Engineering document intelligent association and retrieval method based on knowledge graph

Abstract

The invention discloses a knowledge graph-based engineering document intelligent association and retrieval method, which comprises the following steps of: S1, collecting full-life-cycle multi-mode heterogeneous document data, and generating a standardized data set through semantic alignment; s2, synchronously extracting entities, graphs and metadata by using a LayoutLM model, and outputting an original knowledge graph; s3, aggregating discrete entities based on a hypergraph theory, and constructing a dynamic evolution knowledge graph; s4, mining recessive features by introducing an improved GBNN model of a space-time causal mutual information maximization mechanism, and complementing to generate an enhanced semantic index map; s5, natural language query is converted into intention nodes for sub-graph matching; and S6, calculating a decision support value degree, reordering and outputting a document list. According to the method, the cross-document implicit association mining problem is effectively solved, and the retrieval accuracy and the decision support efficiency are improved.

Description

Technical Field

[0001] This invention relates to the field of intelligent document processing and knowledge graphs, and in particular to a method for intelligent association and retrieval of engineering documents based on knowledge graphs. Background Technology

[0002] Intelligent document association retrieval methods based on hypergraph theory and knowledge graphs have been widely applied in engineering design, construction supervision, operation and maintenance, and other fields in recent years due to their powerful multimodal semantic understanding and complex logical reasoning capabilities, becoming an important development direction for realizing digital management of the entire engineering lifecycle. However, in practical applications, the entire engineering lifecycle management scenario faces many challenges such as heterogeneous document modalities, discrete data semantics, and obscure causal logic, and the actual effectiveness of knowledge graph-based retrieval methods is still constrained by many factors.

[0003] Most current engineering document retrieval methods rely on single-modal feature inputs, making it difficult to fully utilize multi-source information such as design specification text, CAD drawing vector graphics, and construction log time-series data. This results in a lack of comprehensive semantic modeling of engineering entities. Some systems only use simple keyword matching or shallow semantic similarity calculations, ignoring implicit causal tracing paths between documents, upstream and downstream event contexts, and physical evolution logic, thus limiting the depth and decision support capabilities of retrieval results. Furthermore, the retrieval logic process lacks a transparent explanation path, making it difficult to provide engineers with clear reasons for document recommendations, affecting the credibility and usability of retrieval results.

[0004] Furthermore, most existing knowledge graph retrieval technologies are based on static graph structures and fail to utilize dynamic hypergraph neural networks to perform high-order aggregation of discrete entities that cross documents and chapters and point to the same engineering event. They also lack in-depth mining of temporal causal mutual information, which makes it difficult for the model to effectively identify potential semantic associations that are not explicitly labeled. This makes it difficult to adapt to the complex scenarios of constantly evolving engineering events and dynamically growing documents, seriously affecting the practical value and accuracy of intelligent retrieval systems in real engineering practice.

[0005] Therefore, how to provide a knowledge graph-based intelligent association and retrieval method for engineering documents 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 intelligent association and retrieval method for engineering documents based on knowledge graphs. This invention fully integrates key steps such as multimodal semantic parsing, hypergraph evolution network construction, improved GBNN model latent feature mining, and physically guided ranking. It constructs an intelligent document retrieval process with semantic alignment standardization, temporal causal link evolution, latent association probability reasoning, and decision value assessment, achieving deep semantic association and accurate recommendation of engineering documents throughout their entire lifecycle. This invention utilizes the LayoutLM model to achieve deep fusion of graph and text features. By introducing an improved GBNN model with a spatiotemporal causal mutual information maximization mechanism, it forces node features to approximate the global temporal causal structure. It possesses advantages such as comprehensive multimodal information processing, strong latent association mining depth, and high decision support for retrieval results. It can significantly improve document recall, decision value, and interpretability in complex engineering scenarios, thereby effectively solving problems such as lack of document semantic modeling, insufficient latent association mining, and a single ranking strategy in existing methods.

[0007] According to an embodiment of the present invention, an intelligent association and retrieval method for engineering documents based on knowledge graphs includes the following steps: S1. Collect multimodal heterogeneous document data throughout the entire project lifecycle and preprocess it, perform semantic alignment on heterogeneous terms, and generate a standardized document dataset; S2. Input the standardized document dataset into the LayoutLM model for document semantic parsing, and use the joint attention mechanism to simultaneously extract entity, graph geometric features and document metadata, identify the dependency relationships and attribute constraints between entities, and output the original knowledge graph triple data. S3. Construct a document association evolution network based on hypergraph theory, map the original knowledge graph triple data to the hypergraph space, aggregate discrete entities that cross documents and chapters and point to the same project event, and introduce timestamps and causal logic tags to weight the hyperedges to generate a dynamic evolution knowledge graph. S4. Latent feature mining is performed on the dynamic evolution knowledge graph. By introducing an improved GBNN model that maximizes spatiotemporal causal mutual information, the probability of multi-hop path similarity is calculated. Probabilistic reasoning and uncertainty quantification are performed. Combined with confidence thresholds, potential semantic associations are identified, and the enhanced semantic index graph is completed and corrected. S5. Receive the user's natural language query request and convert it into intent graph nodes. Perform subgraph matching retrieval in the enhanced semantic index graph to retrieve a set of candidate document nodes. S6. Reorder the candidate document nodes in the candidate document node set according to their relevance, calculate the decision support value of the candidate document nodes, and output the final list of related documents from high to low based on the decision support value.

[0008] Optionally, S1 specifically includes: S11. Collect multimodal heterogeneous document data throughout the entire lifecycle of the project, including text data contained in the design specification, vector graphics data contained in the CAD drawings, and unstructured text contained in the meeting minutes. Perform noise removal processing on the text data, vector graphics data, and unstructured text through data cleaning, and delete duplicate items and garbled characters in the data. S12. Call the pre-built ontology tree, use the BERT model to perform semantic matching to map the heterogeneous terms in the denoised multimodal heterogeneous document data to the standard terms in the ontology tree, replace the heterogeneous terms with the corresponding standard terms in the ontology tree, perform semantic alignment operation on the heterogeneous terms, and store the semantically aligned text data, vector graphics data and unstructured text according to the preset format to generate a standardized document dataset.

