multi-modal feature vector; s3, generating an archive knowledge graph and a time sequence feature sequence by using the multi-modal feature vector, inputting the archive knowledge graph and the time sequence feature sequence into an HGT graph neural network, and outputting archive structure embedded representation and evolution chain semantic features; s4, the archive structure embedding is compared with the version evolution chain feature, and a potential abnormal record is generated; s5, carrying out risk quantification based on the improved MC-Dropout, and outputting an abnormal type, a position and a risk level; and S6, submitting the abnormal information to an urban construction archive management platform. According to the method, the early warning strategy is optimized through the graph neural network and the uncertainty evaluation mechanism, and efficient and accurate urban construction archive anomaly recognition and early warning are achieved.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and risk warning technology, and in particular to an artificial intelligence-based method for early warning of anomalies in urban construction archives. Background Technology

[0002] In recent years, with the continuous accumulation of urban construction archives, the complexity of their management and the burden of review have increased significantly, leading to the gradual introduction of intelligent methods into the archives governance process. Among these, artificial intelligence (AI) technology, due to its strong feature extraction capabilities and high pattern recognition efficiency, has been widely used in applications such as archive anomaly identification, early warning, and risk assessment. However, in practical deployments, existing AI methods still have significant limitations in areas such as multi-source data fusion, accuracy of anomaly modeling, and the ability to interpret early warnings.

[0003] Currently, most methods for identifying anomalies in urban construction archives rely solely on text comparison or structural rules, making it difficult to fully explore the correlation characteristics of unstructured data such as drawings, approval records, and spatial information. This results in incomplete anomaly identification under complex evolutionary processes. Some studies only use single static judgment results as early warning outputs, ignoring the predictive uncertainty caused by the evolution of archive status, changes in relationships, and frequent data updates. They also lack effective mechanisms to quantify the credibility of anomaly risks. Furthermore, most early warning systems are black-box outputs, lacking interpretable analytical paths and indicator support, making it difficult to assist archive management personnel in making accurate judgments and tracing the source of problems.

[0004] Furthermore, current AI models often fail to incorporate the heterogeneous structural characteristics and temporal evolution patterns of archival knowledge graphs, resulting in insufficient performance in capturing cross-modal dependencies between entities and potential anomaly propagation paths, thus affecting the comprehensiveness and foresight of anomaly warnings.

[0005] Therefore, how to provide an AI-based method for early warning of anomalies in urban construction archives is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose an AI-based method for early warning of anomalies in urban construction archives. This invention fully integrates features such as multi-source archival structural data, semantic information, approval relationships, and evolutionary timelines to construct an anomaly identification and risk warning process with capabilities including heterogeneous structure modeling, semantic relationship perception, time-dependent expression, and uncertainty assessment. By introducing an HGT graph neural network to embed and learn multiple types of entities and relationships in the archival knowledge graph, it accurately captures cross-modal dependency paths between complex archives. Combined with an improved MC-Dropout approach, it introduces temporal consistency and relationship consistency constraints in multiple random forward propagation paths, perturbing and weighting the predicted output of potential anomaly records to achieve joint generation and dynamic control of anomaly scores and uncertainty scores. This invention possesses advantages such as structured expression of relationship modeling, strong robustness in anomaly detection, a reliable and adjustable scoring mechanism, and traceable early warning output. It can significantly improve the accuracy and interpretability of intelligent early warning in urban construction archive scenarios, thereby effectively solving problems such as limitations in anomaly capture, lack of prediction reliability, and opaque early warning logic in existing methods.

[0007] An artificial intelligence-based early warning method for urban construction archive anomalies according to an embodiment of the present invention includes the following steps:

[0008] S1. Collect multi-source archival data, perform preprocessing, and construct an archival feature database;

[0009] S2. Extract the text features, drawing component features, and approval label features from the archive feature database to generate a multimodal feature vector;

[0010] S3. Generate an archive knowledge graph and temporal feature sequence using multimodal feature vectors, and introduce them into the HGT graph neural network along with multimodal node features to output the archive structure embedding representation and evolution chain semantic features;

[0011] S4. Compare the embedded representation of the archive structure with the features of the version evolution chain to generate potential abnormal records;

[0012] S5. Based on the improved MC-Dropout, the structural offset difference, semantic conflict cosine similarity interpolation and temporal order offset difference of potential abnormal records are used to quantify the risk and output the anomaly type, anomaly location and risk level.

[0013] S6. Submit the anomaly type, anomaly location, and risk level to the urban construction archives management platform, record the early warning output and review feedback information, and update the review log and permission records.

[0014] Optionally, S1 specifically includes:

[0015] S11. Read the original files of urban construction archives, including text, images, drawings, approval process records, and location information. Convert the file content according to the character encoding format. Read the image and drawing data. Convert the image pixel format and drawing vector format according to the unified pixel standard and unified vector standard. Convert all images and drawings to a unified size and unified resolution.

[0016] S12. Perform field extraction on the approval process record and location information. Read the approval node name, approval time, approval department and coordinate location in sequence according to the preset field names. Save the extracted fields with unified field names. Perform sentence segmentation and paragraph separation on the urban construction archive text. Separate continuous text into independent text segments according to periods and punctuation marks. Store each text segment in sequence.

[0017] S13. After the urban construction archives text, images, drawings, approval process records and location information have been standardized by uniform encoding, uniform pixel format, uniform vector format and uniform field, they are classified and stored according to data type, and the constructed archive feature library is output.

