Cantilever construction data management system based on knowledge graph
By using a knowledge graph-based data management system and leveraging the BERT model and cross-layer feature alignment technology, the problem of intelligent extraction of unstructured data in cantilever construction was solved, achieving high-precision extraction of entities and relationships, and improving the level of intelligence in construction management and project quality.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- GUIZHOU ROAD & BRIDGE GRP
- Filing Date
- 2026-03-04
- Publication Date
- 2026-06-19
Smart Images
Figure CN121764925B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of bridge construction data management technology, and in particular to a cantilever construction data management system based on knowledge graphs. Background Technology
[0002] In the current construction management of cantilever bridges, there is a common problem of data silos where structured monitoring data is separated from unstructured text data such as construction logs and inspection reports. This makes it difficult to quickly extract and associate key engineering entities and relationships from massive amounts of heterogeneous data, thus limiting the realization of intelligent risk warning and decision support capabilities based on data fusion.
[0003] Currently, Chinese invention patent application number CN202011554675.2 discloses a method and system for monitoring and early warning of unbalanced moments in prestressed concrete continuous beams. The method involves sensors collecting the pressure borne by temporary supports at the pier top and transmitting it to a controller. The controller performs analog-to-digital conversion on the input signal and calculates the unbalanced moments of the cantilever beams on both sides. When the unbalanced moment exceeds the limit, an alarm signal is issued via an alarm device. Simultaneously, the controller sends the data to a computer or mobile device for display and storage. This invention enables real-time monitoring and early warning of unbalanced moments during the cantilever construction of prestressed concrete continuous beam bridges, solving the construction problems of difficulty in quantifying, monitoring, and managing unbalanced moments during cantilever construction. However, the related technologies struggle to extract entities and relationships accurately from unstructured data from cantilever construction, lacking intelligence. Furthermore, the technologies cannot address the semantic shift problem of different levels of features in the BERT model and lack a cross-layer feature alignment and adaptive fusion mechanism based on deformable convolution and correlation weights. Summary of the Invention
[0004] The technical problem solved by this invention is that related technologies have difficulty extracting entities and relationships with high accuracy from unstructured data of cantilever construction, lack intelligence, cannot solve the semantic offset problem of features at different levels in the BERT model, and lack a cross-layer feature alignment and adaptive fusion mechanism based on deformable convolution and correlation weights.
[0005] To solve the above-mentioned technical problems, the present invention provides the following technical solution:
[0006] A knowledge graph-based cantilever construction data management system includes a data acquisition module, a knowledge extraction module, a construction management module, and an application module.
[0007] The data acquisition module is used to collect cantilever construction data;
[0008] The knowledge extraction module is used to extract entities and relationships from unstructured data of cantilever construction to form a structured ternary list.
[0009] The construction management module is used to integrate the structured data acquired by the data acquisition module and the structured ternary list to construct a knowledge graph of cantilever construction.
[0010] The application module is used to generate data queries, risk warnings, and decision reports based on the knowledge graph of cantilever construction.
[0011] As a preferred embodiment of the knowledge graph-based cantilever construction data management system described in this invention, the cantilever construction data includes structured cantilever construction data and unstructured cantilever construction data.
[0012] The structured data for cantilever construction includes sensor monitoring data, construction progress data, material property data, and equipment status data.
[0013] The unstructured data for cantilever construction includes construction logs, technical specifications, design drawings, test reports, and expert evaluation reports.
[0014] As a preferred embodiment of the knowledge graph-based cantilever construction data management system described in this invention, the knowledge extraction module includes a filtering unit and a parsing unit.
[0015] The filtering unit is used to identify and filter unstructured data of cantilever construction using regular expressions to obtain a subset of candidate sentences, which includes entity identifiers and relation identifiers.
[0016] The parsing unit is used to extract information from the subset of candidate sentences to obtain a structured ternary list.
