Domain Knowledge Graph Construction Method for Intelligent Decision-Making in Asphalt Pavement Maintenance
By constructing a maintenance case library, hierarchical ontology model, and knowledge graph based on multi-source heterogeneous information, and combining it with a deep reinforcement learning model, the problems of decision accuracy and interpretability in asphalt pavement maintenance have been solved, realizing intelligent, dynamic, and refined maintenance decision-making.
Patent Information
- Application Number
- CN202611393996.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-09
- Publication Date
- 2026-11-17
AI Technical Summary
Existing technologies lack effective integration of multi-source heterogeneous information and utilization of domain knowledge in asphalt pavement maintenance, resulting in insufficient accuracy and interpretability of maintenance decisions.
A maintenance case library with multi-source heterogeneous information for asphalt pavement maintenance is constructed. Knowledge extraction and entity alignment are performed by combining a hierarchical ontology model. The data is stored in a knowledge graph and dynamically managed through an incremental update mechanism. A knowledge-constrained deep reinforcement learning model is used for maintenance decision-making.
It improves the accuracy and interpretability of maintenance decisions, enhances the model's generalization ability and robustness under different road conditions, and provides scientific, efficient and reliable technical support.
Smart Images

Figure CN123117286A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of asphalt pavement maintenance and management, and in particular relates to a method for constructing a domain knowledge graph for intelligent decision-making in asphalt pavement maintenance. Background Technology
[0002] In recent years, with the continuous advancement of transportation infrastructure construction, the demand for highway maintenance has shown a rapid growth trend. Effective and timely maintenance intervention can delay the development of pavement defects, extend the service life of roads, and improve their service quality, which is of great significance for promoting the sustainable development of transportation infrastructure. Currently, asphalt pavement maintenance management is gradually transforming from a traditional experience-driven model to a data-driven model to meet the growing demand for intelligent, systematic, and refined maintenance.
[0003] At the technological level, a new generation of information technologies, represented by artificial intelligence, are being gradually applied to the field of road maintenance. Purely data-driven methods, especially deep learning technology, have achieved remarkable results in tasks such as pavement distress identification, service condition assessment, long-term performance prediction, and maintenance decision support, demonstrating superior performance compared to traditional empirical models, thanks to their powerful feature extraction and nonlinear fitting capabilities.
[0004] However, the vast amount of data and professional experience generated throughout the entire road lifecycle contains a wealth of knowledge. Existing methods mainly rely on data-driven technologies, lacking effective integration and utilization of domain knowledge, resulting in insufficient accuracy and interpretability of asphalt pavement maintenance decisions. How to organize scattered domain knowledge into a computable, associative, and scalable unified structure has become one of the core issues facing pavement maintenance management. Knowledge graphs, as a graph-structured knowledge representation form with rich semantic expression and strong associative capabilities, provide technical support for knowledge reasoning, traceable querying, and experience reuse. Therefore, constructing a knowledge graph for the field of asphalt pavement maintenance management has significant practical necessity and application value. Summary of the Invention
[0005] Purpose of the Invention: The purpose of this invention is to provide a domain knowledge graph construction method for intelligent decision-making in asphalt pavement maintenance, addressing the problems of difficulty in unifying the organization of multi-source heterogeneous information, insufficient utilization of domain knowledge, and insufficient dynamic adaptability of maintenance decisions. This invention constructs an asphalt pavement maintenance management knowledge graph that integrates multi-source heterogeneous information, and combines it with a knowledge-constrained deep reinforcement learning model to form a data- and knowledge-driven maintenance decision-making method, achieving intelligent, dynamic, and refined maintenance decisions.
[0006] Technical solution: The present invention provides a method for constructing a domain knowledge graph for intelligent decision-making in asphalt pavement maintenance, comprising the following steps:
[0007] Step 1: Build a maintenance case library to store structured and unstructured text data of maintenance cases from multiple sources;
[0008] Step 2: Construct a hierarchical ontology model for the field of asphalt pavement maintenance and management, define entity categories, attributes and semantic relationships, and realize knowledge modeling for maintenance scenarios;
[0009] Step 3: Based on the named entity recognition algorithm, perform knowledge extraction operation on the unstructured text data in the maintenance case library to extract entities corresponding to pavement defects, causes of defects and maintenance measures, and perform entity alignment and disambiguation processing.
[0010] Step 4: Based on the hierarchical ontology model, convert the structured data and standardized text knowledge in the maintenance case library into nodes, attributes, and relationships, and store them in the knowledge graph;
[0011] Step 5: Dynamically manage and continuously optimize the knowledge graph through an incremental update mechanism;
[0012] Step 6: Based on the optimized knowledge graph, implement a deep reinforcement learning model with knowledge constraints, perform maintenance decision reasoning, and output a comprehensive decision scheme covering maintenance measures, optimal maintenance timing, and maintenance costs.
[0013] Furthermore, in step 1, the heterogeneous information of the multi-source maintenance case includes basic road information, pavement structure information, pavement performance information, pavement distress information, traffic load information, climate environment information, and maintenance history information.
[0014] The construction of the maintenance case library specifically involves:
[0015] The acquired multi-source heterogeneous information is preprocessed, including format conversion, unit unification, and identification and deletion of missing and outlier values.
[0016] The processed heterogeneous information is associated and classified into cases according to road segments and inspection times, and a unique case number is assigned to each maintenance case.
[0017] Based on the case number, the pavement performance, defects, traffic, environment and maintenance history information corresponding to the same maintenance case are integrated to form a maintenance case and stored in the case library.
[0018] Furthermore, step 2 specifically involves:
[0019] A hierarchical ontology model for asphalt pavement maintenance and management is constructed, and the hierarchical ontology model is divided into a road segment layer and a case layer.
[0020] Construct the entity categories and data attributes corresponding to the road segment layer to describe the basic information, design information, and pavement performance, traffic load, climate environment, and maintenance history information of the road segment over time;
[0021] Construct the entity categories and data attributes corresponding to the case layer to describe the road surface condition, disease characteristics, disease causes, road surface structure and maintenance measures of the road segment at different detection times;
[0022] Define object attributes between different entity categories, and establish hierarchical relationships between road segments and cases based on the object attributes, as well as temporal relationships between different cases of the same road segment.