[0018] Optionally, S2 specifically includes:

[0019] S21. Read the urban construction archive text in the archive feature library line by line, perform word segmentation on each line of text and record the part of speech, combine the word order into text semantic fragments, output the initial features of the urban construction archive text, match the words in the initial features of the urban construction archive text with the preset word vector table, concatenate the matched word vectors in order into text semantic vectors, replace the unmatched words with preset placeholder vectors, and obtain the archive text features.

[0020] S22. Input the urban construction archive drawings into the image semantic segmentation process, divide the drawings into drawing component areas, perform pixel classification on each area and mark it as component lines, component symbols and component annotations, generate drawing component features, read the approval process records in the order of fields, perform label extraction on the approval node name, approval time and approval department, and generate approval label features.

[0021] S23. Match the document text features, drawing component features, and approval label features according to their respective document numbers, concatenate the three features in sequence into a unified feature vector, and output the concatenated unified feature vector as a multimodal feature vector.

[0022] Optionally, S3 specifically includes:

[0023] S31. Bind the multimodal feature vector to the corresponding urban construction archive text entity, drawing component entity and approval event entity according to the archive number as multimodal node feature. Write the relationship type according to the relationship between the target entities. Record each relationship as an edge and record the time sequence of each edge as a time label to form an archive knowledge graph and time sequence feature sequence.

[0024] S32. In the HGT graph neural network, a dedicated linear transformation matrix for each node type is preset for urban construction archive text entity, drawing component entity and approval event entity. The feature of each multimodal node is multiplied element by element with the corresponding node type linear transformation matrix and the bias is added to obtain the node type-specific representation.

[0025] S33. Pre-set dedicated linear transformation matrices for design relationships, construction dependencies, and approval workflows respectively. Multiply each node type-specific representation with the corresponding relationship type linear transformation matrix element by element and add the offsets to obtain a message vector specific to the relationship type.

[0026] S34. Convert the time label of each edge into a time vector, add the time vector element by element to the relation type-specific message vector to generate a meta-relation message vector with time information. In the HGT graph neural network, perform multi-head attention calculation for each type of meta-relation, multiply the meta-relation message vector with time information element by element with the corresponding preset attention weight and add them to obtain the meta-relation aggregation vector.

[0027] S35. Perform residual join, layer normalization and feedforward transformation on the meta-relation aggregation vector in sequence, and perform residual join and layer normalization on the transformation result again to obtain the node representation after one update.

[0028] S36. Repeatedly perform node updates according to the network layer number, perform relation type-specific linear transformation, multi-head attention calculation and node update operations on all nodes until all nodes generate the final representation;

[0029] S37. Add the updated node representation of the current node element by element to the cumulative vector, move to the next node and repeat the operation until all nodes are processed to obtain the cumulative vector. Use the cumulative vector as the embedded representation of the archive structure. Divide the cumulative vector into multiple continuous segments in chronological order. Store each segment in the evolution chain semantic feature table to form the evolution chain semantic features.

[0030] Optionally, S4 specifically includes:

[0031] S41. Read the archive structure embedding representation and evolution chain semantic features sequentially according to the archive number, use the archive structure embedding representation as node features, and use the evolution chain semantic features as time sequence features and write them into the comparison sequence.

[0032] S42. Perform element-wise subtraction between each node feature and the features of its adjacent nodes, record the positions where the absolute value of the difference is greater than the set difference threshold as structural offset points, and write all structural offset points into the structural offset sequence in node order.

[0033] S43. Perform element-wise cosine similarity calculation between the multimodal feature vector of each node and the multimodal feature vector of the previous time node. Record positions below the set similarity threshold as semantic conflict points, and write all semantic conflict points into the semantic conflict sequence in chronological order.

[0034] S44. Read the temporal order index in the semantic features of the evolutionary chain, calculate the time interval between adjacent records according to the index order, record the positions where the time interval is greater than the set interval threshold as time breakpoints, and write all time breakpoints into the temporal order offset sequence.

[0035] S45. Match the structural offset sequence, semantic conflict sequence, and time order offset sequence sequentially according to node number and time index. When the number of records of any node in any sequence exceeds the set offset threshold, mark the node as a potential anomaly and write the potential anomaly into the potential anomaly record.

[0036] Optionally, S5 specifically includes:

[0037] S51. Input the archive structure embedding representation, evolution chain semantic features and multimodal node features of the identified potential abnormal record nodes into the improved MC-Dropout, perform multi-round sampling inference and risk uncertainty assessment, and output the mean and variance of the abnormal probability distribution of each node.

[0038] S52. Using the mean and variance results of the anomaly probability obtained through the improved MC-Dropout, and combining the three types of indicators—structural offset difference, semantic conflict cosine similarity interpolation, and temporal order offset difference—to calculate the comprehensive risk score, and output the anomaly type, anomaly location, and risk level of the corresponding node.

[0039] Optionally, the improved MC-Dropout specifically includes:

[0040] In each hidden layer of the HGT graph neural network, a Dropout control node is set up to receive the output vector of the previous layer. During each forward propagation, a 0-1 random occlusion vector with the same dimension as the output vector is generated for the current layer. The random occlusion vector is multiplied element by element with the output vector, and the neuron outputs at the unoccluded positions are retained to form a predicted output with a random occlusion network computation path.