[0017] As a preferred embodiment of the knowledge graph-based cantilever construction data management system described in this invention, the information extraction specifically includes:
[0018] The sentences in the candidate sentence subset are preprocessed. The preprocessing includes adding a first special symbol at the beginning of the sentence, adding a second special symbol at the end of the sentence, and adding a third special symbol before the target entity in the sentence. The preprocessed sentences are then subjected to word embedding, position embedding, and segment embedding in sequence to obtain a word vector sequence.
[0019] The word vector sequence is encoded using the BERT model to obtain the hidden state output encoding. For the hidden state output encoding, the global feature representation of each hidden state in the BERT model is extracted, and an L×L symmetric matrix is initialized, where L is the total number of layers in the BERT model. For each element of the matrix, the correlation weight between the hidden states at different levels is calculated using cosine similarity to construct a symmetric original cross-layer correlation matrix. Each row of the original cross-layer correlation matrix is subjected to softmax normalization to obtain the cross-layer correlation matrix. Based on the cross-layer correlation matrix, the entity feature vectors of each level are weighted and fused through feature alignment to obtain the enhanced entity feature vector.
[0020] By using the channel attention mechanism, the weights of the enhanced entity feature vector and the feature vector corresponding to the first special label are redistributed to obtain the comprehensive feature vector;
[0021] The comprehensive feature vector is input into the relation classification layer, and the predicted probability is calculated through a fully connected layer and a softmax function. All relation categories whose predicted probabilities exceed a first threshold are filtered out. Based on the entities that meet the first threshold and their corresponding relation categories, a structured ternary list is constructed. The column format of the structured ternary list is head entity-relation type-tail entity, and the rows are records of each entity that meets the first threshold and its corresponding relation category.
[0022] As a preferred embodiment of the knowledge graph-based cantilever construction data management system described in this invention, the correlation weights between hierarchical hidden states are calculated using cosine similarity, and the calculation expression is as follows:
[0023] ;
[0024] in, This represents the relevance weight between the i-th entity in the target layer and the a-th entity in the reference layer. This represents the feature vector of the i-th entity in the target layer. This represents the feature vector of the a-th entity in the reference layer. This represents the dot product operation of two eigenvectors. This represents the square root of the sum of squares of the elements in the entity's feature vector. Indicates the reference layer entity index. This represents the target layer entity index.
[0025] As a preferred embodiment of the knowledge graph-based cantilever construction data management system described in this invention, the weighted fusion of entity feature vectors at each level through feature alignment includes:
[0026] Extract the target layer entity feature vector and reference layer entity feature vector corresponding to each target entity from the hidden state, map them to the same feature space through DCN, and calculate the feature alignment offset. The expression for the feature alignment offset is:
[0027] ;
[0028] in, This represents the feature offset between the target layer l and the reference layer j. For the target layer entity feature vector, For the reference layer entity feature vector, For feature splicing operations, The function is composed of two convolutional layers. Bilinear interpolation is used to calculate feature sampling at the target entity location to achieve feature alignment. Then, the inter-layer correlation weights in the cross-layer correlation matrix are used to perform a weighted summation of the entity feature vectors at each layer. The expression for this weighted summation is as follows:
[0029] ;
[0030] in, To enhance entity feature vectors, The normalized inter-layer correlation weights The feature vector of reference layer j aligned to and mapped to target layer l. The total number of entities in the reference layer. For reference layer entity index.
[0031] As a preferred embodiment of the knowledge graph-based cantilever construction data management system described in this invention, the construction management module includes a data fusion unit, a graph construction unit, a storage and update unit, and a query service unit.
[0032] The data fusion unit is used to receive a structured ternary list and structured data, traverse the structured data and the structured ternary list, find the corresponding paired data records in the structured ternary list according to each structured data, eliminate duplicates and conflicts, and identify and merge different text mentions and semantically identical but different construction relationship types that refer to the same construction entity through similarity calculation based on embedding vectors, and obtain knowledge graph data. The knowledge graph data includes construction entity type, construction relationship type and construction attribute type corresponding to construction entity type.
