Domain knowledge graph construction method and system for unmanned aerial vehicle inspection

Abstract

The invention discloses a domain knowledge graph construction method and system for unmanned aerial vehicle inspection, and belongs to the technical field of unmanned aerial vehicle inspection. Aiming at the problems of dispersed data organization, insufficient fusion of multi-modal data, non-effective utilization of inspection process information, difficulty in identification of inspection blind areas and the like in the existing method, a space-time trajectory entity is introduced, an unmanned aerial vehicle flight trajectory is used as a knowledge graph core entity, and a space-time association relationship among the trajectory, equipment and defects is established; a multi-modal data fusion technology is adopted to realize deep fusion of visible light, infrared and multispectral data; through space-time density clustering analysis, a high-frequency inspection area and an inspection blind area are automatically identified, and the equipment coverage rate and the inspection timeliness are calculated. According to the invention, full-cycle knowledge representation of the inspection process is realized, the defect identification accuracy is improved, the inspection strategy is optimized, and support is provided for intelligent inspection management.

Description

Technical Field

[0001] This invention relates to the field of unmanned aerial vehicle (UAV) inspection technology, specifically to a method and system for constructing a domain knowledge graph for UAV inspection. Background Technology

[0002] With the rapid development of drone technology, drone inspection has become an important means of equipment maintenance in industries such as power, petroleum, and transportation, offering advantages such as high efficiency, low cost, and good safety. Existing drone inspection systems are typically equipped with sensors such as visible light cameras, infrared thermal imagers, and multispectral cameras, capable of collecting rich multimodal inspection image data. However, the following problems exist in data processing and utilization: First, there is a lack of data organization and correlation. Existing methods adopt a linear process of "collection-identification-reporting," resulting in the storage of inspection data in a fragmented form. Equipment information, defect information, inspection records, and other data are scattered across different systems, making it difficult to establish correlations between data, trace historical data, and accumulate and reuse knowledge.

[0003] Secondly, the multimodal data fusion capability is insufficient. Different modal data such as visible light images, infrared images, and multispectral images are often processed separately, failing to leverage the complementary advantages of multimodal data. This results in low accuracy of single-modal recognition, high false detection and false negative rates, and poor robustness in complex environments.

[0004] Secondly, information from the inspection process is not being effectively utilized. Existing methods focus on inspection results but neglect the value of the inspection process itself. Information such as the drone's flight path, inspection time, observation angle, and coverage area is crucial for assessing inspection quality, identifying blind spots, and optimizing inspection strategies; however, this information is typically not systematically recorded and analyzed.

[0005] Furthermore, the lack of inspection coverage analysis and blind spot identification mechanisms means that existing methods struggle to answer key questions such as "which equipment has been inspected," "which areas have inspection blind spots," and "is the inspection frequency reasonable." The absence of a systematic analysis and evaluation of inspection coverage results in a lack of quantitative assessment and optimization criteria for inspection work.

[0006] To address the aforementioned issues, it is necessary to propose a method and system for constructing a domain knowledge graph for unmanned aerial vehicle (UAV) inspection. Summary of the Invention

[0007] The purpose of this invention is to solve the problems existing in the background technology, and to propose a method and system for constructing a domain knowledge graph for UAV inspection.

[0008] The objective of this invention can be achieved through the following technical solutions: A method and system for constructing a domain knowledge graph for UAV inspection includes a method for constructing a domain knowledge graph for UAV inspection and a system for constructing a domain knowledge graph for UAV inspection.

[0009] In a first aspect, the present invention provides a method for constructing a domain knowledge graph for unmanned aerial vehicle (UAV) inspection, comprising the following steps: Step 1: Multi-source data acquisition; Collect drone flight data, multimodal inspection image data, basic information of the inspection object, industry knowledge documents and historical inspection reports to obtain raw data related to drone inspection.

[0010] Step 2: Entity recognition and extraction; Based on the original data, entity types are extracted and defined, a set of nodes for the knowledge graph is constructed, and entity disambiguation and linking are performed.

[0011] Based on the original data, define entity type ontology, use rule mapping, named entity recognition and computer vision technology to extract entities from different data sources, construct a node set of knowledge graph, and perform entity disambiguation and linking. The entity type ontology includes equipment entities, defect entities, location entities, status entities, personnel entities, time entities, and spatiotemporal trajectory entities; The spatiotemporal trajectory entity is obtained by performing spatiotemporal segmentation processing on the UAV flight trajectory, recording trajectory segment identifiers, spatiotemporal ranges, motion parameters and observation information, and establishing spatiotemporal associations between the trajectory entity and the equipment entity and defect entity; Step 3: Relationship extraction and attribute labeling; Based on the extracted and defined entity types, semantic relationships between entities are identified, and entity attribute information is labeled to construct the edge and node attributes of the knowledge graph.

[0012] Identify semantic relationships between extracted entities, including compositional relationships, causal relationships, temporal relationships, positional relationships, and maintenance relationships, and label static and dynamic attributes for entities to construct the edge and node attributes of the knowledge graph; As a preferred embodiment of the present invention, the specific process of relation extraction includes: Structured data: Relationships are extracted directly from foreign key constraints and related tables; Unstructured text: Relationship classification methods based on rules or deep learning are employed; Image data: Visual relationships are extracted using scene graph generation technology.

[0013] As a preferred embodiment of the present invention, a semantic relationship network between entities is constructed to enrich entity attribute information, including: Define the composition relationship: representing the hierarchical structure and topological connections of devices; Define causal relationship: representing the association between equipment and defects, and between defects and failures; Define temporal relationships: representing the temporal attributes and sequence dependencies of events; Define positional relationships: indicating the spatial location and ownership of equipment; Define maintenance relationships: This refers to the operational relationships between personnel, equipment, and defects.

[0014] Step 4: Knowledge Integration and Alignment; Entities from different data sources are aligned, and the same real-world object is identified based on identifier matching or comprehensive similarity calculation. Relationships are then fused, deduplicated, and conflict resolved to obtain a fused knowledge graph. As a preferred embodiment of the present invention, the specific process of aligning entities from different data sources includes: Identifier alignment: Prefer using unique identifiers for exact matching; Similarity alignment: When unique identifiers are lacking, a comprehensive similarity is calculated, including: Name similarity: Calculates text similarity based on string matching algorithms; Geographical similarity: Calculates spatial distance to determine whether they point to the same object; Attribute similarity: Compares the degree of matching of key attributes; Relationship structure similarity: compare the neighboring nodes and association patterns of entities.