[0023] Furthermore, the entity categories of the road segment layer include road segments, pavement performance time-series data, climate environment time-series data, traffic time-series data, and maintenance history;
[0024] The data attributes of the road segment include road name, road number, starting chainage, ending chainage, driving direction, lane, road grade, year of opening to traffic, design service life, design surface layer thickness, design base layer thickness, design subbase layer thickness, and design materials of each structural layer.
[0025] The data attributes of the time-series pavement performance data include pavement condition index, international roughness index and rutting depth for different years;
[0026] The data attributes of the climate and environmental time series data include the annual average maximum temperature, number of high-temperature days, annual precipitation, number of annual temperature cycles, number of annual freeze-thaw cycles, annual average minimum temperature, and number of low-temperature days for different years.
[0027] The data attributes of the traffic time series data include the equivalent standard axle load, annual average daily truck traffic volume, and annual average daily traffic volume for different years.
[0028] The data attributes of the maintenance history include the maintenance measures and maintenance thickness corresponding to different years.
[0029] Furthermore, the entity categories in the case layer include maintenance cases, pavement defects, causes of defects, pavement structure types, previous maintenance measures, planned maintenance measures, and special areas;
[0030] The data attributes of the maintenance cases include case number, inspection time, road age, design speed, pavement condition index, international roughness index, rutting depth, equivalent standard axle load, average daily truck traffic, average daily traffic volume, average annual maximum temperature, number of high-temperature days per year, annual precipitation, number of temperature cycles per year, number of freeze-thaw cycles per year, average annual minimum temperature, and number of low-temperature days per year.
[0031] The data attributes of the pavement structure type include the current overlay thickness, current surface layer thickness, current base layer thickness, current subbase layer thickness, and current materials of each structural layer;
[0032] The data attributes of the previous maintenance measures include the type of the previous maintenance measures, the time of the previous maintenance, and the thickness of the previous maintenance.
[0033] The data attributes of the planned maintenance measures include the type of planned maintenance measure, the planned maintenance time, and the planned maintenance thickness.
[0034] Furthermore, the object properties include:
[0035] The hierarchical relationship between road segments and case studies;
[0036] Data correlations between time-series data on road segment and pavement performance, climate environment, traffic, and maintenance history;
[0037] The relationship between maintenance cases and pavement distress, pavement structure type, previous maintenance measures, planned maintenance measures, and special areas;
[0038] The causal relationship between road surface defects and their causes;
[0039] And the preceding and following relationships between different cases;
[0040] In this process, based on the road segment to which the maintenance case belongs and its inspection time, the preceding and subsequent relationships between different cases corresponding to the same road segment are established in chronological order, forming a time-series case chain that reflects the changes in pavement condition, pavement defects and maintenance measures over time.
[0041] Furthermore, step 3 specifically involves:
[0042] Based on the hierarchical ontology model, the entity categories and annotation rules that need to be annotated are determined. Unstructured text in the maintenance case library is manually annotated, and the annotation results are reviewed and corrected to form an annotated corpus.
[0043] The named entity recognition model is trained using the labeled corpus, and the trained named entity recognition model is used to extract entities corresponding to road surface defects, causes of defects and maintenance measures from the unstructured text to be processed, as entities to be standardized.
[0044] A domain terminology list organized by entity category is pre-constructed to provide a standardized expression for each entity category, and terms that are the same as the entity category to be standardized are selected from the domain terminology list as candidate standard terms;
[0045] A semantic matching model based on a pre-trained language model is used to calculate the semantic matching score between the entity to be standardized and each candidate standard term.
[0046] The candidate standard term with the highest semantic matching score and reaching the preset matching threshold is selected as the target standard term. The entity to be standardized is mapped to the target standard term to achieve entity alignment and disambiguation, forming a standardized entity, thereby obtaining the final case knowledge.
[0047] Furthermore, step 4 specifically involves:
[0048] Based on the hierarchical ontology model, case knowledge is converted into entity, data attribute, and object attribute, and imported into the asphalt pavement maintenance management knowledge graph and stored in the graph database. Entity categories are mapped to node labels, entity instances are mapped to nodes, data attributes are mapped to node attributes, and object attributes are mapped to relationships between nodes.
[0049] The nodes in the knowledge graph are divided into global nodes and case-scope nodes. The global nodes are uniquely stored in the knowledge graph, and the case-scope nodes are created according to their corresponding maintenance cases and associated with the corresponding maintenance case nodes.
[0050] Furthermore, step 5 specifically involves:
[0051] Acquire newly added or changed maintenance case data, and preprocess, extract entities, and align entities to form case knowledge to be updated;
[0052] Based on the road segment, case number, and inspection time, the case knowledge to be updated is matched and located with the existing case knowledge in the knowledge graph. While retaining the original knowledge graph structure, only the nodes, node attributes, and relationships between nodes corresponding to newly added or changed maintenance cases are locally added, modified, and associated to achieve incremental updates of the knowledge graph.
[0053] When no matching case knowledge exists in the knowledge graph, add corresponding nodes, node attributes, and relationships between nodes;
[0054] When a matching case knowledge exists in the knowledge graph, the corresponding node, node attributes, and relationships between nodes are updated.
[0055] Based on the detection time of the newly added maintenance cases, they are connected to the time-series case chain corresponding to the same road segment, and the preorder and postorder relationships between the newly added maintenance cases and existing maintenance cases are established.
[0056] The updated nodes, node attributes, and relationships between nodes are validated for consistency. Duplicate nodes, conflicting relationships, and abnormal attribute values are identified and processed to achieve incremental updates and dynamic management of the knowledge graph.