[0033] The knowledge graph data is standardized according to the ontology of cantilever construction to obtain standard knowledge graph data, and the knowledge graph data is converted into a standard data format in the field.
[0034] The graph construction unit is used to instantiate the standard knowledge graph data and construct a knowledge graph for cantilever construction.
[0035] The storage update unit is used to map the structured ternary list to nodes and edges in the graph database, where entities are mapped to nodes and relations are mapped to edges, and corresponding construction attribute types are added to the nodes and edges for storage. The knowledge extraction module responds to the newly added unstructured data to obtain the newly added structured ternary list.
[0036] The query service unit is used to provide a map query interface for the application module.
[0037] As a preferred embodiment of the knowledge graph-based cantilever construction data management system described in this invention, the instantiation specifically includes:
[0038] Based on the construction entity types in the knowledge graph data, an empty queue of entity instances is created under the corresponding entity category, and a unique identifier is assigned to each empty queue of entity instances. The construction attribute types in the knowledge graph data are associated with the empty queues of entity instances to obtain an entity instance queue. Based on the construction relationship types in the knowledge graph data, relationship instances are established, and the entity instance queues are connected with the corresponding attribute types to obtain the cantilever construction knowledge graph.
[0039] As a preferred embodiment of the knowledge graph-based cantilever construction data management system described in this invention, the application module includes an interactive query unit, a risk warning unit, and a decision reporting unit:
[0040] The interactive query unit is used to retrieve information from the cantilever construction knowledge graph and provide a traceability path display;
[0041] The risk warning unit is used to monitor changes in entities in the cantilever construction knowledge graph in real time, and to perform risk reasoning and warning based on construction relationship types.
[0042] The decision reporting unit is used to generate analysis reports based on the relationship types in the cantilever construction knowledge graph and to provide decision-making references for comparing construction schemes.
[0043] As a preferred embodiment of the knowledge graph-based cantilever construction data management system described in this invention, the entities and relationships in the cantilever construction knowledge graph are visualized and rendered using an interactive topology graph. Different construction entity types are distinguished by nodes of different colors and shapes, and different construction relationship types are represented by lines. The system also supports drag-and-drop positioning, scaling, and real-time display of key attribute panels for nodes.
[0044] The beneficial effects of this invention are as follows: By constructing a knowledge graph for cantilever construction and employing information extraction techniques based on the BERT model combined with cross-layer feature alignment and fusion, it can efficiently and automatically identify entities and relationships from unstructured text such as construction logs and inspection reports, forming structured triplet knowledge and achieving deep value mining of text data. Through deep fusion of multi-source heterogeneous data, a unified engineering knowledge system is established, enabling sensor monitoring data and construction process records to corroborate each other, providing data support for construction scheme comparison, improving the intelligent management level of cantilever construction, and having significant value in ensuring construction safety and improving project quality and efficiency. Attached Figure Description
[0045] Figure 1 This is a basic flowchart of a knowledge graph-based cantilever construction data management system provided in one embodiment of the present invention. Detailed Implementation
[0046] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.
[0047] Example, refer to Figure 1 As one embodiment of the present invention, a knowledge graph-based cantilever construction data management system is provided, including a data acquisition module, a knowledge extraction module, a construction management module, and an application module:
[0048] The data acquisition module is used to collect data from cantilever construction.
[0049] The knowledge extraction module is used to extract entities and relationships from unstructured data of cantilever construction to form a structured ternary list;
[0050] The construction management module is used to integrate the structured data and structured ternary lists obtained by the data acquisition module to construct a knowledge graph of cantilever construction.
[0051] The application module is used to generate data queries, risk warnings, and decision reports based on the knowledge graph of cantilever construction.
[0052] This invention employs an information extraction technique based on the BERT model, combined with cross-layer feature alignment and fusion, to efficiently and automatically identify entities and relationships from unstructured text such as construction logs and inspection reports, forming structured triple knowledge and achieving deep value mining of text data. By deeply fusing multi-source heterogeneous data, a unified knowledge graph for cantilever construction is established, enabling sensor monitoring data and construction process records to corroborate each other, providing data support for construction scheme comparison, improving the intelligent management level of cantilever construction, and having significant value in ensuring construction safety and improving project quality and efficiency.