[0015] Step 5: Knowledge Graph Storage and Indexing; The fused knowledge graph is stored in a graph database to establish a multi-level index structure, supporting efficient querying, reasoning, and incremental updates. Import the merged entity, relation, and attribute data into a graph database in batches and establish an attribute graph model mapping: entities are mapped to nodes and relations are mapped to edges, or establish an RDF model mapping: entities and relations are represented as triples.

[0016] Step Six: Spatiotemporal Clustering Analysis and Inspection Blind Spot Identification; Spatiotemporal density clustering is performed on historical inspection trajectories to identify high-frequency inspection areas and blind spots, calculate equipment coverage and inspection timeliness, and add the analysis results to the knowledge graph in entity form.

[0017] As a preferred embodiment of the present invention, the specific process of performing spatiotemporal density clustering on historical inspection trajectories includes: Trajectory preprocessing is performed, segmenting continuous flight trajectories by time intervals or spatial distances. Spatial coordinates, timestamps, motion features, observation features, and environmental features are extracted for each trajectory segment. Multiple spatial clusters are output, each representing a high-frequency region of interest. A density-based spatial clustering algorithm is used to cluster trajectory points, identifying high-frequency inspection areas. The neighborhood density of each point is calculated, core points are marked, and clusters are recursively expanded to form clusters.

[0018] Secondly, the present invention provides a domain knowledge graph construction system for UAV inspection, including a data acquisition module, an entity extraction module, a relationship construction module, a knowledge fusion module, a storage and query module, and a spatiotemporal analysis module.

[0019] The data acquisition module is used to collect UAV flight data, multimodal inspection image data, basic information of the inspection objects, industry knowledge documents, and historical inspection reports.

[0020] The entity extraction module is connected to the data acquisition module. It is used to extract and define entity types based on the raw data, construct a set of nodes for the knowledge graph, and perform entity disambiguation and linking.

[0021] The relationship building module is connected to the entity extraction module to identify semantic relationships between entities, label the attribute information of entities, and build the edges and node attributes of the knowledge graph. The knowledge fusion module is connected to the relationship building module and is used to perform entity alignment, identifier matching and similarity calculation on entity types to obtain a fused knowledge graph. The storage and query module is connected to the knowledge fusion module and is used to store the fused knowledge graph in a graph database, create indexes, and perform batch updates.

[0022] The spatiotemporal analysis module is connected to the storage and query module and is used to perform spatiotemporal density clustering on historical inspection trajectories, identify high-frequency inspection areas and blind spots, and evaluate inspection coverage. The spatiotemporal analysis module includes a trajectory preprocessing unit, a density clustering unit, a coverage evaluation unit, a blind spot identification unit, and a knowledge update unit.

[0023] The trajectory preprocessing unit is used to segment the continuous flight trajectory in time and space and extract trajectory point features; The density clustering unit is used to cluster trajectory points using a density-based spatial clustering algorithm to identify high-frequency inspection areas; The coverage assessment unit is used to calculate equipment coverage, timeliness score, and observation angle diversity. The blind spot identification unit is used to identify spatial blind spots and temporal blind spots; The knowledge update unit is used to add the spatiotemporal analysis results to the knowledge graph in entity form, establish relationships, and label risk levels.

[0024] Compared with the prior art, the beneficial effects of the present invention are: 1. This invention introduces spatiotemporal trajectory entities, treating the UAV's flight trajectory as a core entity in a knowledge graph. It records the spatiotemporal range, motion parameters, attitude changes, and observation information of trajectory segments, and establishes spatiotemporal relationships between trajectory entities and equipment and defect entities. This innovative design addresses the problem of existing knowledge graph methods focusing only on inspection objects and results while neglecting the inspection process. 2. This invention addresses the characteristics of UAVs equipped with multiple sensors (visible light camera, infrared thermal imager, multispectral camera) by proposing a multimodal fusion defect identification method based on a cross-modal attention mechanism. Through precise spatiotemporal alignment, deep feature fusion, flight attitude modulation, and temperature-vision-space three-dimensional correlation analysis, the method achieves complementary advantages of multimodal data, solving the problems of false detection, missed detection, and poor robustness in single-modal identification. 3. This invention establishes a multi-dimensional inspection coverage evaluation system by performing spatiotemporal density clustering analysis on historical inspection trajectories, thereby achieving automatic identification of inspection blind spots and intelligent optimization of inspection strategies. It solves the problems of traditional methods relying on human experience, uneven inspection coverage, and difficulty in discovering blind spots, forming an intelligent closed loop of "data-driven – knowledge analysis – strategy optimization". Attached Figure Description

[0025] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to the accompanying drawings: Figure 1 This is a flowchart of a domain knowledge graph construction method for UAV inspection proposed in an embodiment of the present invention; Figure 2 This is a system block diagram of a domain knowledge graph construction system for UAV inspection proposed in an embodiment of the present invention.

[0026] Figure 3 This is a schematic diagram of multimodal data acquisition and spatiotemporal alignment proposed in the embodiments of the present invention; Figure 4 This is a schematic diagram comparing the defect detection effects presented in the embodiments of the present invention; Figure 5 This is one of the schematic diagrams of spatiotemporal trajectory clustering proposed in the embodiments of the present invention; Figure 6 This is the second schematic diagram of spatiotemporal trajectory clustering proposed in the embodiments of the present invention; Figure 7 This is the third schematic diagram of spatiotemporal trajectory clustering proposed in the embodiments of the present invention; Figure 8 This is the fourth schematic diagram of spatiotemporal trajectory clustering proposed in the embodiments of the present invention; Detailed Implementation The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0027] First embodiment: A method for constructing a domain knowledge graph for UAV inspection; Please see Figure 1 As shown, a method for constructing a domain knowledge graph for UAV inspection includes the following steps: Step 1: Multi-source data acquisition; Collect drone flight data, multimodal inspection image data, basic information of the inspection object, industry knowledge documents and historical inspection reports to obtain raw data related to drone inspection.

[0028] Step 2: Entity recognition and extraction; Based on the original data, entity types are extracted and defined, a set of nodes for the knowledge graph is constructed, and entity disambiguation and linking are performed.

[0029] Step 3: Relationship extraction and attribute labeling; Based on the extracted and defined entity types, semantic relationships between entities are identified, and entity attribute information is labeled to construct the edge and node attributes of the knowledge graph.

[0030] Step 4: Knowledge Integration and Alignment; Entity alignment, identifier matching, and similarity calculation are performed on entity types to obtain a fused knowledge graph.

[0031] Step 5: Knowledge Graph Storage and Indexing; The integrated knowledge graph is stored in a graph database, indexed, and updated in batches.

[0032] Step Six: Spatiotemporal Clustering Analysis and Inspection Blind Spot Identification; Spatiotemporal clustering of historical inspection trajectories is performed to identify high-frequency inspection areas and blind spots.