[0057] Furthermore, step 6 specifically involves:
[0058] Obtain the road segment to be decided at the decision time The data includes pavement performance, pavement structure, traffic load, climate environment, and maintenance history. The acquired structured data is then processed through feature extraction, numerical encoding, and standardization to obtain a road condition feature vector. ;
[0059] Entities associated with the road segment to be decided are identified from the optimized knowledge graph. A local knowledge subgraph oriented towards the current decision state is constructed. The entities, entity attributes, and semantic relationships in the local knowledge subgraph are encoded and relationships are aggregated to obtain a knowledge feature vector. ;
[0060] The road condition feature vector With the knowledge feature vector The data is pieced together to construct the road segment to be decided at the decision-making time. Knowledge-enhanced state vector The knowledge is used to enhance the state vector. As the state input to a deep reinforcement learning model, it is represented as =[ , ];
[0061] Constructing the initial action space, including no maintenance and various maintenance measures. According to the road segment to be decided at the decision time Knowledge-enhanced state vector The system retrieves the applicable conditions and constraints for each candidate maintenance measure based on technical specifications, engineering standards, and expert experience, and generates an action mask accordingly. Candidate maintenance actions that do not meet the current state are filtered out to obtain the executable action space. ;
[0062] A comprehensive target reward system is constructed based on pavement service performance, maintenance costs, and environmental impact. Simultaneously, based on the applicability relationships between defect types, defects causes, and maintenance measures represented in the knowledge graph, the matching degree between candidate maintenance actions and the current pavement condition is determined. Based on this, a knowledge-guided reward system is constructed. The comprehensive target reward and the knowledge-guided reward are then weighted and fused to obtain an immediate reward. ;
[0063] A state transition environment is constructed, in which the next decision time is calculated based on the pavement performance evolution model under the condition that no maintenance is implemented. The natural decay state is considered. Under maintenance conditions, the pavement performance state after maintenance intervention is calculated based on the natural decay state and the maintenance effect model corresponding to the candidate maintenance actions, and then the next decision time is determined. Traffic load, climate environment, road age, and maintenance history are used to update other state variables, thus obtaining the next decision time for the road segment to be decided. Knowledge-enhanced state vector ;
[0064] Based on the knowledge-enhanced state vector Current maintenance actions Instant rewards and the knowledge-enhanced state vector at the next decision moment Build state transition experience And store it in the experience replay pool;
[0065] A deep reinforcement learning model is used to batch sample and train the state transition experience. The loss function is calculated by combining the action value in the current state with the target action value in the next state, and the network parameters are iteratively updated to maximize the cumulative discount reward within the planning period until the model meets the preset convergence condition.
[0066] The trained deep reinforcement learning model is used to evaluate the value of each candidate maintenance action in the executable action space. The candidate maintenance action with the highest action value is selected as the target action at the current decision moment. The road surface condition is updated according to the preset decision cycle and continuous maintenance decisions are made. Based on the sequence of target actions at each decision moment within the planning cycle, the maintenance measures, implementation timing and corresponding maintenance costs of the road segment to be decided are determined.
[0067] Beneficial Effects: Compared with existing technologies, this invention has the following significant advantages: First, it constructs a maintenance case library based on multi-source heterogeneous information; second, it constructs a hierarchical ontology model for asphalt pavement maintenance management, and sequentially performs knowledge extraction, entity alignment, and entity disambiguation; subsequently, it stores the processed case knowledge in a knowledge graph, and achieves dynamic management of the knowledge graph through an incremental update mechanism; finally, it uses a knowledge-constrained deep reinforcement learning model for maintenance decision reasoning. This invention, by constructing a knowledge graph for asphalt pavement maintenance management, proposes an intelligent decision-making paradigm that combines data-driven and knowledge-guided approaches. This can improve the accuracy and interpretability of maintenance decisions, enhance the model's generalization ability, scalability, and robustness under different road conditions, and provide scientific, efficient, and reliable technical support for asphalt pavement maintenance management. Attached Figure Description
[0068] Figure 1 This is a flowchart illustrating the domain knowledge graph construction method for intelligent decision-making in asphalt pavement maintenance provided by the present invention.
[0069] Figure 2 This is a schematic diagram of the asphalt pavement maintenance and management knowledge graph provided by the present invention;
[0070] Figure 3 This is a partial schematic diagram of the asphalt pavement maintenance and management knowledge graph provided by the present invention;
[0071] Figure 4 This is a flowchart illustrating the construction and application of the asphalt pavement maintenance management knowledge graph provided by this invention. Detailed Implementation
[0072] The technical solution of the present invention will be further described below with reference to the accompanying drawings.
[0073] This invention provides a method for constructing a domain knowledge graph for intelligent decision-making in asphalt pavement maintenance. To further illustrate the technical solution of this invention, the following is combined with... Figure 1 The flowchart shown further illustrates the present invention. Figure 1 This is a flowchart illustrating the intelligent decision-making and dynamic management method for asphalt pavement maintenance based on knowledge graphs as described in this invention.
[0074] The embodiments of this invention mainly include six steps: case library construction, hierarchical ontology model construction, knowledge extraction and entity alignment, knowledge graph storage, incremental knowledge graph update, and deep reinforcement learning decision-making. The following detailed explanation of each step is provided in conjunction with the specific implementation process.
[0075] Step 1: Build a maintenance case library to store heterogeneous information of maintenance cases from multiple sources.
[0076] The process involves acquiring multi-source heterogeneous information generated during asphalt pavement maintenance and management. This multi-source heterogeneous information includes basic road information, pavement structure information, pavement performance information, pavement distress information, traffic load information, climate environment information, and maintenance history information.
[0077] In this embodiment, the multi-source heterogeneous information originates from road condition inspection reports, defect investigation records, road design data, maintenance design documents, maintenance construction records, traffic volume survey data, meteorological monitoring data, and other road infrastructure information management platforms. The multi-source heterogeneous information includes structured numerical data stored in database fields or spreadsheets, as well as unstructured text data stored in the form of defect descriptions, causal analyses, and maintenance recommendations.
[0078] The acquired multi-source heterogeneous information is preprocessed, including format conversion, unit unification, and identification and deletion of missing and outlier values.
[0079] The multi-source heterogeneous information is classified and processed according to its data format. For numerical data, data from different sources within the same data category are converted to a consistent data format and unit of measurement; for text data, its original content is preserved for subsequent knowledge extraction, entity alignment, and disambiguation processing.