[0041] The above forward propagation is repeated a preset number of times for the same potentially abnormal record, generating different random occlusion vectors each time, obtaining the prediction output of a preset number of paths, and storing them sequentially according to the path number;

[0042] For each path's prediction output, a temporal consistency constraint is introduced. The previous and next time nodes in the semantic features of the evolution chain of potential anomaly records are read. The path prediction values ​​corresponding to the three time nodes are read sequentially. The absolute values ​​of the differences between the current node and the previous node and the next node are calculated. The position where the absolute value of any of the two differences is greater than the set jump threshold is marked as a temporal anomaly point. All prediction outputs are multiplied by a preset reduction coefficient.

[0043] For each path's predicted output, a relational consistency constraint is introduced. Potentially abnormal records are read from the set of nodes connected in the archive knowledge graph. For each adjacent node, the corresponding path prediction value is read. The absolute difference between the current node's prediction value and the prediction values ​​of each adjacent node is calculated and averaged. If the average difference is greater than the set relational deviation threshold, it is marked as a relational abnormal path, and all predicted outputs of the path are multiplied by a reduction factor.

[0044] After the prediction output of all paths has been reduced, the prediction output value of each path is read. The average of the prediction outputs of all paths is used as the final anomaly score. The average of the squared differences between the output values ​​of all paths and the anomaly scores is used as the uncertainty score. The anomaly scores and uncertainty scores are stored in the results list according to the sample number.

[0045] The obtained anomaly score and uncertainty score are added to the structural offset difference, semantic conflict cosine similarity interpolation and temporal order offset difference item by item, and then weighted and combined according to the preset weight coefficient to calculate the comprehensive risk score of potential anomaly records.

[0046] When the overall risk score exceeds the set risk threshold, the node is marked as a high-risk anomaly, and the anomaly type, anomaly location, and overall risk score are written into the early warning output table.

[0047] Optionally, S6 specifically includes: submitting the anomaly type, anomaly location, and risk level to the urban construction archive management platform, which registers and displays the anomaly according to the archive number and pushes it to the corresponding responsible department, simultaneously recording the early warning output, manual review opinions, and handling results, writing the review feedback into the archive update log, and updating the corresponding permission records.

[0048] The beneficial effects of this invention are:

[0049] First, the HGT graph neural network supports unified modeling and graph structure embedding learning of different types of entities and their various heterogeneous relationships in urban construction archives. It can accurately capture structural evolution, semantic diffusion and cross-type dependency paths. While maintaining the semantic integrity of entities, it can realize semantic reasoning and structural coupling recognition among complex urban construction archives, and significantly enhance the graph modeling capability for anomaly recognition.

[0050] Secondly, the proposed improved MC-Dropout dynamically sets random occlusion paths in each hidden layer and introduces temporal consistency constraints and relational consistency constraints in multiple forward propagation paths. This allows for dynamic filtering and weight reduction of the predicted values ​​output by each path, effectively identifying potential anomaly paths such as evolutionary time chain breaks and spectral structure shifts, thus improving the robustness and reliability of anomaly detection. Finally, the multi-path prediction results are fused to output anomaly scores and uncertainty scores, achieving an accurate characterization of the risk level of urban construction archives.

[0051] Furthermore, this invention designs a dynamic weighted combination method for risk scoring and factors such as structural deviation, semantic conflict, and temporal anomaly, which effectively enhances the ability of early warning results to explain the source of risk, improves the practicality, visualization, and management operability of early warning output, and provides technical support for building a highly reliable, highly perceptive, and traceable digital governance mechanism for urban construction archives. Attached Figure Description

[0052] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0053] Figure 1 This is a flowchart of an artificial intelligence-based early warning method for urban construction archives proposed in this invention;

[0054] Figure 2 This is a flowchart of the multimodal entity graph modeling and anomaly feature extraction based on HGT graph neural network proposed in this invention;

[0055] Figure 3 This is a flowchart of the improved MC-Dropout anomaly scoring and uncertainty estimation proposed in this invention. Detailed Implementation

[0056] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0057] refer to Figures 1-3 An AI-based method for early warning of anomalies in urban construction archives includes the following steps:

[0058] S1. Collect multi-source archival data and perform format unification, structured parsing and field standardization processing to build an archival feature database;

[0059] S2. Extract archival text features, drawing component features, and approval label features from the archival feature database using natural language processing, optical character recognition, and image semantic segmentation techniques to generate a multimodal feature vector;

[0060] S3. Input the multimodal feature vector as the attribute of the archive node into the heterogeneous graph construction process. Based on the relationship and time evolution order between the text entities of urban construction archives, the components of drawings and the entities of approval events, generate an archive knowledge graph and time-series feature sequence containing node features and relationship types. Together with the multimodal node features, introduce the HGT graph neural network to perform multi-head attention calculation and dynamic weight learning, and output the archive structure embedding representation and evolution chain semantic features.

[0061] S4. Based on structural consistency, semantic logic rationality, and temporal sequence continuity, the embedded representation of the archive structure is compared with the features of the version evolution chain to generate potential abnormal records.

[0062] S5. Based on the improved MC-Dropout and multi-indicator fusion risk scoring mechanism, the risk quantification of the structural offset difference, semantic conflict cosine similarity interpolation and temporal order offset difference of potential abnormal records is performed, and the abnormality type, abnormality location and risk level are output.