[0033] In this embodiment, step one involves collecting UAV flight data, multimodal inspection image data, basic information of the inspection object, industry knowledge documents, and historical inspection reports. The drone flight data includes GPS positioning coordinates, flight altitude, flight speed, attitude angles (roll angle, pitch angle, heading angle), battery level, and wireless signal strength.

[0034] The process of acquiring multimodal inspection image data involves using sensors mounted on a UAV, such as a visible light camera, an infrared thermal imager, and a multispectral camera, to collect image data of the inspection target. The image resolution is no less than 1920×1080 pixels and includes EXIF ​​information (shooting time, geographical location, camera parameters, etc.). For infrared images, the temperature quantization range and pseudo-color mapping parameters are recorded.

[0035] Specifically, the basic information of the inspection targets consists of static attribute data obtained from the enterprise information management system or geographic information system, including equipment model, installation location, commissioning date, rated parameters, and historical maintenance records. The data is provided in relational database tables or XML format.

[0036] The process of collecting industry knowledge documents involves collecting text documents such as equipment manuals, inspection specifications, fault diagnosis standards, and safe operating procedures from the industry.

[0037] The process of collecting historical inspection reports involves acquiring structured or semi-structured report data generated by manual inspections or previous drone inspections, including defect descriptions, level assessments, handling suggestions, and closed-loop records.

[0038] As one implementation method of this embodiment, to ensure data quality, the collected UAV flight data, multimodal inspection image data, basic information of the inspection object, industry knowledge documents, and historical inspection report data undergo preliminary cleaning: duplicate records are removed, outliers (such as sensor readings outside the physical range) are filtered, the time format is unified to the ISO 8601 standard, and the coordinate system is unified to WGS-84. Data cleaning rules can be configured according to specific application scenarios.

[0039] In this embodiment, in step two, domain entities are extracted to construct a node set for the knowledge graph. The specific process is as follows: Based on the inspection business requirements and domain expert knowledge, predefine entity type ontology, including but not limited to: (1) Equipment entities: transmission lines, substations, towers, transformers, circuit breakers, insulators, conductors, and grounding devices; (2) Defective entities: corrosion, cracks, breakage, overheating, foreign matter, tilting, rust, leakage; (3) Location entity: Substation name, line name, tower number, geographical region (province / city / county), GPS coordinates; (4) Status entities: Normal, Abnormal, Critical, Urgent, Processed, Pending; (5) Personnel entities: Inspectors, auditors, maintenance personnel, and responsible persons; (6) Time entities: Inspection date, discovery time, and handling time; (7) Spatiotemporal trajectory entities: The UAV's flight trajectory is segmented to obtain its trajectory segment ID, start and end GPS coordinates, flight altitude range, timestamp interval, flight speed, attitude angle change sequence, and list of captured photos. Furthermore, a spatiotemporal association relationship is established between trajectory entities and equipment entities and defect entities, with the relationship labels being "inspection coverage," "observed," or "not covered." Using GPS coordinates and camera field of view (FOV), it is determined in which trajectory segments each device is observed.

[0040] As one implementation method of this embodiment, for structured data (such as database tables, JSON, XML), a rule mapping method is used to extract entities. For example, from the data table of the device management system, each row of records is mapped to a device entity, and the fields directly correspond to the entity attributes.

[0041] Example of extraction rules: SELECT device_id AS Entity ID, device_name AS Name, device_type AS Type, rated_voltage AS Rated Voltage, install_date AS Commissioning Date FROM equipment_table; As one implementation method of this embodiment, for unstructured data, including text data such as inspection reports and technical documents, Named Entity Recognition (NER) technology is used to extract entities. The specific process is as follows: Text preprocessing: The text is segmented, part-of-speech tagging is performed, and dependency parsing is conducted. A domain-adaptive Chinese word segmentation tool is used, and a custom dictionary is loaded (containing technical terms such as "SF6 circuit breaker"; "GIS combined electrical appliances"; etc.).

[0042] Sequence labeling: Based on the BERT pre-trained language model, combined with the BiLSTM-CRF sequence labeling network, entity boundaries and types in text are identified. The labeling adopts the BIO labeling system, where B represents device, I represents device, and O represents the start, inside, and outside of a device entity.

[0043] Model Training: The model was fine-tuned using manually annotated domain corpus. The training set contained no fewer than 10,000 annotated samples, covering typical inspection reports and technical documents. Cross-validation was used to evaluate model performance, achieving an F1 score of over 85%.

[0044] Entity normalization: Maps the identified entity names to a standard entity library.

[0045] For example, “10kV line”, “10kV line”, and “10kV transmission line” are normalized to the same entity. The system uses a combination of edit distance, pinyin similarity, and semantic similarity for calculation, and automatically normalizes the data by setting a threshold, which is configurable from 0.8 to 0.95.

[0046] As one implementation method of this embodiment, for image EXIF ​​data, the geographical location is extracted as a location entity, the shooting time is extracted as a time entity, and computer vision technology is used to identify equipment and defect entities in the image. The specific process is as follows: Object detection: YOLOv5 or Faster R-CNN deep learning models are used to detect equipment objects (such as towers, insulators, and conductors) and defect objects (such as damage, rust, and bird nests) in images. The models are trained on no fewer than 50,000 labeled images, achieving a detection accuracy (mAP) of over 90%.

[0047] Image segmentation: For defects that require precise localization (such as cracks or corroded areas), U-Net or Mask R-CNN is used for semantic segmentation or instance segmentation to obtain pixel-level defect boundaries.

[0048] Thermal anomaly detection: For infrared images, temperature threshold analysis and region growing algorithms are used to automatically identify areas of thermal anomaly, extract heat points as defect entities, and record attributes such as maximum temperature, average temperature, and anomaly area.

[0049] OCR text recognition: Performs optical character recognition on equipment nameplates and labels to extract text information such as equipment number and specifications as supplementary entity attributes.

[0050] In this embodiment, entity disambiguation and linking are performed in step two, and the specific process is as follows: Based on unique identifiers: Precise matching is performed using unique identifiers such as device ID and tower number.

[0051] Based on attribute similarity: When unique identifiers are lacking, the similarity of attributes such as entity name, location, and type is calculated comprehensively. TF-IDF vectors and cosine similarity are used to calculate text similarity, and the Haversine formula is used to calculate geographic distance. A combined threshold is set to determine whether they are the same entity.

[0052] Manual review mechanism: For candidate entity pairs with similarity in the fuzzy range (e.g., 0.7-0.85), they are submitted to domain experts for manual judgment, and the judgment results are fed back to the model for continuous learning.