[0009] Optionally, S2 specifically includes: S21. Input the standardized document dataset into the LayoutLM model, convert the contained text data and unstructured text into word vectors through the embedding layer, convert the vector graphics data into image patch feature vectors, and concatenate the word vectors and image patch feature vectors to construct a multimodal input vector sequence. S22. Input the multimodal input vector sequence into the Transformer encoder of the LayoutLM model, and use the self-attention layer to calculate the dot product between each word vector and all image patch feature vectors to obtain the interaction matrix; S23. Normalize each row of the interaction matrix to obtain the weight coefficients, sum the image patch feature vectors according to the weight coefficients, and add them to the word vectors to update the word vectors. S24. Input the updated word vectors into the preset linear layer, calculate the classification score vector of each word vector belonging to the preset entity category, perform an exponential operation on each value in the classification score vector and divide it by the sum of all exponential values ​​to obtain the probability distribution, and select the category with the highest probability as the corresponding entity label. S25. Read the vector node coordinate file in the vector graphics data, parse all horizontal and vertical coordinate values, and encapsulate them into graphic geometric features. Read the timestamp field in the document attribute table, extract the year, month, day, hour, minute, and second values ​​as time data, and combine them into document metadata. S26. Analyze the positions of all word vectors labeled as entity tags in the text, use a bidirectional long short-term memory network to calculate the semantic correlation between every two entity word vectors, and generate a set of directed edges for the dependency relationships between entities. S27. Perform key-value matching between the horizontal and vertical coordinate data extracted from the vector graphic data and the spatial distance entity label. Write the successfully matched coordinate data into the attribute slot of the corresponding entity and attach the graphic geometric features as attribute nodes to the corresponding entity. S28. Perform key-value matching between the extracted year, month, day, hour, minute, and second values ​​and the time event entity tags. Write the successfully matched time data into the attribute slot of the corresponding entity and attach the document metadata as an attribute node to the corresponding entity. S29. Using entities as head nodes and attributes as tail nodes, mark ownership relationships as attribute edges and dependency relationships as entity edges, construct the original knowledge graph triple data in the order of head nodes, relationships, and tail nodes, and output it.

[0010] Optionally, S3 specifically includes: S31. Read all entity nodes in the original knowledge graph triple data, build an index based on the document number and chapter number to which each entity node belongs, map all entity nodes to the set of vertices in the hypergraph space, and establish a one-to-one correspondence between vertex IDs and entity nodes. S32. Traverse each vertex in the vertex set, calculate the cosine similarity between the current vertex and all other vertices in the documents and chapters, select vertices whose cosine similarity values ​​are greater than the preset similarity threshold and point to the same project event, and classify the selected vertices into the same hyperedge to generate a hyperedge set. S33. The word vectors that will be labeled as entity labels are used as vertex feature vectors. Using the hyperedge aggregation mechanism of the dynamic hypergraph neural network, the vertex feature vectors contained in each hyperedge are read. The mean pooling operation is performed on all vertex feature vectors in the same hyperedge, and the mean pooling result is used as the overall feature representation of the corresponding hyperedge. S34. Scan the related documents of the hyperedge containing sentences with preset causal association words, extract the word vectors of the corresponding sentences and take the average value as the causal logic label, read the timestamp of the related documents of each hyperedge, convert the timestamp into a time value vector and concatenate it with the numerical code of the causal logic label in dimension, and use the preset linear transformation layer to map the concatenated vector into a scalar weight value. S35. Multiply the calculated scalar weight values ​​element-wise with the overall feature representation of the corresponding hyperedge to obtain the weighted hyperedge features. Use the weighted hyperedge features as the temporal causal link connecting the vertices to construct a dynamic evolutionary knowledge graph.

[0011] Optionally, S4 specifically includes: S41. Construct an improved GBNN model. Use the weighted hyperedge features of the vertices represented in the dynamic evolution knowledge graph as feature vectors and arrange them in order of node ID to form a feature matrix. Input the feature matrix into the input layer and perform a linear transformation on the feature matrix through a preset weight matrix to obtain the initial node embedding matrix. S42. Input the initial node embedding matrix into the graph convolutional layer, read the adjacency matrix of the dynamic evolution knowledge graph, calculate the product of the initial node embedding matrix and the adjacency matrix, aggregate the neighbor features of each node, and use the ReLU activation function to perform nonlinear mapping on the aggregation result to obtain the neighbor aggregated feature vector. S43. Input the neighbor aggregation feature vector into the gated recurrent layer, calculate the difference vector between the neighbor aggregation feature vector and the hidden state of the node at the current time, input the difference vector into the Sigmoid function to calculate the update gate value, and at the same time calculate the weighted sum vector between the neighbor aggregation feature vector and the hidden state of the node at the current time, input the weighted sum vector into the Tanh function to calculate the candidate node state vector. S44. Calculate the memory retention amount by performing element-wise multiplication of the update gate value with the hidden state of the node at the previous time step, calculate the memory update amount by performing element-wise multiplication of the difference between 1 and the update gate value with the candidate node state vector, perform vector addition operation on the memory retention amount and the memory update amount, update the hidden state of the node at the current time step, and obtain the node embedding matrix. S45. Read all the temporal causal links in the dynamic evolution knowledge graph, arrange the node IDs contained in each temporal causal link in chronological order according to the timestamp, obtain the causal path sequence, input it into the preset bidirectional LSTM network for encoding, extract the concatenation vector of the forward output and the reverse output, and generate the global temporal causal graph structure vector. S46. Introduce a spatiotemporal causal mutual information maximization mechanism, calculate the dot product value of each local node representation in the node embedding matrix and the global temporal causal graph structure vector, use the dot product value as the spatiotemporal mutual information score, construct a loss function with negative spatiotemporal mutual information score as the objective, and use the stochastic gradient descent algorithm to update the parameters of the graph convolutional layer and the gated recurrent layer in reverse, so that the local node representation approximates the global temporal causal graph structure vector in the feature space. S47. Embed the updated node into the matrix input-output layer, calculate the dot product similarity between any two non-connected nodes in the dynamic evolution knowledge graph, perform Softmax normalization on the dot product similarity, and obtain the multi-hop path similarity probability between nodes, i.e. the prediction probability. S48. Use the Monte Carlo sampling method to randomly select 10% of the edges from the dynamic evolution knowledge graph to form a negative sample set. Calculate the multi-hop path similarity probability of each edge in the negative sample set. Use the binary cross-entropy loss function to calculate the error value between the predicted probability and the true label, and use it as a quantitative indicator of the uncertainty of the probability prediction. S49. Traverse all non-connected node pairs, read the multi-hop path similarity probability and uncertainty quantification index, filter out target node pairs with multi-hop path similarity probability greater than the preset confidence threshold and uncertainty quantification index less than the preset uncertainty threshold, identify the target node pairs as potential semantic associations, establish new connection edges between the two target nodes, set the weight of the new connection edge to the corresponding multi-hop path similarity probability, add the new connection edge to the dynamic evolution knowledge graph, and generate an enhanced semantic index graph.