[0080] The data is checked for integrity and outliers. If a maintenance case has missing key data, data format that does not meet preset requirements, or values that exceed preset project ranges, it is determined to be an invalid case and the corresponding entire data record is deleted.
[0081] The processed heterogeneous information is associated and classified into cases according to road segments and inspection times, and a unique case number is assigned to each maintenance case.
[0082] In this embodiment, the case number is generated by combining the road number, starting station number, driving direction, lane and detection year with letters and numbers according to the preset coding rules.
[0083] Based on the case number, the preprocessed structured numerical data corresponding to the same maintenance case are integrated, and the corresponding disease description, cause analysis, maintenance suggestions and other unstructured text data are associated with the maintenance case, thereby forming a maintenance case composed of structured data and unstructured data to be extracted, thus constructing a case library.
[0084] Step 2: Construct a hierarchical ontology model for the field of asphalt pavement maintenance and management, define entity categories, attributes and their semantic relationships, and realize knowledge modeling for maintenance scenarios.
[0085] A hierarchical ontology model for asphalt pavement maintenance and management is constructed, and this hierarchical ontology model is divided into a road segment layer and a case layer. The ontology diagram of the asphalt pavement maintenance and management knowledge graph is shown below. Figure 2 .
[0086] In this embodiment, the hierarchical ontology model is constructed using the Protégé ontology editing tool, and entity categories, data attributes, object attributes, and constraint rules are formally expressed based on the OWL language.
[0087] Construct the entity categories and data attributes corresponding to the road segment layer to describe the basic information, design information, and pavement performance, traffic load, climate environment, and maintenance history information of the road segment over time.
[0088] The entity categories of the road segment layer include road segments, pavement performance time series data, climate environment time series data, traffic time series data, and maintenance history.
[0089] The data attributes of the road segment include road name, road number, starting station number, ending station number, driving direction, lane, road grade, year of opening to traffic, design service life, design surface layer thickness, design base layer thickness, design subbase layer thickness, and design materials of each structural layer.
[0090] The data attributes of the time-series pavement performance data include pavement condition index, international smoothness index, and rut depth for different years.
[0091] The data attributes of the climate and environmental time series data include the annual average maximum temperature, the number of high-temperature days, the annual precipitation, the number of annual temperature cycles, the number of annual freeze-thaw cycles, the annual average minimum temperature, and the number of low-temperature days for different years.
[0092] The data attributes of the traffic time series data include the equivalent standard axle load, annual average daily truck traffic volume, and annual average daily traffic volume for different years.
[0093] The data attributes of the maintenance history include the maintenance measures and maintenance thickness corresponding to different years.
[0094] Construct the entity categories and data attributes corresponding to the case layer to describe the road surface condition, disease characteristics, disease causes, road surface structure and maintenance measures of the road segment at different detection times.
[0095] The entity categories in the case layer include maintenance cases, pavement defects, causes of defects, pavement structure types, previous maintenance measures, planned maintenance measures, and special areas.
[0096] The data attributes of the maintenance cases include case number, inspection time, road age, design speed, pavement condition index, international roughness index, rut depth, equivalent standard axle load, average daily truck traffic, average daily traffic volume, average annual maximum temperature, number of high-temperature days per year, annual precipitation, number of temperature cycles per year, number of freeze-thaw cycles per year, average annual minimum temperature, and number of low-temperature days per year.
[0097] The road surface defects are not assigned data attributes and are used to characterize defects such as alligator cracks, block cracks, longitudinal cracks, transverse cracks, subsidence, ruts, wave-like embracing, potholes, loosening, bleeding, and repair.
[0098] The causes of the disease are not assigned data attributes and are used to characterize environmental, traffic, structural, material, and construction factors.
[0099] The data attributes of the road structure type include the current overlay thickness, current surface layer thickness, current base layer thickness, current subbase layer thickness, and current materials of each structural layer.
[0100] The data attributes of the previous maintenance measures include the type of the previous maintenance measures, the time of the previous maintenance, and the thickness of the previous maintenance.
[0101] The data attributes of the planned maintenance measures include the type of planned maintenance measure, the planned maintenance time, and the planned maintenance thickness.
[0102] The special areas are not assigned data attributes and are used to characterize spatial locations such as bridge decks, tunnels, curves, and intersections.
[0103] Define object attributes between different entity categories. These attributes include: the hierarchical relationship between road segments and cases, defined as `has_case`; the data association relationships between road segments and pavement performance time-series data, climate environment time-series data, traffic time-series data, and maintenance history, defined as `has_pavement_performance_time_series_data`, `has_environment_time_series_data`, `has_traffic_time_series_data`, and `has_maintenance_history`, respectively; the case association relationships between maintenance cases and pavement distress, pavement structure type, previous maintenance measures, planned maintenance measures, and special areas, defined as `has_distress`, `has_pavement_structure_type`, `has_last_maintenance_treatment`, `has_planned_maintenance_treatment`, and `has_special_area`, respectively; the causal relationship between pavement distress and its causes, defined as `is_caused_by`; and the pre- and post-case relationships between different cases, defined as `has_subsequent_case` and `has_preceding_case`, respectively.
[0104] In this process, based on the road segment to which the maintenance case belongs and its inspection time, the preceding and subsequent relationships between different cases corresponding to the same road segment are established in chronological order, forming a time-series case chain that reflects the changes in pavement condition, pavement defects and maintenance measures over time.
[0105] In this embodiment, each maintenance case is only established with the preceding and following maintenance cases that are adjacent in time, so as to reduce redundant relationships and maintain the continuity of the time case chain.
[0106] Step 3: Perform knowledge extraction based on the named entity recognition algorithm to convert unstructured text into structured data, and perform entity alignment and disambiguation processing.
[0107] Based on the ontology model, the entity categories and annotation rules to be annotated are determined. An original corpus is constructed by selecting unstructured text from maintenance design documents, road condition inspection reports, pavement defect investigation records, maintenance construction records, and maintenance evaluation reports. The entity boundaries and entity categories corresponding to pavement defects, their causes, and maintenance measures in the unstructured text are manually annotated. The annotation results are then reviewed and corrected to form an annotated corpus.