[0053] It should be noted that the same entity may appear in different forms in different data sources, requiring entity disambiguation and linking to ensure that the same entity has only one unique representation in the knowledge graph.

[0054] As one implementation method of this embodiment, the entity extraction results are stored in a unified entity representation format, including entity ID (UUID), entity type, entity name, attribute key-value pairs, data source, and confidence level fields. All extracted entities form a candidate entity set, preparing for subsequent relation extraction and knowledge fusion.

[0055] Define attribute templates for each entity type. For example, the attributes of a device entity include: device ID, device name, device model, manufacturer, rated voltage, and commissioning date.

[0056] In this embodiment, in step three, the semantic relationships between entities include: Compositional relationships, causal relationships of faults and defects, temporal relationships, locational relationships, and fault-maintenance relationships.

[0057] In this embodiment, the attribute information of entities is labeled in step three to construct the edge and node attributes of the knowledge graph. The specific process is as follows: Define the composition relationship, which represents the hierarchical structure of equipment, such as a substation containing transformers, a tower containing insulators, and a transmission line connecting towers, etc. Define the causal relationship between faults and defects. The defect relationship represents the association between equipment and defects, such as the presence of damage to insulators or corrosion of wires. Define time relationships, which represent the temporal attributes of events, such as defect discovery time, equipment commissioning time, etc. Define location relationships, which indicate the location of equipment, such as a pole being located in a substation or equipment belonging to a certain line; Define the fault-maintenance relationship, which represents the maintenance operations performed by personnel on equipment or on defects, such as inspectors inspecting equipment and maintenance personnel handling defects. As one implementation method of this embodiment, a relation extraction model is used for complex sentence structures. Entity pairs and their contexts are taken as input, and semantic features are extracted using a BERT encoder. The relation type is then predicted through a classification layer. The model is trained on an annotated corpus containing no fewer than 20,000 annotated relation samples, and the accuracy should reach over 80%.

[0058] In this embodiment, the attribute information of the entity is annotated in step three, and the specific process is as follows: Provides detailed attribute information for entities, including static and dynamic attributes.

[0059] Among them, static attributes are directly extracted from structured data, including equipment model, rated parameters, manufacturer, installation date, etc., and are directly assigned to entity attribute fields; The dynamic attributes are extracted from sensor data and inspection results, including the current status, the most recent inspection time, and the latest number of defects. These dynamic attributes include timestamps.

[0060] In this embodiment, the edge and node attributes of the knowledge graph are constructed in step three, and the specific process is as follows: The results of relation extraction and attribute annotation are stored in the form of triples (head entity, relation, tail entity) or entity-attribute key-value pairs, with additional meta-information such as confidence level, data source, and timestamp, as the basic data of the knowledge graph.

[0061] In this embodiment, entity alignment is performed in step four, which fuses knowledge extracted from different data sources to eliminate redundancy and conflicts, forming a unified and consistent knowledge graph. The specific process is as follows: Identify entities from different data sources that point to the same real-world object and merge and align them, including identifier alignment and similarity alignment.

[0062] The specific process of identifier alignment is as follows: unique identifiers (such as device ID, tower number) are used first for precise matching, and entities from different sources but with the same identifier are merged into one entity node; The specific process of similarity alignment is as follows: when identifiers are lacking, the overall similarity of entities is calculated. The similarity calculation is as follows: Name similarity: The text similarity of entity names is calculated using methods such as edit distance (Levenshtein distance), Jaccard similarity, and cosine similarity. Geographical similarity: For entities with GPS coordinates, calculate the geographic distance. If the distance is less than a set threshold (e.g., 50 meters), it is considered that they may point to the same object. Attribute similarity: Calculates the degree of matching of other key attributes (such as model number and rated parameters); Relationship structure similarity: compare the neighboring entities and associations of two entities. The more similar the structures, the more likely they are to be the same entity.

[0063] As one implementation method of this embodiment, the overall similarity is achieved using a weighted average method, with the weights adjustable according to the application scenario. Entities with similarity exceeding a threshold (e.g., 0.85) are aligned with the candidates, and a manual review interface is provided.

[0064] As one implementation method of this embodiment, a merge alignment is performed. The aligned entities are then merged, retaining all source attribute information. For conflicting attributes (such as different model records of the same device), the following merging strategy is adopted: Select the attribute value with the highest confidence level; Select the attribute value for the latest timestamp; Retain all conflicting attribute values, mark them as multi-valued attributes, and record their origin.

[0065] As one implementation method of this embodiment, relation fusion is performed. For multiple candidate relations of the same entity pair, fusion and deduplication are performed, including: Relation deduplication: Identify duplicate relation triples (head entities, relation types, and tail entities are completely identical or semantically equivalent), retain one relation, and merge the confidence score and source information; Relationship conflict resolution: For contradictory relationships (such as the same device having both "Status: Normal" and "Status: Abnormal"), use time series analysis to retain the latest status; or retain the relationship with higher confidence based on the confidence threshold. Relational reasoning: Relational reasoning is performed based on rules defined in the ontology, completing missing relations. For example, the rule "If A contains B, B contains C, then A contains C" can infer transitive relations. The rule base contains reasoning rules summarized by domain experts, and reasoning is performed using a forward reasoning engine (such as an OWL-based inference engine).

[0066] As one implementation method of this embodiment, the quality of the fused knowledge graph is evaluated to identify and correct quality problems. The specific process is as follows: Completeness check: Check for missing key attributes, such as missing location attributes for equipment entities or missing discovery time for defective entities. For missing attributes, attempt to supplement them from the original data or mark them as needing improvement; Consistency check: Check the consistency of entity attributes and relationships, such as ensuring that the rated voltage of equipment matches the voltage level of its circuit, and that the defect level matches the description. Data that violates constraints is marked as abnormal and manually reviewed. Accuracy assessment: The accuracy of the knowledge graph is checked by sampling using expert review or cross-validation. Precision and recall are calculated: Precision = Number of correct triples / Number of triples drawn; Recall = Number of correct triples / Number of true triples.

[0067] Feedback mechanism: Provides a user feedback interface, allowing domain experts to flag erroneous entities or relationships. The system records the feedback and uses it for continuous model optimization.

[0068] It should be noted that step four, through knowledge fusion and alignment, forms redundant and high-quality core data of the knowledge graph, laying the foundation for subsequent storage and application.