[0063] S6. Submit the anomaly type, anomaly location, and risk level to the urban construction archives management platform, record the early warning output and review feedback information, and update the review log and permission records.

[0064] This implementation method constructs a multimodal archival knowledge graph and temporal feature sequence, combined with multi-head attention computation and dynamic weight learning of an HGT graph neural network, to achieve accurate identification and risk prediction of potential anomalies in urban construction archives. By extracting multimodal features from archival text, drawing components, and approval labels, and inputting these features into a graph neural network for learning and reasoning, this invention can automatically identify complex relationships and evolutionary trends between different entities in the archives, greatly enhancing the ability to perceive and warn of archival anomalies. Introducing an improved MC-Dropout and multi-indicator fusion risk scoring mechanism enables comprehensive risk quantification across multiple dimensions such as structural deviation, semantic conflict, and temporal anomalies, effectively improving the accuracy, response speed, and level matching degree of anomaly warnings. It exhibits higher robustness and intelligence than traditional methods, especially in complex multi-source data scenarios. Finally, the system feeds back the anomaly type, location, and risk level to the urban construction archive management platform, achieving real-time warnings and dynamic optimization, effectively ensuring the accuracy and traceability of archival data management.

[0065] In this embodiment, S1 specifically includes:

[0066] S11. Read the original files of urban construction archives, including text, images, drawings, approval process records, and location information. Convert the file content according to the character encoding format. Read the image and drawing data. Convert the image pixel format and drawing vector format according to the unified pixel standard and unified vector standard. Convert all images and drawings to a unified size and unified resolution.

[0067] S12. Perform field extraction on the approval process record and location information. Read the approval node name, approval time, approval department and coordinate location in sequence according to the preset field names. Save the extracted fields with unified field names. Perform sentence segmentation and paragraph separation on the urban construction archive text. Separate continuous text into independent text segments according to periods and punctuation marks. Store each text segment in sequence.

[0068] S13. After the urban construction archives text, images, drawings, approval process records and location information have been standardized by uniform encoding, uniform pixel format, uniform vector format and uniform field, they are classified and stored according to data type, and the constructed archive feature library is output.

[0069] In this embodiment, S2 specifically includes:

[0070] S21. Read the urban construction archive text in the archive feature library line by line, perform word segmentation on each line of text and record the part of speech, combine the word order into text semantic fragments, output the initial features of the urban construction archive text, match the words in the initial features of the urban construction archive text with the preset word vector table, concatenate the matched word vectors in order into text semantic vectors, replace the unmatched words with preset placeholder vectors, and obtain the archive text features.

[0071] S22. Input the urban construction archive drawings into the image semantic segmentation process, divide the drawings into drawing component areas, perform pixel classification on each area and mark it as component lines, component symbols and component annotations, generate drawing component features, read the approval process records in the order of fields, perform label extraction on the approval node name, approval time and approval department, and generate approval label features.

[0072] S23. Match the document text features, drawing component features, and approval label features according to their respective document numbers, concatenate the three features in sequence into a unified feature vector, and output the concatenated unified feature vector as a multimodal feature vector.

[0073] In this embodiment, S3 specifically includes:

[0074] S31. Bind the multimodal feature vector to the corresponding urban construction archive text entity, drawing component entity and approval event entity according to the archive number as multimodal node feature. Write the relationship type according to the relationship between the target entities. Record each relationship as an edge and record the time sequence of each edge as a time label to form an archive knowledge graph and time sequence feature sequence.

[0075] S32. In the HGT graph neural network, a dedicated linear transformation matrix for each node type is preset for urban construction archive text entity, drawing component entity and approval event entity. The feature of each multimodal node is multiplied element by element with the corresponding node type linear transformation matrix and the bias is added to obtain the node type-specific representation.

[0076] S33. Pre-set dedicated linear transformation matrices for design relationships, construction dependencies, and approval workflows respectively. Multiply each node type-specific representation with the corresponding relationship type linear transformation matrix element by element and add the offsets to obtain a message vector specific to the relationship type.

[0077] S34. Convert the time label of each edge into a time vector, add the time vector element by element to the relation type-specific message vector to generate a meta-relation message vector with time information. In the HGT graph neural network, perform multi-head attention calculation for each type of meta-relation, multiply the meta-relation message vector with time information element by element with the corresponding preset attention weight and add them to obtain the meta-relation aggregation vector.

[0078] S35. Perform residual join, layer normalization and feedforward transformation on the meta-relation aggregation vector in sequence, and perform residual join and layer normalization on the transformation result again to obtain the node representation after one update.

[0079] S36. Repeatedly perform node updates according to the network layer number, perform relation type-specific linear transformation, multi-head attention calculation and node update operations on all nodes until all nodes generate the final representation;

[0080] S37. Add the updated node representation of the current node element by element to the cumulative vector, proceed to the next node and repeat the operation until all nodes are processed to obtain the cumulative vector. Use the cumulative vector as the embedded representation of the archive structure. Divide the cumulative vector into multiple continuous segments in chronological order. Store each segment in the evolutionary chain semantic feature table to form the evolutionary chain semantic features. Store the attention weights of each type of meta-relation generated in the multi-head attention calculation stage into a meta-relation attention value table according to the meta-relation type and save it according to the relation type index.