[0108] In this embodiment, the BIO annotation method is used for entity annotation, where B represents the starting position of the entity, I represents the internal position of the entity, and O represents the non-entity position. Initial annotation is completed by at least two annotators, and the annotation results are cross-checked. When the annotation results from different annotators are inconsistent, an expert in road maintenance is consulted for review and correction.
[0109] A named entity recognition model is trained using the labeled corpus, and the trained named entity recognition model is used to extract entities corresponding to road surface defects, causes of defects and maintenance measures from the unstructured text to be processed, as entities to be standardized.
[0110] In this embodiment, the named entity recognition model adopts the BERT-BiLSTM-CRF model, where BERT is used to extract contextual semantic features, BiLSTM is used to learn the bidirectional dependencies of the text sequence, and CRF is used to obtain entity annotation results that satisfy the label transfer constraint conditions.
[0111] Based on road industry regulations, technical standards, and engineering experience, a domain terminology list organized by entity category is pre-constructed to provide standardized expressions for each entity category. Terms that are the same as the entity category to be standardized are selected from the domain terminology list as candidate standard terms.
[0112] The entity name, entity category, and context information of the entity to be standardized, along with the term name and term category of the candidate standard terms, are input into a semantic matching model based on a pre-trained language model to calculate the semantic matching score between the entity to be standardized and each candidate standard term.
[0113] In this embodiment, the semantic matching model adopts a cross-encoder model based on the BERT pre-trained model.
[0114] The candidate standard term with the highest semantic matching score and reaching the preset matching threshold is selected as the target standard term. The entity to be standardized is mapped to the target standard term to achieve entity alignment and disambiguation, forming a standardized entity, thereby obtaining the final case knowledge.
[0115] In this embodiment, the matching threshold is determined based on the F1 value on the validation set.
[0116] When the highest semantic matching score is lower than the matching threshold, the corresponding entity is marked as an entity to be manually confirmed, and it is mapped to existing standard terms or written as a new standard term in the domain terminology table based on the manual confirmation result.
[0117] Step 4: Store the standardized case knowledge, after knowledge extraction, entity alignment, and disambiguation, into the knowledge graph.
[0118] Based on the hierarchical ontology model, mapping rules are established between the case library and entities, data attributes, and object attributes in the ontology model. Following these mapping rules, structured numerical data is directly converted into corresponding entity instances and attribute information. For unstructured text data, after knowledge extraction, entity alignment, and disambiguation, the extracted and standardized entities are mapped to corresponding entity instances and attribute information; semantic relationships between entity instances are then constructed.
[0119] After the above transformation is completed, case knowledge conforming to the constraints of the hierarchical ontology model is generated based on the unique case number of the maintenance case, and then imported into the asphalt pavement maintenance management knowledge graph and stored in the graph database. Specifically, entity categories are mapped to node labels, entity instances are mapped to nodes, data attributes are mapped to node attributes, and object attributes are mapped to relationships between nodes.
[0120] In this embodiment, Neo4j graph database is used as the storage platform for the knowledge graph. Cypher language is used to write the mapped case knowledge into the graph database. The schematic diagram of the constructed asphalt pavement maintenance management knowledge graph is shown below. Figure 3 As shown.
[0121] The nodes in the knowledge graph are divided into global nodes and case-scope nodes. The global nodes are uniquely stored in the knowledge graph to avoid the same global node being created repeatedly. The case-scope nodes are created according to their corresponding maintenance cases and are associated with the corresponding maintenance case nodes.
[0122] In this embodiment, road segments, maintenance cases, and associated pavement performance time-series data, climate environment time-series data, traffic time-series data, and maintenance history are set as global nodes; pavement defects, causes of defects, pavement structure types, previous maintenance measures, planned maintenance measures, and special areas are set as case scope nodes.
[0123] Step 5: Dynamically manage and continuously optimize the knowledge graph through an incremental update mechanism.
[0124] Acquire newly added or changed maintenance case data, and preprocess, extract entities, and align entities to form case knowledge to be updated.
[0125] Based on the road segment, case number, and inspection time, the knowledge of the case to be updated is matched and located with the existing case knowledge in the knowledge graph. While retaining the original knowledge graph structure, only the nodes, node attributes, and relationships between nodes corresponding to the newly added or changed maintenance cases are locally added, modified, and associated to achieve incremental updates of the knowledge graph without overall reconstruction or full replacement of the knowledge graph.
[0126] When no matching case knowledge exists in the knowledge graph, a corresponding maintenance case node is created, and related nodes are created or reused based on the case knowledge to be updated, establishing semantic relationships between nodes. Specifically, for nodes already existing in the knowledge graph, the relationship between the newly added maintenance case and the node is directly established to avoid creating duplicate nodes.
[0127] When matching case knowledge exists in the knowledge graph, the case knowledge to be updated is compared with the existing case knowledge. For newly added nodes, attributes, or semantic relationships, they are written into the knowledge graph; for nodes, attributes, or semantic relationships that have changed, they are modified; and for nodes, attributes, or semantic relationships that are no longer valid, they are deleted.
[0128] Based on the detection time of the newly added maintenance case, it is connected to the time-series case chain corresponding to the same road segment, establishing the pre- and post-series relationships between the newly added maintenance case and existing maintenance cases. When the newly added maintenance case is at the end of the time-series case chain, a pre- and post-series relationship is established between it and existing end cases; when the newly added maintenance case is at the beginning of the time-series case chain, a pre- and post-series relationship is established between it and existing beginning cases; when the newly added maintenance case is between two existing maintenance cases, the time-series relationship between the original adjacent cases is adjusted, and relationships between the newly added maintenance case and its pre- and post-series cases are established respectively, in order to maintain the continuity and temporal order of the time-series case chain.