[0069] In this embodiment, in step five, the constructed knowledge graph data is stored in a graph database, and an index is created to support efficient querying and reasoning. The specific process is as follows: Create a graph database (such as Neo4j or JanusGraph) and configure its storage path, memory allocation, transaction management, and other parameters. Import the merged entity, relationship, and attribute data into the graph database in batches using the database's provided batch import tools (such as Neo4j's neo4j-admin import or the Cypher LOAD CSV command) to improve import efficiency. Convert the data format to CSV or JSON, importing entities as nodes and relationships as edges.

[0070] In a graph database, an attribute graph model mapping is established: entities are mapped to nodes, node labels represent entity types, and node properties store entity attributes. Relationships are mapped to edges, edge types represent relationship types, and edge attributes store relationship attributes (such as timestamps and confidence levels).

[0071] Establish RDF model mappings in graph databases: entities and literal attributes represent subjects and objects, while relations and attribute names represent predicates, forming RDF triples (Subject-Predicate-Object). Use ontology IRIs (such as custom namespaces) to identify resources.

[0072] As one implementation method of this embodiment, an index is created in the graph database to support efficient querying and reasoning. The specific process is as follows: Create node attribute indexes for attributes that are frequently used query conditions, including device ID, device name, location, and timestamp.

[0073] For example, in Neo4j, you can create an index using CREATE INDEX ON :Device(deviceId); Build a full-text index to support keyword search. Integrate with the graph database using a full-text search engine such as Lucene or Elasticsearch; Create a spatial index to support range queries and nearest neighbor queries.

[0074] For example, query all poles and towers in a certain area, or the substation closest to a certain point.

[0075] Build graph structure indexes: Create graph topology indexes to accelerate graph traversal and path lookup. For example, pre-compute neighbors for nodes with high height (nodes connected by multiple edges) to reduce real-time traversal overhead; Establish data version management: Knowledge graphs evolve dynamically over time, requiring a version management mechanism.

[0076] Establish temporal attributes to add temporal attributes to entities and relationships, record the valid time range (validFrom, validTo), and support historical queries. For example, when a device's status changes, instead of deleting the old status, update the valid time and create a new status record.

[0077] Establish version snapshots and periodically create snapshots of the knowledge graph to save the complete state at a specific point in time, which facilitates backtracking and comparative analysis.

[0078] As one implementation method of this embodiment, incremental updates are performed on the graph database. For newly added inspection data, an incremental import method is adopted, adding or updating nodes and edges through the transaction API of the graph database to avoid full re-import. Unique constraints (such as device ID as primary key) are designed to prevent duplicate imports.

[0079] It should be noted that through continuous updates and maintenance, the knowledge graph can adapt to business changes and data growth, always maintaining high quality and high value, and providing reliable knowledge support for drone inspections.

[0080] As one implementation method of this embodiment, in step six, spatiotemporal clustering is performed on the historical inspection trajectories to identify high-frequency inspection areas and inspection blind spots. The specific process is as follows: The DBSCAN or OPTICS algorithm is used to perform spatial density clustering of trajectory points to identify high-frequency inspection areas. Through time dimension analysis, inspection blind spots are identified, and the clustering results are added to the knowledge graph as "inspection hotspot areas" and "inspection blind spots" entities, with risk levels marked.

[0081] Second embodiment: A domain knowledge graph construction system for UAV inspection; Please see Figure 2 As shown, a domain knowledge graph construction system for UAV inspection includes a data acquisition module, an entity extraction module, a relationship construction module, a knowledge fusion module, a storage and query module, and a spatiotemporal analysis module.

[0082] The data acquisition module is used to collect UAV flight data, multimodal inspection image data, basic information of the inspection objects, industry knowledge documents, and historical inspection reports.

[0083] The entity extraction module is connected to the data acquisition module. It is used to extract and define entity types based on the raw data, construct a set of nodes for the knowledge graph, and perform entity disambiguation and linking.

[0084] The relationship building module is connected to the entity extraction module to identify semantic relationships between entities, label the attribute information of entities, and build the edges and node attributes of the knowledge graph. The knowledge fusion module is connected to the relationship building module and is used to perform entity alignment, identifier matching and similarity calculation on entity types to obtain a fused knowledge graph. The storage and query module is connected to the knowledge fusion module and is used to store the fused knowledge graph in a graph database, create indexes, and perform batch updates.

[0085] The spatiotemporal analysis module is connected to the storage and query module and is used to perform spatiotemporal density clustering on historical inspection trajectories, identify high-frequency inspection areas and blind spots, and evaluate inspection coverage. The spatiotemporal analysis module includes a trajectory preprocessing unit, a density clustering unit, a coverage evaluation unit, a blind spot identification unit, and a knowledge update unit.

[0086] In this embodiment, the data acquisition module includes a flight data acquisition unit, an image data acquisition unit, a basic information acquisition unit, a document acquisition unit, and a report acquisition unit.

[0087] Among them, the flight data acquisition unit is used to collect the UAV's GPS positioning coordinates, flight altitude, flight speed, attitude angle, battery level and wireless signal strength; The image data acquisition unit is used to acquire image data of the inspection target through the visible light camera, infrared thermal imager and multispectral camera carried by the UAV, with an image resolution of not less than 1920×1080 pixels; The basic information collection unit is used to obtain static attribute data of the inspection objects from the enterprise information management system or geographic information system; The document collection unit is used to collect text documents such as equipment manuals, inspection specifications, fault diagnosis standards, and safe operating procedures from the industry. The report acquisition unit is used to acquire structured or semi-structured report data generated by manual inspections or previous drone inspections.

[0088] In this embodiment, the entity extraction module includes a structured data processing unit, a text data processing unit, and an image data processing unit.

[0089] The structured data processing unit is used to extract entities from database tables, JSON, and XML using rule mapping methods, mapping each row of records in the data table to an entity, with fields corresponding to entity attributes. The text data processing unit is used to extract entities from text data such as inspection reports and technical documents using named entity recognition technology. Specifically, it includes text preprocessing, entity recognition based on BERT pre-trained language model combined with BiLSTM-CRF sequence labeling network, model training, and entity normalization. The image data processing unit is used to identify equipment and defective entities in images using computer vision technology, including target detection using YOLO v5 or Faster R-CNN, image segmentation using U-Net or Mask R-CNN, thermal anomaly detection in infrared images, and OCR text recognition on equipment nameplates.

[0090] In this embodiment, the entity extraction module further includes an entity disambiguation unit, which is used to perform precise matching based on unique identifiers, or to comprehensively calculate the similarity of attributes such as entity name, location, and type based on attribute similarity, so as to ensure that the same entity has only one unique representation in the knowledge graph.

[0091] In this embodiment, the relationship construction module includes a relationship definition unit, a relationship extraction unit, and an attribute annotation unit.