[0012] Optionally, the step of calculating the error value between the predicted probability and the true label using the binary cross-entropy loss function specifically includes: Read the corresponding real label values ​​in the dynamic evolution knowledge graph, substitute the real label values ​​into the real variable terms in the binary cross-entropy loss function formula, where a real label value of 1 represents a connection between nodes and a real label value of 0 represents no connection between nodes. Logarithmic operation is performed on the predicted probability to calculate the natural logarithm of the predicted probability. The true label value is multiplied by the natural logarithm to obtain the first product term. At the same time, the difference between the value 1 and the true label value is calculated. Negative logarithmic operation is performed on the predicted probability to calculate the natural logarithm of the difference between the value 1 and the predicted probability. The difference is multiplied by the natural logarithm of the difference to obtain the second product term. Add the first product term to the second product term, calculate the negative of the sum of the two product terms, and define the negative of the sum as the prediction error value for a single sample.

[0013] Optionally, S5 specifically includes: S51. Receive the user's natural language query request, input the natural language query request into the preset BERT model for semantic encoding, extract the word vector features of the natural language query request, input the word vector features into the preset fully connected layer to calculate the classification score, select the category with the highest classification score as the query intent category, and create the corresponding intent graph node according to the query intent category. S52. Perform precise matching of the keyword entities in the natural language query request with the entity nodes in the enhanced semantic index graph using entity name tags. Take the entity nodes that successfully match the keyword entities as anchor nodes, expand the first-level neighbor nodes around the anchor nodes, extract the subgraph structure formed by the anchor nodes and the first-level neighbor nodes as candidate subgraphs, traverse all nodes in the candidate subgraphs, filter out the nodes whose node type belongs to the document type, and mark them as candidate document nodes. S53. Read all connection paths between anchor nodes and each candidate document node in the candidate subgraph, traverse each connection path, perform cumulative calculation on the weight values ​​of all connection edges contained in the connection path to obtain the total path weight of each connection path, filter out the target path with a total path weight less than the preset path threshold, and extract the candidate document node corresponding to the end of the target path as the initial screening node. S54. Extract word vector features from intent graph nodes and node embedding vectors from preliminary screening nodes. Calculate the Euclidean distance between word vector features and node embedding vectors in the multidimensional feature space. Define the Euclidean distance as the semantic distance between intent graph nodes and preliminary screening nodes. S55. Sort all the initial screening nodes in ascending order of semantic distance value, select the top-ranked initial screening nodes for aggregation, remove duplicate nodes from the list, and generate a candidate document node set.

[0014] Optionally, S6 specifically includes: S61. Read the document metadata of each candidate document node in the candidate document node set, extract the timestamp field containing the year, month, day, hour, minute and second values, and arrange all candidate document nodes linearly according to the order of the timestamp values ​​to construct a timeline sequence. S62. Traverse any two candidate document nodes in the candidate document node set, query the causal tracing path connecting the two candidate document nodes in the enhanced semantic index graph, read the pointing relationship of all connecting edges on the causal tracing path, and determine the upstream and downstream positions of the two candidate document nodes in the project event context. S63. Assign values ​​to candidate document nodes based on their upstream and downstream positions. If a candidate document node is located at the upstream starting point in the project event network, assign the preset basic decision score to the candidate document node. If a candidate document node is located at the downstream end, calculate the number of path hops from the upstream starting point node to the current node, divide the basic decision score by the number of path hops to obtain the attenuated score, and use it as the decision support value of the candidate document node. S64. Sort all candidate document nodes in descending order of decision support value, generate a final list of associated documents based on the sorting order, read the decision support value of each candidate document node in the final list of associated documents, and take the documents corresponding to the candidate document nodes whose values ​​exceed the preset value threshold as recommended documents, and output the final list of associated documents.

[0015] The beneficial effects of this invention are: This invention addresses the challenges of multimodal heterogeneity of document data, difficulty in latent semantic association mining, and low retrieval decision value throughout the entire project lifecycle by constructing a document association evolution network based on hypergraph theory and deploying an improved GBNN model. It employs the LayoutLM model for document semantic parsing, combined with a joint attention mechanism to simultaneously extract entity, graphical geometric features, and metadata, outputting a raw knowledge graph triple. By mapping a standardized document dataset to a hypergraph space, and utilizing hyperedge aggregation and temporal causal weighting, a dynamic evolutionary knowledge graph containing temporal causal links is constructed. The dynamic evolutionary knowledge graph is then input into an improved GBNN model with a spatiotemporal causal mutual information maximization mechanism, performing local topological aggregation and global causal structure approximation, outputting the multi-hop path similarity probability and uncertainty quantification index between nodes. During the latent feature mining stage, potential semantic associations are identified by combining confidence thresholds and uncertainty indices, completing and correcting the knowledge graph. Furthermore, a physically guided adaptive sorting mechanism is used to analyze causal tracing paths, calculate decision support value, and output the final list of associated documents. Ultimately, it achieves deep semantic understanding, automatic discovery of implicit associations, and decision-level accurate recommendation of multimodal heterogeneous engineering documents, effectively improving the comprehensiveness of document retrieval, the accuracy of model reasoning, and the ability to support engineering management decisions. Attached Figure Description

[0016] 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: Figure 1 This is a flowchart of a knowledge graph-based intelligent association and retrieval method for engineering documents proposed in this invention; Figure 2 This is a flowchart of the multimodal engineering document semantic parsing and original knowledge graph construction based on the LayoutLM model proposed in this invention; Figure 3 This is a flowchart of the latent feature mining and graph completion process of the improved GBNN model based on the mechanism of maximizing spatiotemporal causal mutual information proposed in this invention. Detailed Implementation

[0017] 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.

[0018] refer to Figures 1-3 A knowledge graph-based intelligent association and retrieval method for engineering documents includes the following steps: S1. Collect multimodal heterogeneous document data throughout the entire lifecycle of the project. The multimodal heterogeneous document data includes text data in the design specification, vector graphics data in the CAD drawings, and unstructured text in the meeting minutes. The data is then preprocessed, and ontology trees are used to semantically align heterogeneous terms to generate a standardized document dataset. S2. Input the standardized document dataset into the LayoutLM model for document semantic parsing, and use the joint attention mechanism to simultaneously extract entity, graph geometric features and document metadata, identify the dependency relationships and attribute constraints between entities, and output the original knowledge graph triple data containing entity nodes, attribute nodes and initial relation edges. S3. Construct a document association evolution network based on hypergraph theory, map the original knowledge graph triple data to the hypergraph space, use the hyperedge aggregation mechanism of dynamic hypergraph neural network to aggregate discrete entities that cross documents and chapters and point to the same project event, and introduce timestamps and causal logic labels to weight the hyperedges to generate a dynamic evolution knowledge graph containing temporal causal links. S4. Latent feature mining is performed on the dynamic evolution knowledge graph. By introducing an improved GBNN model that maximizes spatiotemporal causal mutual information, the probability of multi-hop path similarity is calculated. Probabilistic reasoning and uncertainty quantification are performed. Combined with confidence thresholds, potential semantic associations are identified, and the enhanced semantic index graph is completed and corrected. S5. Receive the user's natural language query request and convert it into intent graph nodes. Perform subgraph matching retrieval in the enhanced semantic index graph. By calculating the semantic distance and path weight between the query intent node and the graph node, retrieve the candidate document node set. S6. Reorder the candidate document nodes in the candidate document node set according to their relevance. Analyze the upstream and downstream positions of the candidate documents in the engineering event context based on the causal tracing path in the enhanced semantic index graph. Calculate the decision support value of the candidate document nodes and output the final list of associated documents from high to low based on the decision support value.