[0053] Cantilever construction data includes structured cantilever construction data and unstructured cantilever construction data;
[0054] Structured data for cantilever construction includes sensor monitoring data, construction progress data, material property data, and equipment status data;
[0055] Unstructured data for cantilever construction includes construction logs, technical specifications, design drawings, test reports, and expert evaluation reports.
[0056] In a specific embodiment, the cantilever construction data is clearly divided into structured and unstructured data, and its definition and scope are detailed.
[0057] The knowledge extraction module includes a filtering unit and a parsing unit;
[0058] The filtering unit is used to identify and filter unstructured data of cantilever construction using regular expressions to obtain a subset of candidate sentences, which includes entity identifiers and relation identifiers.
[0059] The parsing unit is used to extract information from a subset of candidate sentences to obtain a structured ternary list.
[0060] Information extraction specifically includes:
[0061] The sentences in the candidate sentence subset are preprocessed. The preprocessing includes adding a first special symbol at the beginning of the sentence, a second special symbol at the end of the sentence, and a third special symbol before the target entity in the sentence. The preprocessed sentences are then subjected to word embedding, position embedding, and segment embedding in sequence to obtain a word vector sequence.
[0062] The word vector sequence is encoded using the BERT model to obtain the hidden state output encoding. For the hidden state output encoding, the global feature representation of each hidden state of the BERT model is extracted, and an L×L symmetric matrix is initialized, where L is the total number of layers in the BERT model. For each element of the matrix, the correlation weight between the hidden states of different layers is calculated using cosine similarity to construct a symmetric original cross-layer correlation matrix. Each row of the original cross-layer correlation matrix is subjected to softmax normalization to obtain the cross-layer correlation matrix. Based on the cross-layer correlation matrix, the entity feature vectors of each layer are weighted and fused through feature alignment to obtain the enhanced entity feature vector.
[0063] By using the channel attention mechanism, the weights of the enhanced entity feature vector and the feature vector corresponding to the first special label are redistributed to obtain the comprehensive feature vector;
[0064] The comprehensive feature vector is input into the relation classification layer, and the predicted probability is calculated through a fully connected layer and a softmax function. All relation categories whose predicted probabilities exceed the first threshold are filtered out. Based on the entities that meet the first threshold and their corresponding relation categories, a structured triple list is constructed. The column format of the structured triple list is head entity-relation type-tail entity, and the rows are records of each entity that meets the first threshold and its corresponding relation category.
[0065] In a specific embodiment, the first threshold is a fixed quality control parameter that serves as a confidence threshold for determining the probability of relation classification. This threshold effectively balances recall and precision by filtering relation categories whose predicted probabilities exceed the set value: predictions above the threshold are adopted as valid relations, significantly improving the reliability of the extraction results; while predictions below the threshold are considered uncertain relations and discarded.
[0066] The BERT model encodes word vector sequences and outputs the hidden states of each layer. To fully utilize the feature information from different layers, the system calculates the cosine similarity between all hidden layers, constructs a cross-layer correlation matrix, and obtains inter-layer weights through Softmax normalization. Based on these weights, feature alignment and weighted fusion are performed on the entity feature vectors distributed across each layer to generate enhanced entity feature vectors that more accurately represent entity semantics. This process effectively solves the problem of feature semantic mismatch between different layers.
[0067] Furthermore, through a channel attention mechanism, the importance of the enhanced entity feature vector and the sentence global semantic vector carried by the first special symbol is redistributed and fused into a comprehensive feature vector. Finally, this vector is input into the relation classification layer, and the probability of it belonging to each relation class is calculated through a fully connected network and the Softmax function. For example, if the probability of construction activity exceeds the first threshold, the system determines that there is a relation between the head entity segment 7 and the tail entity tensioning, and generates a standard structured triple list record as segment 7, construction activity, tensioning. Finally, a list composed of many such triples is output, completing the transformation from unstructured text to structured knowledge.