[0092] Among them, the relation definition unit is used to define the types of relations between entities, including composition relations, fault-defect causal relations, time relations, location relations, and fault-maintenance relations; The relation extraction unit is used to directly extract relations from foreign key constraints and association tables in structured data, extract relations from text data using trigger word-based rule extraction or deep learning-based relation classification methods, and extract relations from image data using scene graph generation technology. The attribute annotation unit is used to annotate static and dynamic attributes of entities. Static attributes are extracted directly from structured data, while dynamic attributes are extracted from sensor data and inspection results and include timestamps.

[0093] In this embodiment, the relation extraction unit adopts a relation extraction model, takes entity pairs and their context as input, uses a BERT encoder to extract semantic features, and predicts the relation type through a classification layer.

[0094] In this embodiment, the knowledge fusion module includes an entity alignment unit, a relationship fusion unit, and a quality assessment unit.

[0095] The entity alignment unit is used to identify entities from different data sources that point to the same real-world object and perform merging and alignment, including exact matching based on unique identifiers and alignment based on comprehensive similarity, which includes name similarity, geographic location similarity, attribute similarity and relational structure similarity. The relation fusion unit is used to fuse and deduplicate multiple candidate relations for the same entity pair, including identifying duplicate relation triples, resolving contradictory relations, and performing relation reasoning based on rules defined by ontology. The quality assessment unit is used to assess the quality of the fused knowledge graph, including integrity checks, consistency checks, and accuracy assessments, and provides a user feedback interface.

[0096] In this embodiment, the storage query module includes a graph database storage unit, an index creation unit, and an incremental update unit.

[0097] The graph database storage unit is used to establish a graph database, and to import the merged entity, relation and attribute data into the graph database in batches. Entities are mapped to nodes and relations are mapped to edges. The index building unit is used to build node attribute indexes, full-text indexes, spatial indexes, and graph structure indexes to support efficient querying and reasoning; The incremental update unit is used to import new inspection data incrementally, adding or updating nodes and edges through the transaction API of the graph database, avoiding full re-import.

[0098] In this embodiment, the index building unit is also used to establish a data version management mechanism, including adding temporal attributes to entities and relationships to record the effective time range, and periodically creating version snapshots to save the complete state at a specific point in time.

[0099] In this embodiment, the trajectory preprocessing unit is used to perform spatiotemporal segmentation of the continuous flight trajectory and extract trajectory point features; The density clustering unit is used to cluster trajectory points using a density-based spatial clustering algorithm to identify high-frequency inspection areas; The coverage assessment unit is used to calculate equipment coverage, timeliness score, and observation angle diversity. The blind spot identification unit is used to identify spatial blind spots and temporal blind spots; The knowledge update unit is used to add the spatiotemporal analysis results to the knowledge graph in entity form, establish relationships, and label risk levels.

[0100] It should be noted that the data acquisition module, entity extraction module, relationship construction module, knowledge fusion module, storage and query module, and spatiotemporal analysis module are connected in sequence to form a complete knowledge graph construction process, realizing the automated construction from raw data to knowledge graph.

[0101] Third embodiment: A defect intelligent identification method based on multimodal fusion; Please see Figure 3 As shown in the diagram, multimodal data acquisition and spatiotemporal alignment are illustrated. In step three, a deep learning model is used to detect defects in device objects within the image. The specific process is as follows: Extract the intrinsic and extrinsic parameters (focal length, principal point coordinates, distortion coefficients, rotation matrix, translation vector) of each sensor, and establish the mapping relationship between the pixel coordinate system and the world coordinate system through camera calibration. Employ a feature-point-based registration method, using SIFT or ORB algorithms to extract feature points in visible light images, and set the descriptor dimension to 128 or 256 dimensions. In infrared images, corresponding feature points are extracted. Since the texture of infrared images is weak, edge enhancement preprocessing (Sobel operator or Canny operator) is used before feature extraction. Use the RANSAC algorithm to remove mismatched points, set the inlier threshold to 2 pixels, and iterate at least 1000 times. Based on the calculation of the homography matrix H (3×3 matrix) of the matching point pairs, the infrared image perspective transformation is transformed to the visible light image coordinate system; Repeat the above process for the multispectral image to achieve pixel-level alignment of the three modalities, with the registration error controlled within 3 pixels.

[0102] As one implementation method of this embodiment, an end-to-end multimodal defect detection neural network is designed, and the network structure includes: Visible light branch: ResNet-50 is used as the backbone network, with an input image size of 640×640 pixels and an output feature map size of 20×20×2048; Feature pyramids (FPN) are constructed at the outputs of layers 3, 4, and 5, corresponding to multi-scale features of 40×40, 20×20, and 10×10, respectively. Infrared Branch: To address the characteristics of weak texture and low contrast in infrared images, a lightweight feature extraction network is designed. The network structure is: Conv(64)-Conv(128)-Conv(256)-Conv(512), with each layer followed by BatchNorm and ReLU activation, and the output feature map size is 20×20×512; Multispectral branch: For multispectral images with 6 or more channels, 1×1 convolution is used to reduce the channel dimension to 3 channels, and then the MobileNetV3 network is input to extract features. The output feature map size is 20×20×960. Flight attitude branch: The flight altitude, speed, roll angle, pitch angle and yaw angle are combined into a 5-dimensional vector, which is input into a two-layer fully connected network (FC-128-FC-64) and outputs a 64-dimensional attitude feature vector.

[0103] As one implementation method of this embodiment, the multi-head attention mechanism of the Transformer architecture is used to fuse the features of each branch. The feature maps of visible light, infrared, and multispectral light are flattened into a sequence with a sequence length L=20×20=400 and a feature dimension d=512. Position encoding is added, and the 2D spatial position is encoded using a sine-cosine function. As one implementation method of this embodiment, an attitude modulation mechanism is introduced into the Transformer architecture: attention weights are dynamically adjusted based on flight attitude characteristics. When the roll angle or pitch angle is greater than 15°, the weight of the frame image is reduced; when the flight speed is greater than 10m / s, a motion blur detection branch is added.

[0104] As one implementation method of this embodiment, the fused features are input into the detection head, and the Faster R-CNN architecture is used to obtain the output defect category labels, including: corrosion, crack, breakage, heat generation, foreign matter, tilt, rust, leakage, loosening, deformation, dirt, and discharge. Please see Figure 4 As shown in the diagram, the defect detection results are compared. The first row, from left to right, represents the defect categories: corrosion, cracks, damage, and overheating. The second row, from left to right, lists the defect categories: foreign objects, tilting, corrosion, and leakage. The third row, from left to right, represents the defect categories: loosening, deformation, dirt, and discharge.