[0019] This implementation significantly improves the accuracy of engineering document retrieval and the efficiency of decision support. By collecting multimodal data throughout its entire lifecycle and using the LayoutLM model for parsing, unified semantic modeling of text, graphics, and metadata is achieved, effectively solving the information silo problem caused by heterogeneous data sources. A dynamic evolutionary knowledge graph based on hypergraph theory is constructed, using a hyperedge aggregation mechanism to associate discrete entities across documents and introducing temporal causal logic, greatly enhancing the ability to express the complex context of engineering events. Through an improved GBNN model for latent feature mining and spatiotemporal mutual information calculation, potential semantic relationships can be accurately identified, filling gaps in the knowledge graph. Finally, the decision support value is calculated by combining causal tracing paths and then reordered to ensure that the output document list highly matches the actual needs of engineering projects, significantly reducing manual screening costs and providing efficient and accurate knowledge services for engineering construction.

[0020] In this embodiment, S1 specifically includes: S11. Collect multimodal heterogeneous document data throughout the entire lifecycle of the project, including text data contained in the design specification, vector graphics data contained in the CAD drawings, and unstructured text contained in the meeting minutes. Perform noise removal processing on the text data, vector graphics data, and unstructured text through data cleaning, and delete duplicate items and garbled characters in the data. S12. Call the pre-built ontology tree, use the BERT model to perform semantic matching to map the heterogeneous terms in the denoised multimodal heterogeneous document data to the standard terms in the ontology tree, replace the heterogeneous terms with the corresponding standard terms in the ontology tree, perform semantic alignment operation on the heterogeneous terms, and store the semantically aligned text data, vector graphics data and unstructured text according to the preset format to generate a standardized document dataset.

[0021] In this embodiment, S2 specifically includes: S21. Input the standardized document dataset into the LayoutLM model, and convert the contained text data and unstructured text into word vectors through the embedding layer, and convert the vector graphics data into image patch feature vectors. Concatenate the word vectors and image patch feature vectors to construct a unified multimodal input vector sequence. S22. Input the multimodal input vector sequence into the Transformer encoder of the LayoutLM model, and use the self-attention layer to calculate the dot product between each word vector and all image patch feature vectors to obtain the text-graphic interaction matrix. S23. Normalize each row of the interaction matrix to obtain the weight coefficients. Then, sum the image patch feature vectors according to the weight coefficients and add them to the word vectors to update the word vectors so that the text features contain the visual context information of the graphics. S24. Input the updated word vectors into the preset linear layer, calculate the classification score vector of each word vector belonging to the preset entity category, perform an exponential operation on each value in the classification score vector and divide it by the sum of all exponential values ​​to obtain the probability distribution, and select the category with the highest probability as the corresponding entity label. S25. Read the vector node coordinate file in the vector graphics data, parse all horizontal and vertical coordinate values, and encapsulate them into graphic geometric features. Read the timestamp field in the document attribute table, extract the year, month, day, hour, minute, and second values ​​as time data, and combine them into document metadata. S26. Analyze the positions of all word vectors labeled as entity tags in the text, use a bidirectional long short-term memory network to calculate the semantic correlation between every two entity word vectors, and generate a set of directed edges for the dependency relationships between entities. S27. Perform key-value matching between the horizontal and vertical coordinate data extracted from the vector graphic data and the spatial distance entity label. Write the successfully matched coordinate data into the attribute slot of the corresponding entity and attach the graphic geometric features as attribute nodes to the corresponding entity. S28. Perform key-value matching between the extracted year, month, day, hour, minute, and second values ​​and the time event entity tags. Write the successfully matched time data into the attribute slot of the corresponding entity and attach the document metadata as an attribute node to the corresponding entity. S29. Using entities as head nodes and attributes as tail nodes, mark ownership relationships as attribute edges and dependency relationships as entity edges, construct the original knowledge graph triple data in the order of head nodes, relationships, and tail nodes, and output it.

[0022] This invention achieves deep semantic parsing and structured knowledge extraction from heterogeneous engineering documents by introducing the LayoutLM model combined with a multimodal feature fusion mechanism. Text word vectors and vector graphic image patch features are concatenated, and an interaction matrix is ​​calculated using a self-attention layer to update the feature representation, enabling text features to be integrated with the visual context information of the graphics, effectively eliminating the cross-modal semantic gap. A BiLSTM network is used to calculate the semantic correlation between entities and construct directed edges. Key-value matching accurately maps the geometric coordinates of the graphics and document metadata to entity attribute slots. This invention can automatically identify entity dependencies and attribute constraints, solving the problem of the difficulty in quantifying unstructured information. Finally, an original knowledge graph containing entity nodes, attribute edges, and entity edges is generated, providing high-precision triplet data for subsequent graph construction, significantly improving the comprehensiveness and accuracy of document knowledge extraction.

[0023] In this embodiment, S3 specifically includes: S31. Read all entity nodes in the original knowledge graph triple data, build an index based on the document number and chapter number to which each entity node belongs, map all entity nodes to the set of vertices in the hypergraph space, and establish a one-to-one correspondence between vertex IDs and entity nodes. S32. Traverse each vertex in the vertex set, calculate the cosine similarity between the current vertex and all other vertices in the documents and chapters, select vertices whose cosine similarity values ​​are greater than the preset similarity threshold and point to the same project event, and classify the selected vertices into the same hyperedge to generate a hyperedge set containing multiple related vertices. S33. The word vectors that will be labeled as entity labels are used as vertex feature vectors. Using the hyperedge aggregation mechanism of the dynamic hypergraph neural network, the vertex feature vectors contained in each hyperedge are read. The mean pooling operation is performed on all vertex feature vectors in the same hyperedge, and the mean pooling result is used as the overall feature representation of the corresponding hyperedge. S34. Scan the related documents of the hyperedge containing sentences with preset causal association words, extract the word vectors of the corresponding sentences and take the average value as the causal logic label, read the timestamp of the related documents of each hyperedge, convert the timestamp into a time value vector and concatenate it with the numerical code of the causal logic label in dimension, and use the preset linear transformation layer to map the concatenated vector into a scalar weight value. S35. Multiply the calculated scalar weight values ​​element-wise with the overall feature representation of the corresponding hyperedge to obtain the weighted hyperedge features. Use the weighted hyperedge features as the temporal causal link connecting the vertices to construct a dynamic evolutionary knowledge graph.