[0129] The updated nodes, node attributes, and relationships between nodes undergo consistency checks, including node uniqueness checks, attribute data type checks, attribute value range checks, and time sequence relationship checks. Duplicate nodes are merged, conflicting relationships are corrected, and abnormal attribute values are marked, corrected, or deleted, thus achieving incremental updates and dynamic management of the knowledge graph. The flowchart illustrating the construction and application of the asphalt pavement maintenance management knowledge graph provided by this invention is shown below. Figure 4 As shown.
[0130] Step 6: Based on the knowledge graph, a deep reinforcement learning model with knowledge constraints is used to implement maintenance decision reasoning and output a comprehensive decision solution covering maintenance measures, optimal maintenance timing, and maintenance costs.
[0131] Historical maintenance case data are extracted from the case library to construct deep reinforcement learning training samples for maintenance decision-making.
[0132] Obtain the road segment to be decided at the decision time The data includes pavement performance, pavement structure, traffic load, climate environment, and maintenance history. The acquired structured data is then processed through feature extraction, numerical encoding, and standardization to obtain a road condition feature vector. .
[0133] The road condition feature vector This includes pavement condition index, international roughness index, rut depth, pavement age, current overlay thickness, current surface layer thickness, current base course thickness, current subbase course thickness, current materials of each structural layer, equivalent standard axle load, annual average daily truck traffic volume, annual average daily traffic volume, annual average maximum temperature, number of high-temperature days per year, annual precipitation, annual temperature cycle count, annual freeze-thaw cycle count, annual average minimum temperature, number of low-temperature days per year, and characteristics corresponding to the previous maintenance measures.
[0134] In this embodiment, numerical data is processed using the Min-Max normalization method to map each feature value to the [0,1] interval; categorical data is numerically represented using the One-Hot Encoding method to obtain unified feature data that can be used for model calculation.
[0135] Entities associated with the road segment to be decided are identified from the asphalt pavement maintenance and management knowledge graph. A local knowledge subgraph oriented towards the current decision state is constructed. The entities, entity attributes, and semantic relationships in the local knowledge subgraph are encoded and relationships are aggregated to obtain a knowledge feature vector. .
[0136] In this embodiment, a relational graph convolutional network is used to encode the retrieved knowledge subgraph, and the entity features associated with the road segment to be decided and their neighborhood relationship information are aggregated to obtain the feature representation of the knowledge subgraph.
[0137] The road condition feature vector With the knowledge feature vector The data is pieced together to construct the road segment to be decided at the decision-making time. Knowledge-enhanced state vector The knowledge is used to enhance the state vector. As the state input to a deep reinforcement learning model, it is represented as =[ , ].
[0138] Constructing the initial action space, including no maintenance and various maintenance measures. According to the road segment to be decided at the decision time Knowledge-enhanced state vector The system retrieves the applicable conditions and constraints for each candidate maintenance measure based on technical specifications, engineering standards, and expert experience, and generates an action mask accordingly. Candidate maintenance actions that do not meet the current state are filtered out to obtain the executable action space. .
[0139] In this embodiment, for candidate maintenance actions that meet the current requirements, their action mask value is set to 1; for candidate maintenance actions that do not meet the current requirements, their action mask value is set to 0.
[0140] A comprehensive target reward based on pavement service performance, maintenance costs, and environmental impact is constructed. Simultaneously, based on the applicability relationship between defect type, defect cause, and maintenance measures represented in the knowledge graph, the matching degree between candidate maintenance actions and the current pavement condition is determined, and a knowledge-guided reward is constructed accordingly. The comprehensive target reward and the knowledge-guided reward are then weighted and fused to obtain an immediate reward. , is represented as:
[0141]
[0142] in, Indicates the moment of decision Instant rewards; This indicates the degree of improvement in pavement service performance after the implementation of candidate maintenance actions; This indicates the maintenance cost corresponding to the execution of the candidate maintenance action; Indicate the environmental benefits resulting from performing candidate maintenance actions; This indicates the degree of knowledge matching for candidate maintenance actions; , , and This represents the corresponding weighting coefficient.
[0143] In this embodiment, the degree of improvement in pavement service performance is determined based on the changes in pavement condition index, international roughness index, and rutting depth before and after the implementation of the candidate maintenance action; the maintenance cost is determined based on the type of planned maintenance measures, the thickness of planned maintenance measures, the amount of maintenance work, and the corresponding expenses; and the environmental benefits are determined based on at least one of the carbon emissions, energy consumption, material consumption, and waste generation generated during the implementation of the candidate maintenance action.
[0144] In this embodiment, the knowledge matching degree is used to characterize the applicability between candidate maintenance actions and the current state of the road segment to be decided. Based on the applicability relationship between candidate maintenance measures and disease types, disease causes, pavement structure, pavement performance status, traffic load and climate environment in the knowledge graph, the matching score of candidate maintenance actions for each applicable condition is calculated, and weighted fusion is performed according to preset weights to obtain the knowledge matching degree of candidate maintenance actions.
[0145] A state transition environment is constructed, in which the next decision time is calculated based on the pavement performance evolution model under the condition that no maintenance is implemented. The natural decay state is considered. Under maintenance conditions, the pavement performance state after maintenance intervention is calculated based on the natural decay state and the maintenance effect model corresponding to the candidate maintenance actions, and then the next decision time is determined. Traffic load, climate environment, road age, and maintenance history are used to update other state variables, thus obtaining the next decision time for the road segment to be decided. Knowledge-enhanced state vector .
[0146] Based on the knowledge-enhanced state vector Current maintenance actions Instant rewards and the knowledge-enhanced state vector at the next decision moment Build state transition experience And store them in the experience replay pool.
[0147] A deep reinforcement learning model is used to batch sample and train the state transition experience. The loss function is calculated by combining the action value in the current state with the target action value in the next state, and the network parameters are iteratively updated to maximize the cumulative discount reward within the planning period until the model meets the preset convergence condition.
[0148] In this embodiment, the deep reinforcement learning model can be trained using Dueling Double Deep Q-Network (D3QN) and combined with experience replay and target network update mechanisms to improve the stability of action value estimation and decision performance.