[0081] This implementation effectively enhances the anomaly identification and early warning capabilities of urban construction archives by introducing a Hidden Graph Transformer (HGT) neural network, combined with multimodal feature vectors and time-stamped information. First, by binding multimodal feature vectors to urban construction archive text entities, drawing component entities, and approval event entities, and constructing an archive knowledge graph based on the relationships between entities, a temporal feature sequence containing chronological order is formed. Then, the HGT neural network performs dynamic weight learning and message aggregation for different types of nodes and relationships through a pre-set dedicated linear transformation matrix and multi-head attention computation, thereby obtaining accurate node representations. After multi-layer updates and residual connection operations, a final archive structure embedding representation is generated and segmented into multiple continuous segments according to chronological order, forming evolutionary chain semantic features, further enhancing the archive's temporal reasoning and evolutionary trend analysis capabilities. This invention, by introducing temporal information and relationship-type-specific message vectors, makes the identification of abnormal records more accurate and comprehensive in complex archive management scenarios, significantly improving the model's robustness and adaptability, and effectively supporting the real-time optimization and efficient operation of the intelligent early warning system for urban construction archives.

[0082] In this embodiment, S4 specifically includes:

[0083] S41. Read the archive structure embedding representation and evolution chain semantic features sequentially according to the archive number, use the archive structure embedding representation as node features, and use the evolution chain semantic features as time sequence features and write them into the comparison sequence.

[0084] S42. Perform element-wise subtraction between each node feature and the features of its adjacent nodes, record the positions where the absolute value of the difference is greater than the set difference threshold as structural offset points, and write all structural offset points into the structural offset sequence in node order.

[0085] S43. Perform element-wise cosine similarity calculation between the multimodal feature vector of each node and the multimodal feature vector of the previous time node. Record positions below the set similarity threshold as semantic conflict points, and write all semantic conflict points into the semantic conflict sequence in chronological order.

[0086] S44. Read the temporal order index in the semantic features of the evolutionary chain, calculate the time interval between adjacent records according to the index order, record the positions where the time interval is greater than the set interval threshold as time breakpoints, and write all time breakpoints into the temporal order offset sequence.

[0087] S45. Match the structural offset sequence, semantic conflict sequence, and time order offset sequence sequentially according to node number and time index. When the number of records of any node in any sequence exceeds the set offset threshold, mark the node as a potential anomaly and write the potential anomaly into the potential anomaly record.

[0088] In this embodiment, S5 specifically includes:

[0089] S51. Input the archive structure embedding representation, evolution chain semantic features and multimodal node features of the identified potential abnormal record nodes into the improved MC-Dropout, perform multi-round sampling inference and risk uncertainty assessment, and output the mean and variance of the abnormal probability distribution of each node.

[0090] S52. Using the mean and variance results of the anomaly probability obtained through the improved MC-Dropout, and combining the three types of indicators—structural offset difference, semantic conflict cosine similarity interpolation, and temporal order offset difference—to calculate the comprehensive risk score, and output the anomaly type, anomaly location, and risk level of the corresponding node.

[0091] In this embodiment, the improved MC-Dropout specifically includes:

[0092] In each hidden layer of the HGT graph neural network, a Dropout control node is set up to receive the output vector of the previous layer. During each forward propagation, a 0-1 random occlusion vector with the same dimension as the output vector is generated for the current layer. The random occlusion vector is multiplied element by element with the output vector, and the neuron outputs at the unoccluded positions are retained to form a predicted output with a random occlusion network computation path.

[0093] The above forward propagation is repeated a preset number of times for the same potentially abnormal record, generating different random occlusion vectors each time, obtaining the prediction output of a preset number of paths, and storing them sequentially according to the path number;

[0094] A temporal consistency constraint is introduced for the prediction output of each path. The previous and next time nodes in the semantic features of the evolution chain of potential anomaly records are read. The path prediction values ​​corresponding to the three time nodes are read sequentially. The absolute values ​​of the differences between the current node and the previous node and the next node are calculated. The position where the absolute value of any difference between the two differences is greater than the set jump threshold is marked as a temporal anomaly point. All prediction outputs are multiplied by a preset reduction coefficient to reduce the contribution weight of the path.

[0095] For each path's predicted output, a relational consistency constraint is introduced. Potentially abnormal records are read from the set of nodes connected in the archive knowledge graph. For each adjacent node, the corresponding path prediction value is read. The absolute difference between the current node's prediction value and the prediction values ​​of each adjacent node is calculated and averaged. If the average difference is greater than the set relational deviation threshold, it is marked as a relational abnormal path. All predicted outputs of the path are multiplied by a reduction factor to further reduce the path contribution.

[0096] After the prediction output of all paths has been reduced, the prediction output value of each path is read. The average of the prediction outputs of all paths is used as the final anomaly score. The average of the squared differences between the output values ​​of all paths and the anomaly scores is used as the uncertainty score. The anomaly scores and uncertainty scores are stored in the results list according to the sample number.

[0097] The obtained anomaly score and uncertainty score are added to the structural offset difference, semantic conflict cosine similarity interpolation and temporal order offset difference item by item, and then weighted and combined according to the preset weight coefficient to calculate the comprehensive risk score of potential anomaly records.

[0098] When the overall risk score exceeds the set risk threshold, the node is marked as a high-risk anomaly, and the anomaly type, anomaly location, and overall risk score are written into the early warning output table.