[0024] In this embodiment, S4 specifically includes: S41. Construct an improved GBNN model that includes an input layer, a graph convolutional layer, a gated recurrent layer, and an output layer. Use the weighted hyperedge features of the vertices represented in the dynamically evolving knowledge graph as feature vectors, arrange them in order of node ID to form a feature matrix, input the feature matrix into the input layer, and perform a linear transformation on the feature matrix through a preset weight matrix to obtain the initial node embedding matrix. S42. Input the initial node embedding matrix into the graph convolutional layer, read the adjacency matrix of the dynamic evolution knowledge graph, calculate the product of the initial node embedding matrix and the adjacency matrix, aggregate the neighbor features of each node, and use the ReLU activation function to perform nonlinear mapping on the aggregation result to obtain the neighbor aggregated feature vector containing the local topology. S43. Input the neighbor aggregation feature vector into the gated recurrent layer, calculate the difference vector between the neighbor aggregation feature vector and the hidden state of the node at the current time, input the difference vector into the Sigmoid function to calculate the update gate value, and at the same time calculate the weighted sum vector between the neighbor aggregation feature vector and the hidden state of the node at the current time, input the weighted sum vector into the Tanh function to calculate the candidate node state vector. S44. Calculate the memory retention amount by performing element-wise multiplication of the update gate value with the hidden state of the node at the previous time step, calculate the memory update amount by performing element-wise multiplication of the difference between 1 and the update gate value with the candidate node state vector, perform vector addition operation on the memory retention amount and the memory update amount, update the hidden state of the node at the current time step, and obtain the node embedding matrix containing the local node representation. S45. Read all the temporal causal links in the dynamic evolution knowledge graph, arrange the node IDs contained in each temporal causal link in chronological order according to the timestamp, obtain the causal path sequence, input it into the preset bidirectional LSTM network for encoding, extract the concatenation vector of the forward output and the reverse output, and generate the global temporal causal graph structure vector. S46. Introduce a spatiotemporal causal mutual information maximization mechanism, calculate the dot product value of each local node representation in the node embedding matrix and the global temporal causal graph structure vector, use the dot product value as the spatiotemporal mutual information score, construct a loss function with negative spatiotemporal mutual information score as the objective, and use the stochastic gradient descent algorithm to update the parameters of the graph convolutional layer and the gated recurrent layer in reverse, so that the local node representation approximates the global temporal causal graph structure vector in the feature space. S47. Embed the updated node into the matrix input-output layer, calculate the dot product similarity between any two non-connected nodes in the dynamic evolution knowledge graph, perform Softmax normalization on the dot product similarity, and obtain the multi-hop path similarity probability between nodes, i.e. the prediction probability. S48. Use the Monte Carlo sampling method to randomly select 10% of the edges from the dynamic evolution knowledge graph to form a negative sample set. Calculate the multi-hop path similarity probability of each edge in the negative sample set. Use the binary cross-entropy loss function to calculate the error value between the predicted probability and the true label, and use it as a quantitative indicator of the uncertainty of the probability prediction. S49. Traverse all non-connected node pairs, read the multi-hop path similarity probability and uncertainty quantification index, filter out target node pairs with multi-hop path similarity probability greater than the preset confidence threshold and uncertainty quantification index less than the preset uncertainty threshold, identify the target node pairs as potential semantic associations, establish new connection edges between the two target nodes, set the weight of the new connection edge to the corresponding multi-hop path similarity probability, add the new connection edge to the dynamic evolution knowledge graph, and generate an enhanced semantic index graph.

[0025] This invention introduces an improved GBNN model combined with a spatiotemporal causal mutual information maximization mechanism to achieve deep feature mining and implicit association completion in dynamically evolving knowledge graphs. The weighted hyperedge feature matrix is ​​aggregated with neighbor topology information via graph convolution, and temporal states are fused using a gated recurrent layer to generate node embeddings containing local structures. A global structure vector is obtained by encoding temporal causal links using bidirectional LSTM, and mutual information maximization loss is introduced to optimize model parameters, making the local representation approximate the global causal structure. This invention can accurately calculate the multi-hop path similarity probability of unconnected nodes and uses uncertainty quantification indicators to screen high-confidence associations. While ensuring the reliability of inference, it effectively identifies potential semantic associations and automatically establishes new connections, generating an enhanced semantic index graph, significantly improving the knowledge completeness and complex semantic relationship inference capabilities of the graph.

[0026] In this embodiment, the step of calculating the error value between the predicted probability and the true label using the binary cross-entropy loss function specifically includes: Read the corresponding real label values ​​in the dynamic evolution knowledge graph, substitute the real label values ​​into the real variable terms in the binary cross-entropy loss function formula, where a real label value of 1 represents a connection between nodes and a real label value of 0 represents no connection between nodes. Logarithmic operation is performed on the predicted probability to calculate the natural logarithm of the predicted probability. The true label value is multiplied by the natural logarithm to obtain the first product term. At the same time, the difference between the value 1 and the true label value is calculated. Negative logarithmic operation is performed on the predicted probability to calculate the natural logarithm of the difference between the value 1 and the predicted probability. The difference is multiplied by the natural logarithm of the difference to obtain the second product term. Add the first product term to the second product term, calculate the negative of the sum of the two product terms, and define the negative of the sum as the prediction error value for a single sample.