[0068] The correlation weights between hidden states at different levels are calculated using cosine similarity, and the calculation expression is as follows:
[0069] ;
[0070] in, This represents the relevance weight between the i-th entity in the target layer and the a-th entity in the reference layer. This represents the feature vector of the i-th entity in the target layer. This represents the feature vector of the a-th entity in the reference layer. This represents the dot product operation of two eigenvectors. This represents the square root of the sum of squares of the elements in the entity's feature vector. Indicates the reference layer entity index. This represents the target layer entity index.
[0071] In a specific embodiment, the semantic consistency of different network layers in representing the same entity is quantitatively evaluated, providing a scientific weighting basis for subsequent feature fusion.
[0072] Weighted fusion of entity feature vectors at various levels through feature alignment includes:
[0073] Extract the target layer entity feature vector and reference layer entity feature vector corresponding to each target entity from the hidden state, map them to the same feature space through DCN, and calculate the feature alignment offset. The expression for the feature alignment offset is:
[0074] ;
[0075] in, This represents the feature offset between the target layer and the reference layer. For the target layer entity feature vector, For the reference layer entity feature vector, For feature splicing operations, The function consists of two convolutional layers. Bilinear interpolation is used to calculate feature samples of the target entity location to achieve feature alignment. Then, the inter-layer correlation weights in the cross-layer correlation matrix are used to perform a weighted summation of the entity feature vectors at each layer. The expression for the weighted summation is as follows:
[0076] ;
[0077] in, To enhance entity feature vectors, The normalized inter-layer correlation weights The feature vector is aligned from the reference layer to the target layer and then mapped. The total number of entities in the reference layer. For reference layer entity index.
[0078] In a specific embodiment, feature alignment is used to weightedly fuse entity feature vectors at various levels to address the semantic gap and spatial mismatch issues existing in different levels of features in the BERT model, thereby achieving accurate feature fusion. This is achieved by calculating the feature alignment offset between the target layer and the reference layer entity feature vectors through a deformable convolutional network. This offset precisely quantifies the positional differences of features at different levels in the semantic space. Spatial transformation and resampling are performed on the reference layer features to achieve precise alignment with the target layer features in the semantic space, effectively eliminating the inconsistency in feature representation caused by network depth. Based on the inter-layer correlation weights provided by the cross-layer correlation matrix, all these precisely aligned layer features are adaptively weighted and fused to generate an enhanced entity feature vector with more complete information and clearer semantics. This process ensures that the feature contribution from different network depths is proportional to its semantic relevance, thus significantly improving the accuracy and robustness of subsequent relation classification.
[0079] The construction and management module includes a data fusion unit, a graph construction unit, a storage and update unit, and a query service unit;
[0080] The data fusion unit receives structured ternary lists and structured data, traverses the structured data and structured ternary lists, finds the corresponding paired data records in the structured ternary lists based on each structured data, eliminates duplicates and conflicts, and identifies and merges different textual references and semantically identical but differently expressed construction relationship types that refer to the same construction entity through similarity calculation based on embedding vectors, and obtains knowledge graph data. The knowledge graph data includes construction entity types, construction relationship types, and construction attribute types corresponding to construction entity types.
[0081] The knowledge graph data is standardized according to the ontology of cantilever construction to obtain standard knowledge graph data, and the knowledge graph data is converted into the standard data format in the field.
[0082] The graph construction unit is used to instantiate standard knowledge graph data and construct a knowledge graph for cantilever construction.
[0083] The storage update unit is used to map the structured ternary list to nodes and edges in the graph database, where entities are mapped to nodes and relations are mapped to edges. It also adds corresponding construction attribute types to the nodes and edges for storage. The knowledge extraction module responds to the newly added unstructured data to obtain the newly added structured ternary list.
[0084] The query service unit is used to provide a map query interface for application modules.