[0099] This implementation method achieves efficient risk quantification and accurate identification of potential anomaly records by introducing an improved MC-Dropout algorithm and a multi-dimensional risk scoring mechanism. In this process, the structural embedding representation, evolutionary chain semantic features, and multimodal node features of the archive nodes are first input into the improved MC-Dropout algorithm. Multiple rounds of sampling inference and uncertainty assessment are performed, outputting the mean and variance of the anomaly probability distribution for each node. Then, combining three types of indicators—structural offset difference, semantic conflict cosine similarity interpolation, and temporal order offset difference—a comprehensive risk score is calculated, ultimately outputting the anomaly type, anomaly location, and risk level. The improved MC-Dropout further enhances the stability and accuracy of path prediction by setting Dropout control nodes in each layer of the HGT graph neural network and introducing temporal consistency and relational consistency constraints. By weightedly combining the prediction results of multiple paths, the final anomaly score and uncertainty score are generated, providing a comprehensive risk assessment and intelligent early warning for potential anomaly records. This invention can significantly improve the accuracy and timeliness of anomaly identification in urban construction archive management, providing strong technical support for intelligent early warning systems.

[0100] 8. The method for early warning of urban construction archive anomalies based on artificial intelligence according to claim 1, wherein step S6 specifically includes: submitting the anomaly type, anomaly location, and risk level to the urban construction archive management platform, which registers and displays the anomaly according to the archive number and pushes it to the corresponding responsible department, synchronously recording the early warning output, manual review opinions and handling results, writing the review feedback into the archive update log and updating the corresponding permission records.

[0101] Example 1:

[0102] To verify the feasibility and accuracy of this invention in urban construction archives management, it was deployed in the urban construction archives information management center of a provincial capital city. This center is responsible for the integration and risk monitoring of archives information for over 200 large-scale municipal infrastructure projects under construction and already completed throughout the city. This includes various types of archives such as planning approvals, as-built drawing reviews, structural change records, material acceptance forms, construction image records, and survey and design documents, totaling over 280TB of multimodal archive data resources. The daily update volume exceeds 1.5GB, involving cross-data submissions from hundreds of units, including construction companies, supervision companies, and planning bureaus.

[0103] Traditional document management systems primarily rely on field keywords and file structure verification to check the consistency and compliance of submitted materials. However, when faced with large amounts of mixed text and graphics data, unstructured documents, and chronological change information, they suffer from insufficient structured representation, missing graph-based associations, rigid anomaly identification rules, and delayed update responses. Especially in scenarios involving design changes, construction organization adjustments, and the supplementation of as-built documentation, potential risks such as inconsistencies, duplicate archiving, and missing key milestones often arise, but these risks are difficult to identify and warn of in advance.

[0104] In practical deployment, this invention decomposes existing and new data in the central archive system according to project dimensions, constructing a multimodal heterogeneous graph with nodes representing document metadata, text content summaries, structural diagram feature extraction, expert annotations, and image OCR recognition content. It also constructs temporal relationship edges, semantic content similarity edges, and responsibility role constraint edges between documents based on project stages and process status. During graph modeling, an HGT graph neural network is used to embed different types of nodes, and an improved MC-Dropout is introduced, adding Dropout control nodes to each hidden layer. Multiple forward propagation paths are sampled, and risk and uncertainty scores are estimated for the prediction results. After each change in archive data, the system extracts its graph substructure and inputs it into the model for inference, determining whether there are inconsistencies, omissions, expiration, or logical conflicts between the new data and existing archive data, and quantifying its potential anomaly risk level.

[0105] In three months of practical application, this invention assisted the system in automatically mapping and providing early warning assessments for 3,168 new archives, covering 32 large-scale municipal projects and involving over 420,000 document nodes. The table below shows the statistical comparison of key indicators between the method of this invention and traditional rule-based comparison strategies in archive anomaly identification during this period:

[0106] Table 1. Performance Comparison Data Between the Invention and Traditional Methods

[0107] Archive exception type Method type Total number of exception documents Hit times Accuracy (%) Average response time (seconds) Exception level matching degree (%) False positive rate (%) False negative rate (%) Content logic conflict Traditional method 452 301 66.6 42.3 61.2 15.3 10.8 Inventive method 447 421 94.2 13.5 90.7 3.8 1.7 Image-text inconsistency recognition Traditional method 208 146 70.2 37.1 59.8 12.4 8.7 Inventive method 202 192 95.0 12.7 92.3 2.9 1.5 Multi-version archive conflict Traditional method 184 124 67.4 38.9 63.1 13.8 9.5 Inventive method 177 168 94.9 11.3 91.6 3.2 2.1 Missed key attachment Traditional method 126 83 65.9 44.5 58.7 14.6 10.2 Inventive method 119 113 95.0 10.9 89.8 3.4 2.5

[0108] Based on the comparative data shown in Table 1, the urban construction archive anomaly early warning method proposed in this invention, based on HGT graph neural network and improved MC-Dropout, significantly outperforms traditional rule-based methods in several key performance indicators, particularly in early warning accuracy, response timeliness, risk level matching degree, and control of false alarms and missed alarms. Firstly, regarding early warning accuracy, the method of this invention maintains a high accuracy rate of over 94% in four typical archive anomaly scenarios, significantly higher than the approximately 70% of traditional methods, effectively solving the recognition limitations of traditional methods when processing complex multimodal data.