[0027] In this embodiment, S5 specifically includes: S51. Receive the user's natural language query request, input the natural language query request into the preset BERT model for semantic encoding, extract the word vector features of the natural language query request, input the word vector features into the preset fully connected layer to calculate the classification score, select the category with the highest classification score as the query intent category, and create the corresponding intent graph node according to the query intent category. S52. Perform precise matching of the keyword entities in the natural language query request with the entity nodes in the enhanced semantic index graph using entity name tags. Take the entity nodes that successfully match the keyword entities as anchor nodes, expand the first-level neighbor nodes around the anchor nodes, extract the subgraph structure formed by the anchor nodes and the first-level neighbor nodes as candidate subgraphs, traverse all nodes in the candidate subgraphs, filter out the nodes whose node type belongs to the document type, and mark them as candidate document nodes. S53. Read all connection paths between anchor nodes and each candidate document node in the candidate subgraph, traverse each connection path, perform cumulative calculation on the weight values ​​of all connection edges contained in the connection path to obtain the total path weight of each connection path, filter out the target path with a total path weight less than the preset path threshold, and extract the candidate document node corresponding to the end of the target path as the initial screening node. S54. Extract word vector features from intent graph nodes and node embedding vectors from preliminary screening nodes. Calculate the Euclidean distance between word vector features and node embedding vectors in the multidimensional feature space. Define the Euclidean distance as the semantic distance between intent graph nodes and preliminary screening nodes. S55. Sort all the initial screening nodes in ascending order of semantic distance value, select the top-ranked initial screening nodes for aggregation, remove duplicate nodes from the list, and generate a candidate document node set.

[0028] In this embodiment, S6 specifically includes: S61. Read the document metadata of each candidate document node in the candidate document node set, extract the timestamp field containing the year, month, day, hour, minute and second values, and arrange all candidate document nodes linearly according to the order of the timestamp values ​​to construct a timeline sequence. S62. Traverse any two candidate document nodes in the candidate document node set, query the causal tracing path connecting the two candidate document nodes in the enhanced semantic index graph, read the pointing relationship of all connecting edges on the causal tracing path, and determine the upstream and downstream positions of the two candidate document nodes in the project event context. S63. Assign values ​​to candidate document nodes based on their upstream and downstream positions. If a candidate document node is located at the upstream starting point in the project event network, assign the preset basic decision score to the candidate document node. If a candidate document node is located at the downstream end, calculate the number of path hops from the upstream starting point node to the current node, divide the basic decision score by the number of path hops to obtain the attenuated score, and use it as the decision support value of the candidate document node. S64. Sort all candidate document nodes in descending order of decision support value, generate a final list of associated documents based on the sorting order, read the decision support value of each candidate document node in the final list of associated documents, and take the documents corresponding to the candidate document nodes whose values ​​exceed the preset value threshold as recommended documents, and output the final list of associated documents.

[0029] Example 1: To verify the feasibility of this invention in practice, it was applied to the management platform of a large-scale intelligent hydropower project under a provincial energy group. This project is an ultra-high arch dam hydropower project with an investment exceeding 50 billion yuan and a construction period of up to 8 years, encompassing complex work areas such as dam pouring and underground powerhouse excavation. During the project construction, a massive amount of engineering documents were generated, including design drawings, construction plans, supervision logs, monitoring reports, and equipment manuals, totaling over 2 million documents. Due to the large number of participating units, the documents were scattered in origin, inconsistent in standards, and contained complex semantic relationships but lacked structured links. For example, when cracks appeared during dam pouring, personnel found it difficult to immediately link to corresponding design change orders or historical handling solutions, leading to delays in risk management.

[0030] To address the aforementioned pain points, the method of this invention has been deeply deployed and applied in this intelligent hydropower engineering management platform. First, the platform uses ETL tools to aggregate standardized document datasets scattered across various business systems into a data lake. Then, it utilizes the LayoutLM model proposed in this invention to perform multimodal semantic parsing of the documents. The model can automatically identify key entities in the documents, such as "dam foundation corridor," "curtain grouting," and "Unit 2," and accurately extract the graphic geometric features and coordinate information of components from vector graphics. Simultaneously, it associates the timestamps and metadata in the document attribute table to construct an original knowledge graph containing entities, attributes, and relationships. Subsequently, the system uses an improved GBNN model to perform deep feature mining on the dynamically evolving knowledge graph. Through a mechanism that maximizes spatiotemporal causal mutual information, the local node representation can approximate the global temporal causal graph structure, thereby identifying implicit semantic relationships between documents. For example, the system can automatically establish a potential association between a real-time monitoring report on "foundation corridor seepage" and a "geological defect treatment plan" from five years ago, and recommend it to engineering technicians, thus activating and utilizing tacit knowledge. The table below shows the measured performance comparison data between the present invention and traditional methods in the task of intelligent association and retrieval of engineering documents: Table 1. Performance Comparison Data Between the Invention and Traditional Methods in Intelligent Retrieval of Hydropower Engineering Documents

[0031] Based on the comparative data shown in Table 1, it can be seen that the knowledge graph-based intelligent association and retrieval method for engineering documents proposed in this invention exhibits significant performance advantages over traditional retrieval methods in the management of large-scale hydropower projects, especially in key indicators such as retrieval accuracy, response speed, implicit association discovery capability, and decision support efficiency.

[0032] In terms of retrieval accuracy, this invention maintains a high level of over 95% in all four typical business scenarios, far exceeding the average accuracy of traditional systems at approximately 56%. For example, in the scenario of "underground cavern surrounding rock stability monitoring," traditional methods, limited by the mechanical logic of keyword matching, only achieve an accuracy of 50.0%. In contrast, this invention, by integrating the multimodal semantic understanding of the LayoutLM model with the associative reasoning mechanism of knowledge graphs, achieves an accuracy of 95.6%, effectively solving the problem of semantic recognition bias under complex engineering terminology.

[0033] Regarding response timeliness, this invention optimizes the graph index and query path through an improved GBNN model and a mechanism for maximizing spatiotemporal causal mutual information, significantly shortening response time. In four business scenarios, the average response time is reduced from 8.4 seconds using traditional methods to approximately 1.3 seconds, a nearly six-fold speedup, ensuring that engineering technicians can instantly obtain critical information from massive amounts of data. For example, the traditional response time for "metal structure installation quality inspection" is 7.8 seconds, while this invention only requires 1.1 seconds, greatly improving the real-time nature of on-site construction decisions.

[0034] In terms of latent association discovery and decision support, this invention demonstrates capabilities unmatched by traditional methods. The cumulative number of latent associations discovered by this invention far exceeds that of traditional methods, particularly in the area of ​​"temperature control and crack prevention in dam concrete pouring," where 385 latent associations were found, compared to only 12 in traditional methods. This directly boosted decision satisfaction from approximately 62% to over 93%, while significantly reducing manual screening time from an average of 24 minutes to about 4.4 minutes. Overall, this invention, by constructing a dynamically evolving knowledge graph, achieves efficient, accurate, and intelligent engineering knowledge management, effectively filling the technical gaps in latent knowledge mining and real-time decision support offered by traditional methods.