[0105] As one implementation of this embodiment, for each defect category label's detection frame, the temperature statistics of the corresponding area in the infrared image and the spectral features of the multispectral image are output simultaneously. The temperature statistics include the highest temperature, lowest temperature, average temperature, and standard deviation, while the spectral features of the multispectral image include the NDVI vegetation index and reflectance curve.

[0106] As one implementation of this embodiment, the severity rating and geometric features of the output device object defects are relative to each defect category label.

[0107] The severity rating output rules are as follows: if the crack length is >10cm or the area is >50cm², it is rated as "serious defect"; if the crack length is 5cm < 10cm, it is rated as "moderate defect"; if the crack length is ≤5cm, it is rated as "minor defect". The geometric features include: area, number of pixels, perimeter, aspect ratio, circularity, and fractal dimension for each defect category.

[0108] As one implementation method of this embodiment, the training process of the deep learning model is as follows: Collect no less than 100,000 sets of multimodal image data (each set includes a tripartite of visible light, infrared, and multispectral images), and manually annotate the defect bounding boxes, categories, and severity levels. Annotation quality control: each image is independently annotated by 3 annotators, and the final annotation is determined by voting or IOU thresholding (IOU>0.7). During training, the AdamW optimizer was used with an initial learning rate of 1e-4 and a cosine annealing learning rate decay strategy. The training qualification criterion was set as: defect classification accuracy ≥ 90%.

[0109] Fourth embodiment: A method for identifying blind spots in inspections based on spatiotemporal trajectory clustering; In step six, the DBSCAN or OPTICS algorithm is used to perform spatial density clustering of the trajectory points to identify high-frequency inspection areas. The specific process is as follows: Please see Figures 5 to 8 As shown, this is a schematic diagram of spatiotemporal trajectory clustering; Figure 5 In the middle, the colored trajectory lines represent 15 drone flight paths, with red cylinders representing 3 high-frequency inspection areas and blue dots representing random low-frequency inspection points; Figure 6 In the middle, the yellow-orange-red gradient heatmap represents the trajectory density distribution density heatmap and equipment coverage, the black contour lines are density isolines, the red dashed circles represent 3 high-frequency clustering areas with a radius of 600 meters, the green squares represent 39 covered devices, and the gray X marks represent 11 uncovered devices in the inspection blind spots; Figure 7 In the diagram, gray represents blind spots during inspections, and red represents high-frequency areas. Figure 8 In the chart, the purple bars represent the number of daily inspections, the red dashed lines represent the average number of inspections, and the gray area represents the time blind spot, i.e., the number of days with zero inspections.

[0110] The continuous flight trajectory is segmented according to time intervals or spatial distances. The specific process is as follows: Segmentation based on time threshold: If the time interval between adjacent trajectory points is greater than the preset time threshold of 300 seconds, the trajectory is divided into different segments to represent different inspection tasks; for flight trajectories with a flight time of less than 300 seconds that cannot be segmented based on time threshold, segmentation is based on spatial threshold. Spatial threshold segmentation: If the distance between adjacent trajectory points is greater than the preset spatial distance of 500 meters, the trajectory points are divided into different segments, indicating cross-regional flight; As one implementation method of this embodiment, each trajectory segment is assigned a unique identifier Trajectory_ID, and the start and end times, start and end coordinates, total length and average speed are recorded.

[0111] Furthermore, feature extraction of trajectory points is performed, and a feature vector is constructed for each GPS trajectory point, including spatial coordinates, timestamp, instantaneous velocity, acceleration, heading angle, pitch angle, field of view, number of photos taken, number of detected devices, and ambient light intensity.

[0112] The DBSCAN (Density-Based Spatial Clustering of Applications with Noise) algorithm is used to spatially cluster the trajectory points. It iterates through all trajectory points and calculates the number of trajectory points n within a preset radius r for each point. If the number of identified trajectory points n is greater than a preset threshold, that point is marked as a core point. Furthermore, starting from each core point, the cluster is recursively expanded: all density-reachable points are added to the same cluster, and points that do not belong to any cluster are marked as noise points.

[0113] The output consists of several spatial clusters, each representing a high-frequency inspection area. The center coordinates (centroid), radius (distance from the farthest point to the centroid), number of points, and time span are recorded for each cluster. All equipment entities requiring inspection are retrieved from the knowledge graph to obtain a list of equipment. For each equipment entity, its associated trajectory segments are searched to determine whether it has been observed.

[0114] If a device entity is identified that meets all of the following conditions: The horizontal distance between the trajectory point and the actual equipment is ≤50 meters; The camera's field of view covers the entire physical device. There are photographic records taken within the time period corresponding to this trajectory point; Then the physical device will be included in the high-frequency inspection area; If a device entity is identified that meets all of the following conditions: The horizontal distance between the trajectory point and the actual equipment is greater than 50 meters. The camera's field of view cannot cover any part of the physical device. No photos were taken during the time period corresponding to this trajectory point; The physical device will then be included in the inspection blind spot.

[0115] It should be understood that the terms “comprising” and “including” used in this disclosure and claims indicate the presence of the described features, integrals, steps, operations, elements and / or components, but do not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or collections thereof.

[0116] It should also be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the disclosure. As used in this disclosure and claims, the singular forms “a,” “an,” and “the” are intended to include the plural forms unless the context clearly indicates otherwise. It should also be understood that the term “and / or” as used in this disclosure and claims means any combination and all possible combinations of one or more of the associated listed items, and includes such combinations; The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described above. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. A method for constructing a domain knowledge graph for unmanned aerial vehicle (UAV) inspection, characterized in that, The method comprises the following steps: Step one, multi-source data collection; Collecting unmanned aerial vehicle flight data, multi-modal inspection image data, inspection object basic information, industry knowledge documents and historical inspection records to obtain original data related to unmanned aerial vehicle inspection; Step two, entity recognition and extraction; Defining an entity type ontology according to the original data, extracting entities from different data sources using rule mapping, named entity recognition and computer vision technology, constructing a node set of a knowledge graph, and performing entity disambiguation and linking; Step three, relationship extraction and attribute annotation; Identifying semantic relationships between extracted entities, including composition relationship, causal relationship, time relationship, location relationship and maintenance relationship, and annotating static attributes and dynamic attributes for entities, constructing edges and node attributes of the knowledge graph; Step four, knowledge fusion and alignment; Aligning entities from different data sources, identifying the same real object based on identifier matching or comprehensive similarity calculation, and fusing and resolving conflicts of relationships to obtain a fused knowledge graph; Step five, knowledge graph storage and indexing; Storing the fused knowledge graph into a graph database, establishing a multi-level indexing structure, and supporting efficient query, reasoning and incremental update; Step six, spatio-temporal clustering analysis and inspection blind area identification; Performing spatio-temporal density clustering on historical inspection trajectories, identifying high-frequency inspection areas and inspection blind areas, and adding the analysis results to the knowledge graph in the form of entities.