[0149] The trained deep reinforcement learning model is used to evaluate the value of each candidate maintenance action in the executable action space. The candidate maintenance action with the highest action value is selected as the target action at the current decision moment. The road surface condition is updated according to the preset decision cycle and continuous maintenance decisions are made. Based on the sequence of target actions at each decision moment within the planning cycle, the maintenance measures, implementation timing and corresponding maintenance costs of the road segment to be decided are determined.
[0150] In this embodiment, the planning period can be set to 5 to 20 years according to the road maintenance and management needs, and 1 year is used as a decision cycle. When the target action of a certain decision cycle is a specific maintenance measure, the corresponding decision time is determined as the implementation time of the maintenance measure.
[0151] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A method for constructing a domain knowledge graph for intelligent decision-making in asphalt pavement maintenance, characterized in that, Includes the following steps: Step 1: Build a maintenance case library to store structured and unstructured text data of maintenance cases from multiple sources; Step 2: Construct a hierarchical ontology model for the field of asphalt pavement maintenance and management, define entity categories, attributes and semantic relationships, and realize knowledge modeling for maintenance scenarios; Step 3: Based on the named entity recognition algorithm, perform knowledge extraction operation on the unstructured text data in the maintenance case library to extract entities corresponding to pavement defects, causes of defects and maintenance measures, and perform entity alignment and disambiguation processing. Step 4: Based on the hierarchical ontology model, convert the structured data and standardized text knowledge in the maintenance case library into nodes, attributes, and relationships, and store them in the knowledge graph; Step 5: Dynamically manage and continuously optimize the knowledge graph through an incremental update mechanism; Step 6: Based on the optimized knowledge graph, implement a deep reinforcement learning model with knowledge constraints, perform maintenance decision reasoning, and output a comprehensive decision scheme covering maintenance measures, optimal maintenance timing, and maintenance costs.
2. The domain knowledge graph construction method for intelligent decision-making in asphalt pavement maintenance according to claim 1, characterized in that, In step 1, the multi-source maintenance case includes basic road information, pavement structure information, pavement performance information, pavement distress information, traffic load information, climate environment information, and maintenance history information; The construction of the maintenance case library specifically involves: The acquired multi-source heterogeneous information is preprocessed, including format conversion, unit unification, and identification and deletion of missing and outlier values. The processed multi-source maintenance cases are associated and classified according to road sections and inspection times, and a unique case number is assigned to each maintenance case. Based on the case number, the pavement performance, defects, traffic, environment and maintenance history information corresponding to the same maintenance case are integrated to form a maintenance case and stored in the case library.
3. The domain knowledge graph construction method for intelligent decision-making in asphalt pavement maintenance according to claim 1, characterized in that, Step 2 is as follows: A hierarchical ontology model for asphalt pavement maintenance and management is constructed, and the hierarchical ontology model is divided into a road segment layer and a case layer. Construct the entity categories and data attributes corresponding to the road segment layer to describe the basic information, design information, and pavement performance, traffic load, climate environment, and maintenance history information of the road segment over time; Construct the entity categories and data attributes corresponding to the case layer to describe the road surface condition, disease characteristics, disease causes, road surface structure and maintenance measures of the road segment at different detection times; Define object attributes between different entity categories, and establish hierarchical relationships between road segments and cases based on the object attributes, as well as temporal relationships between different cases of the same road segment.
4. The domain knowledge graph construction method for intelligent decision-making in asphalt pavement maintenance according to claim 3, characterized in that, The entity categories of the road segment layer include road segments, pavement performance time series data, climate environment time series data, traffic time series data, and maintenance history; The data attributes of the road segment include road name, road number, starting chainage, ending chainage, driving direction, lane, road grade, year of opening to traffic, design service life, design surface layer thickness, design base layer thickness, design subbase layer thickness, and design materials of each structural layer. The data attributes of the time-series pavement performance data include pavement condition index, international roughness index and rutting depth for different years; The data attributes of the climate and environmental time series data include the annual average maximum temperature, number of high-temperature days, annual precipitation, number of annual temperature cycles, number of annual freeze-thaw cycles, annual average minimum temperature, and number of low-temperature days for different years. The data attributes of the traffic time series data include the equivalent standard axle load, annual average daily truck traffic volume, and annual average daily traffic volume for different years. The data attributes of the maintenance history include the maintenance measures and maintenance thickness corresponding to different years.
5. The domain knowledge graph construction method for intelligent decision-making in asphalt pavement maintenance according to claim 3, characterized in that, The entity categories in the case layer include maintenance cases, pavement defects, causes of defects, pavement structure types, previous maintenance measures, planned maintenance measures, and special areas; The data attributes of the maintenance cases include case number, inspection time, road age, design speed, pavement condition index, international roughness index, rutting depth, equivalent standard axle load, average daily truck traffic, average daily traffic volume, average annual maximum temperature, number of high-temperature days per year, annual precipitation, number of temperature cycles per year, number of freeze-thaw cycles per year, average annual minimum temperature, and number of low-temperature days per year. The data attributes of the pavement structure type include the current overlay thickness, current surface layer thickness, current base layer thickness, current subbase layer thickness, and current materials of each structural layer; The data attributes of the previous maintenance measures include the type of the previous maintenance measures, the time of the previous maintenance, and the thickness of the previous maintenance. The data attributes of the planned maintenance measures include the type of planned maintenance measure, the planned maintenance time, and the planned maintenance thickness.
6. The domain knowledge graph construction method for intelligent decision-making in asphalt pavement maintenance according to claim 3, characterized in that, The object properties include: The hierarchical relationship between road segments and case studies; Data correlations between time-series data on road segment and pavement performance, climate environment, traffic, and maintenance history; The relationship between maintenance cases and pavement distress, pavement structure type, previous maintenance measures, planned maintenance measures, and special areas; The causal relationship between road surface defects and their causes; And the preceding and following relationships between different cases; In this process, based on the road segment to which the maintenance case belongs and its inspection time, the preceding and subsequent relationships between different cases corresponding to the same road segment are established in chronological order, forming a time-series case chain that reflects the changes in pavement condition, pavement defects and maintenance measures over time.