[0085] In a specific embodiment, the knowledge graph data is standardized according to the ontology of cantilever construction to obtain standard knowledge graph data, and the knowledge graph data is converted into a standard data format within the domain.
[0086] The construction and management module achieves complete lifecycle management from multi-source data to a dynamic knowledge graph through four core units: The data fusion unit is responsible for deeply integrating the parsed structured ternary lists with the original structured data, identifying and merging different textual expressions referring to the same entity through similarity calculation based on embedded vectors, such as identifying "block 7" and "block 7#" as the same entity and semantically equivalent relationship types, while eliminating data conflicts, and then standardizing the data through the cantilever construction domain ontology to form unified and standardized knowledge graph data; The graph construction unit instantiates the standardized data into specific knowledge nodes and relation edges, formally constructing the cantilever construction knowledge graph; The storage and update unit is responsible for persistently storing the graph data to the graph database, where entities are mapped to nodes and relations are mapped to edges, and continuously responding to new data to realize the dynamic expansion and update of the graph; The query service unit encapsulates the graph query interface, providing efficient and convenient knowledge retrieval and service support for upper-layer application modules.
[0087] Instantiation specifically includes:
[0088] Based on the construction entity types in the knowledge graph data, an empty queue of entity instances is created under the corresponding entity category, and a unique identifier is assigned to each empty queue of entity instances. The construction attribute types in the knowledge graph data are associated with the empty queues of entity instances to obtain an entity instance queue. Based on the construction relationship types in the knowledge graph data, relationship instances are established, and the entity instance queues are connected with the corresponding attribute types to obtain the cantilever construction knowledge graph.
[0089] In a specific embodiment, the instantiation process is a key step in transforming standardized knowledge graph data into an operable knowledge network. Its role is to concretize abstract entity types and relation types into knowledge nodes and associated edges that contain actual engineering data.
[0090] The application modules include an interactive query unit, a risk warning unit, and a decision reporting unit.
[0091] The interactive query unit is used to retrieve information from the cantilever construction knowledge graph and provide a traceability path display;
[0092] The risk warning unit is used to monitor changes in entities in the cantilever construction knowledge graph in real time and to perform risk reasoning and warning based on construction relationship types.
[0093] The decision report unit is used to generate analysis reports based on the relationship types in the cantilever construction knowledge graph and to provide decision-making references for the comparison of construction schemes.
[0094] In specific embodiments, the interactive query unit provides intuitive knowledge retrieval and traceability display, enabling managers to quickly locate all related information such as "block 7", including material properties, construction progress, and related monitoring data, and present the complete data source chain; the risk warning unit performs real-time risk reasoning based on the cantilever construction knowledge graph relationship network. When a change in the entity state of "stress exceeding limit" is detected, it can automatically associate it with the corresponding construction procedures, material batches, and equipment status through relationships such as "located" and "used", realizing multi-dimensional risk assessment and early warning; the decision report unit uses the rich entity relationships in the graph to automatically generate a comprehensive report containing related data comparison and risk trend analysis, providing multi-source decision support covering historical cases, current status, and specification requirements for construction scheme selection.
[0095] The entities and relationships in the knowledge graph of cantilever construction are visualized and rendered using an interactive topology graph. Different types of construction entities are distinguished by nodes of different colors and shapes, and different types of construction relationships are represented by lines. The graph also supports drag-and-drop positioning, scaling, and real-time display of key attribute panels for nodes.
[0096] In a specific embodiment, the entities and relationships in the cantilever construction knowledge graph are visualized using an interactive topology graph. Intuitive identification of engineering elements is achieved through differentiated visual encoding: blue rectangular nodes represent beam segments, red circular nodes represent monitoring points, green triangular nodes represent construction equipment, and yellow diamond nodes represent material batches. Relationship types are distinguished by different line types: solid lines represent spatial relationships such as "located," dashed lines represent technological relationships such as "used," and dotted lines represent data association relationships such as "monitoring." Users can rearrange the graph structure by dragging nodes, focus on specific construction areas using the zoom function, and display a key attribute panel in real time when any node is selected. This visualization design significantly lowers the barrier to understanding the knowledge graph.