[0109] In terms of response timeliness, the method of this invention significantly shortens the early warning response time. This improvement is due to the real-time adjustment and weighting of the deep Q-network, which optimizes the early warning response link and greatly enhances the real-time risk intervention capability.

[0110] Regarding risk level matching accuracy, this invention maintains a accuracy rate above 90%, significantly higher than the approximately 60% level of traditional methods. Traditional methods rely on static threshold judgments, which are prone to misjudging minor deviations or underestimating serious risks. In contrast, this invention, through dynamically constructed knowledge graphs and comparisons with historical trends, makes level judgments more refined and accurate, ensuring the precision and targeting of early warning strategies. In terms of false positive and false negative rates, this invention controls the false positive rate to below 4% and the false negative rate to around 2%, significantly lower than the 13% false positive rate and nearly 10% false negative rate of traditional methods. Reducing false positives and false negatives not only lowers the cost of manual intervention but also improves the stability and reliability of the system, ensuring efficient operation.

[0111] In summary, this invention, by combining HGT graph neural network with improved MC-Dropout, overcomes the limitations of traditional methods in complex data environments, significantly improves anomaly identification capabilities, response timeliness, and risk assessment accuracy, and has broad practical application value and promotion prospects.

[0112] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for early warning of anomalies in urban construction archives based on artificial intelligence, characterized in that, Includes the following steps: S1. Collect multi-source archival data, perform preprocessing, and construct an archival feature database; S2. Extract the text features, drawing component features, and approval label features from the archive feature database to generate a multimodal feature vector; S3. Generate an archive knowledge graph and temporal feature sequence using multimodal feature vectors, and introduce them into the HGT graph neural network along with multimodal node features to output the archive structure embedding representation and evolution chain semantic features; S4. Compare the embedded representation of the archive structure with the features of the version evolution chain to generate potential abnormal records; S5. Based on the improved MC-Dropout, the structural offset difference, semantic conflict cosine similarity interpolation and temporal order offset difference of potential abnormal records are used to quantify the risk and output the anomaly type, anomaly location and risk level. S6. Submit the anomaly type, anomaly location, and risk level to the urban construction archives management platform, record the early warning output and review feedback information, and update the review log and permission records.

2. The method for early warning of urban construction archive anomalies based on artificial intelligence according to claim 1, characterized in that, S1 specifically includes: S11. Read the original files of urban construction archives, including text, images, drawings, approval process records, and location information. Convert the file content according to the character encoding format. Read the image and drawing data. Convert the image pixel format and drawing vector format according to the unified pixel standard and unified vector standard. Convert all images and drawings to a unified size and unified resolution. S12. Perform field extraction on the approval process record and location information. Read the approval node name, approval time, approval department and coordinate location in sequence according to the preset field names. Save the extracted fields with unified field names. Perform sentence segmentation and paragraph separation on the urban construction archive text. Separate continuous text into independent text segments according to periods and punctuation marks. Store each text segment in sequence. S13. After the urban construction archives text, images, drawings, approval process records and location information have been standardized by uniform encoding, uniform pixel format, uniform vector format and uniform field, they are classified and stored according to data type, and the constructed archive feature library is output.

3. The method for early warning of urban construction archive anomalies based on artificial intelligence according to claim 1, characterized in that, S2 specifically includes: S21. Read the urban construction archive text in the archive feature library line by line, perform word segmentation on each line of text and record the part of speech, combine the word order into text semantic fragments, output the initial features of the urban construction archive text, match the words in the initial features of the urban construction archive text with the preset word vector table, concatenate the matched word vectors in order into text semantic vectors, replace the unmatched words with preset placeholder vectors, and obtain the archive text features. S22. Input the urban construction archive drawings into the image semantic segmentation process, divide the drawings into drawing component areas, perform pixel classification on each area and mark it as component lines, component symbols and component annotations, generate drawing component features, read the approval process records in the order of fields, perform label extraction on the approval node name, approval time and approval department, and generate approval label features. S23. Match the document text features, drawing component features, and approval label features according to their respective document numbers, concatenate the three features in sequence into a unified feature vector, and output the concatenated unified feature vector as a multimodal feature vector.

4. The method for early warning of urban construction archive anomalies based on artificial intelligence according to claim 1, characterized in that, S3 specifically includes: S31. Bind the multimodal feature vector to the corresponding urban construction archive text entity, drawing component entity and approval event entity according to the archive number as multimodal node feature. Write the relationship type according to the relationship between the target entities. Record each relationship as an edge and record the time sequence of each edge as a time label to form an archive knowledge graph and time sequence feature sequence. S32. In the HGT graph neural network, a dedicated linear transformation matrix for each node type is preset for urban construction archive text entity, drawing component entity and approval event entity. The feature of each multimodal node is multiplied element by element with the corresponding node type linear transformation matrix and the bias is added to obtain the node type-specific representation. S33. Pre-set dedicated linear transformation matrices for design relationships, construction dependencies, and approval workflows respectively. Multiply each node type-specific representation with the corresponding relationship type linear transformation matrix element by element and add the offsets to obtain a message vector specific to the relationship type. S34. Convert the time label of each edge into a time vector, add the time vector element by element to the relation type-specific message vector to generate a meta-relation message vector with time information. In the HGT graph neural network, perform multi-head attention calculation for each type of meta-relation, multiply the meta-relation message vector with time information element by element with the corresponding preset attention weight and add them to obtain the meta-relation aggregation vector. S35. Perform residual join, layer normalization and feedforward transformation on the meta-relation aggregation vector in sequence, and perform residual join and layer normalization on the transformation result again to obtain the node representation after one update. S36. Repeatedly perform node updates according to the network layer number, perform relation type-specific linear transformation, multi-head attention calculation and node update operations on all nodes until all nodes generate the final representation; S37. Add the updated node representation of the current node element by element to the cumulative vector, move to the next node and repeat the operation until all nodes are processed to obtain the cumulative vector. Use the cumulative vector as the embedded representation of the archive structure. Divide the cumulative vector into multiple continuous segments in chronological order. Store each segment in the evolution chain semantic feature table to form the evolution chain semantic features.