[0035] 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 intelligent association and retrieval of engineering documents based on knowledge graphs, characterized in that, Includes the following steps: S1. Collect multimodal heterogeneous document data throughout the entire project lifecycle and preprocess it, perform semantic alignment on heterogeneous terms, and generate a standardized document dataset; S2. Input the standardized document dataset into the LayoutLM model for document semantic parsing, and use the joint attention mechanism to simultaneously extract entity, graph geometric features and document metadata, identify the dependency relationships and attribute constraints between entities, and output the original knowledge graph triple data. S3. Construct a document association evolution network based on hypergraph theory, map the original knowledge graph triple data to the hypergraph space, aggregate discrete entities that cross documents and chapters and point to the same project event, and introduce timestamps and causal logic tags to weight the hyperedges to generate a dynamic evolution knowledge graph. S4. Latent feature mining is performed on the dynamic evolution knowledge graph. By introducing an improved GBNN model that maximizes spatiotemporal causal mutual information, the probability of multi-hop path similarity is calculated. Probabilistic reasoning and uncertainty quantification are performed. Combined with confidence thresholds, potential semantic associations are identified, and the enhanced semantic index graph is completed and corrected. S5. Receive the user's natural language query request and convert it into intent graph nodes. Perform subgraph matching retrieval in the enhanced semantic index graph to retrieve a set of candidate document nodes. S6. Reorder the candidate document nodes in the candidate document node set according to their relevance, calculate the decision support value of the candidate document nodes, and output the final list of related documents from high to low based on the decision support value.

2. The method for intelligent association and retrieval of engineering documents based on knowledge graphs according to claim 1, characterized in that, S1 specifically includes: S11. Collect multimodal heterogeneous document data throughout the entire lifecycle of the project, including text data contained in the design specification, vector graphics data contained in the CAD drawings, and unstructured text contained in the meeting minutes. Perform noise removal processing on the text data, vector graphics data, and unstructured text through data cleaning, and delete duplicate items and garbled characters in the data. S12. Call the pre-built ontology tree, use the BERT model to perform semantic matching to map the heterogeneous terms in the denoised multimodal heterogeneous document data to the standard terms in the ontology tree, replace the heterogeneous terms with the corresponding standard terms in the ontology tree, perform semantic alignment operation on the heterogeneous terms, and store the semantically aligned text data, vector graphics data and unstructured text according to the preset format to generate a standardized document dataset.

3. The method for intelligent association and retrieval of engineering documents based on knowledge graphs according to claim 1, characterized in that, S2 specifically includes: S21. Input the standardized document dataset into the LayoutLM model, convert the contained text data and unstructured text into word vectors through the embedding layer, convert the vector graphics data into image patch feature vectors, and concatenate the word vectors and image patch feature vectors to construct a multimodal input vector sequence. S22. Input the multimodal input vector sequence into the Transformer encoder of the LayoutLM model, and use the self-attention layer to calculate the dot product between each word vector and all image patch feature vectors to obtain the interaction matrix; S23. Normalize each row of the interaction matrix to obtain the weight coefficients, sum the image patch feature vectors according to the weight coefficients, and add them to the word vectors to update the word vectors. S24. Input the updated word vectors into the preset linear layer, calculate the classification score vector of each word vector belonging to the preset entity category, perform an exponential operation on each value in the classification score vector and divide it by the sum of all exponential values ​​to obtain the probability distribution, and select the category with the highest probability as the corresponding entity label. S25. Read the vector node coordinate file in the vector graphics data, parse all horizontal and vertical coordinate values, and encapsulate them into graphic geometric features. Read the timestamp field in the document attribute table, extract the year, month, day, hour, minute, and second values ​​as time data, and combine them into document metadata. S26. Analyze the positions of all word vectors labeled as entity tags in the text, use a bidirectional long short-term memory network to calculate the semantic correlation between every two entity word vectors, and generate a set of directed edges for the dependency relationships between entities. S27. Perform key-value matching between the horizontal and vertical coordinate data extracted from the vector graphic data and the spatial distance entity label. Write the successfully matched coordinate data into the attribute slot of the corresponding entity and attach the graphic geometric features as attribute nodes to the corresponding entity. S28. Perform key-value matching between the extracted year, month, day, hour, minute, and second values ​​and the time event entity tags. Write the successfully matched time data into the attribute slot of the corresponding entity and attach the document metadata as an attribute node to the corresponding entity. S29. Using entities as head nodes and attributes as tail nodes, mark ownership relationships as attribute edges and dependency relationships as entity edges, construct the original knowledge graph triple data in the order of head nodes, relationships, and tail nodes, and output it.

4. The method for intelligent association and retrieval of engineering documents based on knowledge graphs according to claim 1, characterized in that, S3 specifically includes: S31. Read all entity nodes in the original knowledge graph triple data, build an index based on the document number and chapter number to which each entity node belongs, map all entity nodes to the set of vertices in the hypergraph space, and establish a one-to-one correspondence between vertex IDs and entity nodes. S32. Traverse each vertex in the vertex set, calculate the cosine similarity between the current vertex and all other vertices in the documents and chapters, select vertices whose cosine similarity values ​​are greater than the preset similarity threshold and point to the same project event, and classify the selected vertices into the same hyperedge to generate a hyperedge set. S33. The word vectors that will be labeled as entity tags are used as vertex feature vectors. Using the hyperedge aggregation mechanism of the dynamic hypergraph neural network, the vertex feature vectors contained in each hyperedge are read. The mean pooling operation is performed on all vertex feature vectors in the same hyperedge, and the mean pooling result is used as the overall feature representation of the corresponding hyperedge. S34. Scan the related documents of the hyperedge containing sentences with preset causal association words, extract the word vectors of the corresponding sentences and take the average value as the causal logic label, read the timestamp of the related documents of each hyperedge, convert the timestamp into a time value vector and concatenate it with the numerical code of the causal logic label in dimension, and use the preset linear transformation layer to map the concatenated vector into a scalar weight value. S35. Multiply the calculated scalar weight values ​​element-wise with the overall feature representation of the corresponding hyperedge to obtain the weighted hyperedge features. Use the weighted hyperedge features as the temporal causal link connecting the vertices to construct a dynamic evolutionary knowledge graph.