[0097] This invention achieves deep integration and intelligent application of multi-source heterogeneous engineering data by constructing a knowledge graph for cantilever construction. It adopts knowledge extraction technology based on the BERT model and cross-layer feature alignment to automatically extract entity relationships from unstructured text. Through data fusion and dynamic update mechanisms, a unified engineering knowledge system is established. Finally, based on a visual interactive interface, it provides construction managers with accurate data query, real-time risk warning and intelligent decision support, significantly improving the digital management level of cantilever construction.
[0098] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product implemented on one or more computer-usable storage media containing computer-usable program code. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.
[0099] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1. A knowledge graph-based cantilever construction data management system, characterized in that, It includes a data acquisition module, a knowledge extraction module, a build management module, and an application module: The data acquisition module is used to collect cantilever construction data; The knowledge extraction module is used to extract entities and relationships from unstructured data of cantilever construction to form a structured ternary list, including: The word vector sequence is encoded using the BERT model to obtain the hidden state output encoding. For the hidden state output encoding, the global feature representation of each hidden state of the BERT model is extracted, and an L×L symmetric matrix is initialized, where L is the total number of layers in the BERT model. For each element of the matrix, the correlation weight between the hidden states of different layers is calculated using cosine similarity to construct a symmetric original cross-layer correlation matrix. Each row of the original cross-layer correlation matrix is subjected to softmax normalization to obtain the cross-layer correlation matrix. Based on the cross-layer correlation matrix, the entity feature vectors of each layer are weighted and fused through feature alignment to obtain the enhanced entity feature vector. By using the channel attention mechanism, the weights of the enhanced entity feature vector and the feature vector corresponding to the first special label are redistributed to obtain the comprehensive feature vector; Weighted fusion of entity feature vectors at various levels through feature alignment includes: Extract the target layer entity feature vector and reference layer entity feature vector corresponding to each target entity from the hidden state, map them to the same feature space through DCN, and calculate the feature alignment offset. The expression for the feature alignment offset is: ； in, This represents the feature offset between the target layer l and the reference layer j. For the target layer entity feature vector, For the reference layer entity feature vector, For feature splicing operations, The function is composed of two convolutional layers. Bilinear interpolation is used to calculate feature sampling at the target entity location to achieve feature alignment. Then, the inter-layer correlation weights in the cross-layer correlation matrix are used to perform a weighted summation of the entity feature vectors at each layer. The expression for this weighted summation is as follows: ； in, To enhance entity feature vectors, The normalized inter-layer correlation weights The feature vector of reference layer j aligned to and mapped to target layer l. The total number of entities in the reference layer. For reference layer entity index; The construction management module is used to integrate the structured data acquired by the data acquisition module and the structured ternary list to construct a knowledge graph of cantilever construction. The application module is used to generate data queries, risk warnings, and decision reports based on the knowledge graph of cantilever construction.
2. The knowledge graph-based cantilever construction data management system of claim 1, wherein, The cantilever construction data includes structured cantilever construction data and unstructured cantilever construction data; The structured data for cantilever construction includes sensor monitoring data, construction progress data, material property data, and equipment status data. The unstructured data for cantilever construction includes construction logs, technical specifications, design drawings, test reports, and expert evaluation reports.
3. The knowledge graph-based cantilever construction data management system of claim 2, wherein, The knowledge extraction module includes a filtering unit and a parsing unit; The filtering unit is used to identify and filter unstructured data of cantilever construction using regular expressions to obtain a subset of candidate sentences, which includes entity identifiers and relation identifiers. The parsing unit is used to extract information from the subset of candidate sentences to obtain a structured ternary list.