5. The method for early warning of urban construction archive anomalies based on artificial intelligence according to claim 1, characterized in that, S4 specifically includes: S41. Read the archive structure embedding representation and evolution chain semantic features sequentially according to the archive number, use the archive structure embedding representation as node features, and use the evolution chain semantic features as time sequence features and write them into the comparison sequence. S42. Perform element-wise subtraction between each node feature and the features of its adjacent nodes, record the positions where the absolute value of the difference is greater than the set difference threshold as structural offset points, and write all structural offset points into the structural offset sequence in node order. S43. Perform element-wise cosine similarity calculation between the multimodal feature vector of each node and the multimodal feature vector of the previous time node. Record positions below the set similarity threshold as semantic conflict points, and write all semantic conflict points into the semantic conflict sequence in chronological order. S44. Read the temporal order index in the semantic features of the evolutionary chain, calculate the time interval between adjacent records according to the index order, record the positions where the time interval is greater than the set interval threshold as time breakpoints, and write all time breakpoints into the temporal order offset sequence. S45. Match the structural offset sequence, semantic conflict sequence, and time order offset sequence sequentially according to node number and time index. When the number of records of any node in any sequence exceeds the set offset threshold, mark the node as a potential anomaly and write the potential anomaly into the potential anomaly record.

6. The method for early warning of urban construction archive anomalies based on artificial intelligence according to claim 1, characterized in that, S5 specifically includes: S51. Input the archive structure embedding representation, evolution chain semantic features and multimodal node features of the identified potential abnormal record nodes into the improved MC-Dropout, perform multi-round sampling inference and risk uncertainty assessment, and output the mean and variance of the abnormal probability distribution of each node. S52. Using the mean and variance results of the anomaly probability obtained through the improved MC-Dropout, and combining the three types of indicators—structural offset difference, semantic conflict cosine similarity interpolation, and temporal order offset difference—to calculate the comprehensive risk score, and output the anomaly type, anomaly location, and risk level of the corresponding node.

7. The method for early warning of urban construction archive anomalies based on artificial intelligence according to claim 6, characterized in that, The improved MC-Dropout specifically includes: In each hidden layer of the HGT graph neural network, a Dropout control node is set up to receive the output vector of the previous layer. During each forward propagation, a 0-1 random occlusion vector with the same dimension as the output vector is generated for the current layer. The random occlusion vector is multiplied element by element with the output vector, and the neuron outputs at the unoccluded positions are retained to form a predicted output with a random occlusion network computation path. The above forward propagation is repeated a preset number of times for the same potentially abnormal record, generating different random occlusion vectors each time, obtaining the prediction output of a preset number of paths, and storing them sequentially according to the path number; For each path's prediction output, a temporal consistency constraint is introduced. The previous and next time nodes in the semantic features of the evolution chain of potential anomaly records are read. The path prediction values ​​corresponding to the three time nodes are read sequentially. The absolute values ​​of the differences between the current node and the previous node and the next node are calculated. The position where the absolute value of any of the two differences is greater than the set jump threshold is marked as a temporal anomaly point. All prediction outputs are multiplied by a preset reduction coefficient. For each path's predicted output, a relational consistency constraint is introduced. Potentially abnormal records are read from the set of nodes connected in the archive knowledge graph. For each adjacent node, the corresponding path prediction value is read. The absolute difference between the current node's prediction value and the prediction values ​​of each adjacent node is calculated and averaged. If the average difference is greater than the set relational deviation threshold, it is marked as a relational abnormal path, and all predicted outputs of the path are multiplied by a reduction factor. After the prediction output of all paths has been reduced, the prediction output value of each path is read. The average of the prediction outputs of all paths is used as the final anomaly score. The average of the squared differences between the output values ​​of all paths and the anomaly scores is used as the uncertainty score. The anomaly scores and uncertainty scores are stored in the results list according to the sample number. The obtained anomaly score and uncertainty score are added to the structural offset difference, semantic conflict cosine similarity interpolation and temporal order offset difference item by item, and then weighted and combined according to the preset weight coefficient to calculate the comprehensive risk score of potential anomaly records. When the overall risk score exceeds the set risk threshold, the node is marked as a high-risk anomaly, and the anomaly type, anomaly location, and overall risk score are written into the early warning output table.

8. The method for early warning of urban construction archive anomalies based on artificial intelligence according to claim 1, characterized in that, Specifically, S6 includes: submitting the anomaly type, anomaly location, and risk level to the urban construction archive management platform, which registers and displays the anomaly according to the archive number and pushes it to the corresponding responsible department, simultaneously recording the early warning output, manual review opinions, and handling results, writing the review feedback into the archive update log, and updating the corresponding permission records.

Images