7. The method for constructing a domain knowledge graph for intelligent decision-making in asphalt pavement maintenance according to claim 1, characterized in that, Step 3 specifically involves: Based on the hierarchical ontology model, the entity categories and annotation rules that need to be annotated are determined. Unstructured text in the maintenance case library is manually annotated, and the annotation results are reviewed and corrected to form an annotated corpus. The named entity recognition model is trained using the labeled corpus, and the trained named entity recognition model is used to extract entities corresponding to road surface defects, causes of defects and maintenance measures from the unstructured text to be processed, as entities to be standardized. A domain terminology list organized by entity category is pre-constructed to provide a standardized expression for each entity category, and terms that are the same as the entity category to be standardized are selected from the domain terminology list as candidate standard terms; A semantic matching model based on a pre-trained language model is used to calculate the semantic matching score between the entity to be standardized and each candidate standard term. The candidate standard term with the highest semantic matching score and reaching the preset matching threshold is selected as the target standard term. The entity to be standardized is mapped to the target standard term to achieve entity alignment and disambiguation, forming a standardized entity, thereby obtaining the final case knowledge.
8. The method for constructing a domain knowledge graph for intelligent decision-making in asphalt pavement maintenance according to claim 1, characterized in that, Step 4 specifically involves: Based on the hierarchical ontology model, case knowledge is converted into entity, data attribute, and object attribute, and imported into the asphalt pavement maintenance management knowledge graph and stored in the graph database. Entity categories are mapped to node labels, entity instances are mapped to nodes, data attributes are mapped to node attributes, and object attributes are mapped to relationships between nodes. The nodes in the knowledge graph are divided into global nodes and case-scope nodes. The global nodes are uniquely stored in the knowledge graph, and the case-scope nodes are created according to their corresponding maintenance cases and associated with the corresponding maintenance case nodes.
9. A method for constructing a domain knowledge graph for intelligent decision-making in asphalt pavement maintenance according to claim 1, characterized in that, Step 5 specifically involves: Acquire newly added or changed maintenance case data, and preprocess, extract entities, and align entities to form case knowledge to be updated; Based on the road segment, case number, and inspection time, the case knowledge to be updated is matched and located with the existing case knowledge in the knowledge graph. While retaining the original knowledge graph structure, only the nodes, node attributes, and relationships between nodes corresponding to newly added or changed maintenance cases are locally added, modified, and associated to achieve incremental updates of the knowledge graph. When no matching case knowledge exists in the knowledge graph, add corresponding nodes, node attributes, and relationships between nodes; When a matching case knowledge exists in the knowledge graph, the corresponding node, node attributes, and relationships between nodes are updated. Based on the detection time of the newly added maintenance cases, they are connected to the time-series case chain corresponding to the same road segment, and the preorder and postorder relationships between the newly added maintenance cases and existing maintenance cases are established. The updated nodes, node attributes, and relationships between nodes are validated for consistency. Duplicate nodes, conflicting relationships, and abnormal attribute values are identified and processed to achieve incremental updates and dynamic management of the knowledge graph.
10. A method for constructing a domain knowledge graph for intelligent decision-making in asphalt pavement maintenance according to claim 1, characterized in that, Step 6 specifically involves: Obtain the road segment to be decided at the decision time The data includes pavement performance, pavement structure, traffic load, climate environment, and maintenance history. The acquired structured data is then processed through feature extraction, numerical encoding, and standardization to obtain a road condition feature vector. ; Entities associated with the road segment to be decided are identified from the optimized knowledge graph. A local knowledge subgraph oriented towards the current decision state is constructed. The entities, entity attributes, and semantic relationships in the local knowledge subgraph are encoded and relationships are aggregated to obtain a knowledge feature vector. ; The road condition feature vector With the knowledge feature vector The data is pieced together to construct the road segment to be decided at the decision-making time. Knowledge-enhanced state vector The knowledge is used to enhance the state vector. As the state input to a deep reinforcement learning model, it is represented as =[ , ]; Constructing the initial action space, including no maintenance and various maintenance measures. According to the road segment to be decided at the decision time Knowledge-enhanced state vector The system retrieves the applicable conditions and constraints for each candidate maintenance measure based on technical specifications, engineering standards, and expert experience, and generates an action mask accordingly. Candidate maintenance actions that do not meet the current state are filtered out to obtain the executable action space. ; A comprehensive target reward system is constructed based on pavement service performance, maintenance costs, and environmental impact. Simultaneously, based on the applicability relationships between defect types, defects causes, and maintenance measures represented in the knowledge graph, the matching degree between candidate maintenance actions and the current pavement condition is determined. Based on this, a knowledge-guided reward system is constructed. The comprehensive target reward and the knowledge-guided reward are then weighted and fused to obtain an immediate reward. ; A state transition environment is constructed, in which the next decision time is calculated based on the pavement performance evolution model under the condition that no maintenance is implemented. The natural decay state is considered. Under maintenance conditions, the pavement performance state after maintenance intervention is calculated based on the natural decay state and the maintenance effect model corresponding to the candidate maintenance actions, and then the next decision time is determined. The traffic load, climate environment, road age, and maintenance history are used to update other state variables, thus obtaining the road segment to be decided at the next decision time. Knowledge-enhanced state vector ; Based on the knowledge-enhanced state vector Current maintenance actions Instant rewards and the knowledge-enhanced state vector at the next decision moment Build state transition experience And stored in the experience replay pool; A deep reinforcement learning model is used to batch sample and train the state transition experience. The loss function is calculated by combining the action value in the current state with the target action value in the next state, and the network parameters are iteratively updated to maximize the cumulative discount reward within the planning period until the model meets the preset convergence condition. The trained deep reinforcement learning model is used to evaluate the value of each candidate maintenance action in the executable action space. The candidate maintenance action with the highest action value is selected as the target action at the current decision moment. The road surface condition is updated according to the preset decision cycle and continuous maintenance decisions are made. Based on the sequence of target actions at each decision moment within the planning cycle, the maintenance measures, implementation timing and corresponding maintenance costs of the road segment to be decided are determined.