4. The knowledge graph-based cantilever construction data management system of claim 3, wherein, The information extraction specifically includes: The sentences in the candidate sentence subset are preprocessed. The preprocessing includes adding a first special symbol at the beginning of the sentence, adding a second special symbol at the end of the sentence, and adding a third special symbol before the target entity in the sentence. The preprocessed sentences are then subjected to word embedding, position embedding, and segment embedding in sequence to obtain a word vector sequence. The comprehensive feature vector is input into the relation classification layer, and the predicted probability is calculated through a fully connected layer and a softmax function. All relation categories whose predicted probabilities exceed a first threshold are filtered out. Based on the entities that meet the first threshold and their corresponding relation categories, a structured ternary list is constructed. The column format of the structured ternary list is head entity-relation type-tail entity, and the rows are records of each entity that meets the first threshold and its corresponding relation category.
5. A knowledge graph-based cantilever construction data management system as described in claim 4, characterized in that, The correlation weights between hidden states at different levels are calculated using cosine similarity, and the calculation expression is as follows: ； in, This represents the relevance weight between the i-th entity in the target layer and the a-th entity in the reference layer. This represents the feature vector of the i-th entity in the target layer. This represents the feature vector of the a-th entity in the reference layer. This represents the dot product operation of two eigenvectors. This represents the square root of the sum of squares of the elements in the entity's feature vector. Indicates the reference layer entity index. This represents the target layer entity index.
6. The knowledge graph-based cantilever construction data management system of claim 5, wherein, The construction management module includes a data fusion unit, a graph construction unit, a storage and update unit, and a query service unit; The data fusion unit is used to receive a structured ternary list and structured data, traverse the structured data and the structured ternary list, find the corresponding paired data records in the structured ternary list according to each structured data, eliminate duplicates and conflicts, and identify and merge different text mentions and semantically identical but different construction relationship types that refer to the same construction entity through similarity calculation based on embedding vectors, and obtain knowledge graph data. The knowledge graph data includes construction entity type, construction relationship type and construction attribute type corresponding to construction entity type. The knowledge graph data is standardized according to the ontology of cantilever construction to obtain standard knowledge graph data, and the knowledge graph data is converted into a standard data format in the field. The graph construction unit is used to instantiate the standard knowledge graph data and construct a knowledge graph for cantilever construction. The storage update unit is used to map the structured ternary list to nodes and edges in the graph database, where entities are mapped to nodes and relations are mapped to edges, and corresponding construction attribute types are added to the nodes and edges for storage. The knowledge extraction module responds to the newly added unstructured data to obtain the newly added structured ternary list. The query service unit is used to provide a map query interface for the application module.
7. A knowledge graph-based cantilever construction data management system as described in claim 6, characterized in that, The instantiation specifically includes: Based on the construction entity types in the knowledge graph data, an empty queue of entity instances is created under the corresponding entity category, and a unique identifier is assigned to each empty queue of entity instances. The construction attribute types in the knowledge graph data are associated with the empty queues of entity instances to obtain an entity instance queue. Based on the construction relationship types in the knowledge graph data, relationship instances are established, and the entity instance queues are connected with the corresponding attribute types to obtain the cantilever construction knowledge graph.
8. The knowledge graph-based cantilever construction data management system of claim 7, wherein, The application module includes an interactive query unit, a risk warning unit, and a decision reporting unit: The interactive query unit is used to retrieve information from the cantilever construction knowledge graph and provide a traceability path display; The risk warning unit is used to monitor changes in entities in the cantilever construction knowledge graph in real time, and to perform risk reasoning and warning based on construction relationship types. The decision reporting unit is used to generate analysis reports based on the relationship types in the cantilever construction knowledge graph and to provide decision-making references for comparing construction schemes.
9. The knowledge graph-based cantilever construction data management system of claim 8, wherein, The entities and relationships in the cantilever construction knowledge graph are visualized and rendered using an interactive topology graph. Different types of construction entities are distinguished by nodes of different colors and shapes, and different types of construction relationships are represented by lines. The graph also supports drag-and-drop positioning, scaling, and real-time display of key attribute panels for nodes.