5. The method for intelligent association and retrieval of engineering documents based on knowledge graphs according to claim 1, characterized in that, S4 specifically includes: S41. Construct an improved GBNN model. Use the weighted hyperedge features of the vertices represented in the dynamic evolution knowledge graph as feature vectors and arrange them in order of node ID to form a feature matrix. Input the feature matrix into the input layer and perform a linear transformation on the feature matrix through a preset weight matrix to obtain the initial node embedding matrix. S42. Input the initial node embedding matrix into the graph convolutional layer, read the adjacency matrix of the dynamic evolution knowledge graph, calculate the product of the initial node embedding matrix and the adjacency matrix, aggregate the neighbor features of each node, and use the ReLU activation function to perform nonlinear mapping on the aggregation result to obtain the neighbor aggregated feature vector. S43. Input the neighbor aggregation feature vector into the gated recurrent layer, calculate the difference vector between the neighbor aggregation feature vector and the hidden state of the node at the current time, input the difference vector into the Sigmoid function to calculate the update gate value, and at the same time calculate the weighted sum vector between the neighbor aggregation feature vector and the hidden state of the node at the current time, input the weighted sum vector into the Tanh function to calculate the candidate node state vector. S44. Calculate the memory retention amount by performing element-wise multiplication of the update gate value with the hidden state of the node at the previous time step, calculate the memory update amount by performing element-wise multiplication of the difference between 1 and the update gate value with the candidate node state vector, perform vector addition operation on the memory retention amount and the memory update amount, update the hidden state of the node at the current time step, and obtain the node embedding matrix. S45. Read all the temporal causal links in the dynamic evolution knowledge graph, arrange the node IDs contained in each temporal causal link in chronological order according to the timestamp, obtain the causal path sequence, input it into the preset bidirectional LSTM network for encoding, extract the concatenation vector of the forward output and the reverse output, and generate the global temporal causal graph structure vector. S46. Introduce a spatiotemporal causal mutual information maximization mechanism, calculate the dot product value of each local node representation in the node embedding matrix and the global temporal causal graph structure vector, use the dot product value as the spatiotemporal mutual information score, construct a loss function with negative spatiotemporal mutual information score as the objective, and use the stochastic gradient descent algorithm to update the parameters of the graph convolutional layer and the gated recurrent layer in reverse, so that the local node representation approximates the global temporal causal graph structure vector in the feature space. S47. Embed the updated node into the matrix input-output layer, calculate the dot product similarity between any two non-connected nodes in the dynamic evolution knowledge graph, perform Softmax normalization on the dot product similarity, and obtain the multi-hop path similarity probability between nodes, i.e. the prediction probability. S48. Use the Monte Carlo sampling method to randomly select 10% of the edges from the dynamic evolution knowledge graph to form a negative sample set. Calculate the multi-hop path similarity probability of each edge in the negative sample set. Use the binary cross-entropy loss function to calculate the error value between the predicted probability and the true label, and use it as a quantitative indicator of the uncertainty of the probability prediction. S49. Traverse all non-connected node pairs, read the multi-hop path similarity probability and uncertainty quantification index, filter out target node pairs with multi-hop path similarity probability greater than the preset confidence threshold and uncertainty quantification index less than the preset uncertainty threshold, identify the target node pairs as potential semantic associations, establish new connection edges between the two target nodes, set the weight of the new connection edge to the corresponding multi-hop path similarity probability, add the new connection edge to the dynamic evolution knowledge graph, and generate an enhanced semantic index graph.

6. The method for intelligent association and retrieval of engineering documents based on knowledge graphs according to claim 5, characterized in that, The calculation of the error value between the predicted probability and the true label using the binary cross-entropy loss function specifically includes: Read the corresponding real label values ​​in the dynamic evolution knowledge graph, substitute the real label values ​​into the real variable terms in the binary cross-entropy loss function formula, where a real label value of 1 represents a connection between nodes and a real label value of 0 represents no connection between nodes. Logarithmic operation is performed on the predicted probability to calculate the natural logarithm of the predicted probability. The true label value is multiplied by the natural logarithm to obtain the first product term. At the same time, the difference between the value 1 and the true label value is calculated. Negative logarithmic operation is performed on the predicted probability to calculate the natural logarithm of the difference between the value 1 and the predicted probability. The difference is multiplied by the natural logarithm of the difference to obtain the second product term. Add the first product term to the second product term, calculate the negative of the sum of the two product terms, and define the negative of the sum as the prediction error value for a single sample.

7. The method for intelligent association and retrieval of engineering documents based on knowledge graphs according to claim 1, characterized in that, S5 specifically includes: S51. Receive the user's natural language query request, input the natural language query request into the preset BERT model for semantic encoding, extract the word vector features of the natural language query request, input the word vector features into the preset fully connected layer to calculate the classification score, select the category with the highest classification score as the query intent category, and create the corresponding intent graph node according to the query intent category. S52. Perform precise matching of the keyword entities in the natural language query request with the entity nodes in the enhanced semantic index graph using entity name tags. Take the entity nodes that successfully match the keyword entities as anchor nodes, expand the first-level neighbor nodes around the anchor nodes, extract the subgraph structure formed by the anchor nodes and the first-level neighbor nodes as candidate subgraphs, traverse all nodes in the candidate subgraphs, filter out the nodes whose node type belongs to the document type, and mark them as candidate document nodes. S53. Read all connection paths between anchor nodes and each candidate document node in the candidate subgraph, traverse each connection path, perform cumulative calculation on the weight values ​​of all connection edges contained in the connection path to obtain the total path weight of each connection path, filter out the target path with a total path weight less than the preset path threshold, and extract the candidate document node corresponding to the end of the target path as the initial screening node. S54. Extract word vector features from intent graph nodes and node embedding vectors from preliminary screening nodes. Calculate the Euclidean distance between word vector features and node embedding vectors in the multidimensional feature space. Define the Euclidean distance as the semantic distance between intent graph nodes and preliminary screening nodes. S55. Sort all the initial screening nodes in ascending order of semantic distance value, select the top-ranked initial screening nodes for summary, remove duplicate nodes from the list, and generate a candidate document node set.

8. The method for intelligent association and retrieval of engineering documents based on knowledge graphs according to claim 1, characterized in that, S6 specifically includes: S61. Read the document metadata of each candidate document node in the candidate document node set, extract the timestamp field containing the year, month, day, hour, minute and second values, and arrange all candidate document nodes linearly according to the order of the timestamp values ​​to construct a timeline sequence. S62. Traverse any two candidate document nodes in the candidate document node set, query the causal tracing path connecting the two candidate document nodes in the enhanced semantic index graph, read the pointing relationship of all connecting edges on the causal tracing path, and determine the upstream and downstream positions of the two candidate document nodes in the project event context. S63. Assign values ​​to candidate document nodes based on their upstream and downstream positions. If a candidate document node is located at the upstream starting point in the project event network, assign the preset basic decision score to the candidate document node. If a candidate document node is located at the downstream end, calculate the number of path hops from the upstream starting point node to the current node, divide the basic decision score by the number of path hops to obtain the attenuated score, and use it as the decision support value of the candidate document node. S64. Sort all candidate document nodes in descending order of decision support value, generate a final list of associated documents based on the sorting order, read the decision support value of each candidate document node in the final list of associated documents, and take the documents corresponding to the candidate document nodes whose values ​​exceed the preset value threshold as recommended documents, and output the final list of associated documents.

Images