2. The method according to claim 1, wherein: In step one, the unmanned aerial vehicle flight data includes spatial positioning coordinates, flight height, flight speed, attitude angle, battery capacity and wireless signal strength; The multi-modal inspection image data is collected by visible light sensors, infrared thermal imaging sensors and multi-spectral sensors, and the image data includes metadata information such as shooting time, geographic location and sensor parameters; The inspection object basic information is obtained from an information management system or a geographic information system, including device model, installation location, commissioning date, rated parameters and historical maintenance records; The historical inspection records include structured or semi-structured inspection report data, including defect description, level assessment, treatment suggestion and closed-loop record.

3. The method according to claim 1, wherein: In step two, the entity type ontology includes device entity, defect entity, location entity, state entity, personnel entity, time entity and spatio-temporal trajectory entity; the spatio-temporal trajectory entity is obtained by spatio-temporal segmentation processing of the unmanned aerial vehicle flight trajectory, records trajectory segment identifier, spatio-temporal range, motion parameters and observation information, and establishes a spatio-temporal association relationship between the trajectory entity and the device entity and the defect entity. 4.The method of claim 1, wherein, In step two, the specific process of entity extraction includes: For structured data, a rule mapping method is used to directly map data table records to entities, and fields correspond to entity attributes; For unstructured text data, a named entity recognition technology is used to identify entity boundaries and types based on a pre-trained language model and a sequence labeling network, and to perform entity normalization. For image data, computer vision techniques are used to identify devices and defect entities in images, including using deep learning object detection models to detect device and defect objects, using image segmentation techniques to obtain pixel-level boundaries of defects, performing temperature analysis on infrared images to identify hot abnormal areas, and performing optical character recognition on device identifiers to extract text information. 5.The method of claim 1, wherein, The specific process of relationship extraction in step three includes: For structured data, relationships are directly extracted from foreign key constraints and associated tables; For unstructured text, rule-based trigger word matching or deep learning-based relationship classification methods are used, taking entity pairs and their context as input, using pre-trained language models to extract semantic features, and predicting relationship types through classification layers; For image data, scene graph generation techniques are used to extract visual relationships; The static attributes are directly extracted from structured data, including device model, rated parameters, manufacturer, and installation date; the dynamic attributes are extracted from sensor data and inspection results, including current state, recent inspection time, and latest defect quantity, with timestamps. 6.The method of claim 1, wherein, The specific process of entity alignment is as follows: Identifier alignment: preferentially use unique identifiers for accurate matching, and merge entities from different data sources with the same identifier into one entity node; Similarity alignment: when there is a lack of unique identifiers, calculate the comprehensive similarity of entities, including name similarity, geographic location similarity, attribute similarity, and relationship structure similarity, and use a weighted fusion method to align and merge entities with a comprehensive similarity exceeding a certain threshold; The relationship fusion includes identifying and removing duplicate or semantically equivalent relationship triples, resolving conflicts between mutually contradictory relationships using temporal analysis or confidence comparison, and completing missing relationships based on ontology rules.

7. The method of claim 1, wherein, The specific process of spatio-temporal clustering analysis and blind area identification is as follows: Trajectory preprocessing: segment continuous flight trajectories into segments based on a pre-set time interval of 300 seconds or a pre-set spatial distance of 500 meters, and extract spatial coordinates, timestamps, motion features, observation features, and environmental features for each trajectory point; For flight trajectories with a flight time less than 300 seconds, which cannot be segmented based on the time threshold, segment them based on the spatial threshold; Spatio-temporal density clustering: use a density-based spatial clustering algorithm to cluster trajectory points, calculate the neighborhood density of each point, mark core points, recursively expand to form clusters, and output multiple spatial clusters as high-frequency inspection areas; Inspection coverage evaluation: query device entities in the knowledge graph to determine whether each device is covered by trajectory observations, calculate device coverage, calculate timeliness scores based on the latest inspection time, and count the number of observations from different directions to calculate observation angle diversity; Inspection blind area identification: grid the inspection area, count trajectory point density to identify spatial blind areas, and perform time series analysis on historical inspection intervals to identify temporal blind areas; Knowledge graph update: add the clustering results, including high-frequency areas, blind areas, and abnormal patterns, to the knowledge graph in the form of entities, establish the association between regions and devices and trajectories, and label risk levels.

8. An unmanned aerial vehicle inspection-oriented domain knowledge graph construction system, configured to implement an unmanned aerial vehicle inspection-oriented domain knowledge graph construction method according to any one of claims 1-7. The system comprises a data collection module, an entity extraction module, a relationship construction module, a knowledge fusion module, a storage query module, and a space-time analysis module. The data collection module is configured to collect unmanned aerial vehicle flight data, multi-modal inspection image data, inspection object basic information, industry knowledge documents, and historical inspection reports. The entity extraction module is connected to the data collection module and is configured to extract and define entity types based on raw data, construct a node set of a knowledge graph, and perform entity disambiguation and linking. The relationship construction module is connected to the entity extraction module and is configured to identify semantic relationships between entities, label attribute information of the entities, and construct edges and node attributes of the knowledge graph. The knowledge fusion module is connected to the relationship construction module and is configured to perform entity alignment, identifier matching, and similarity calculation on entity types to obtain a fused knowledge graph. The storage query module is connected to the knowledge fusion module and is configured to store the fused knowledge graph in a graph database, establish an index, and perform batch updates. The space-time analysis module is connected to the storage query module and is configured to perform space-time density clustering on historical inspection trajectories, identify high-frequency inspection areas and inspection blind areas, and evaluate inspection coverage.

9. The domain knowledge graph construction system for unmanned aerial vehicle inspection according to claim 8, wherein: The entity extraction module comprises a structured data processing unit, a text data processing unit, an image data processing unit, and an entity disambiguation unit. The structured data processing unit is configured to extract entities from database tables, JSON, and XML using a rule mapping method. The text data processing unit is configured to extract entities from text data such as inspection reports and technical documents using named entity recognition technology, including text preprocessing, entity recognition based on a pre-trained language model combined with a sequence labeling network, model training, and entity normalization. The image data processing unit is configured to identify device and defect entities in images using computer vision technology, including object detection, image segmentation, thermal anomaly detection, and optical character recognition. The entity disambiguation unit is configured to perform precise matching based on unique identifiers or fuzzy matching based on comprehensive similarity calculation to ensure that the same entity is uniquely represented in the knowledge graph. 10.The domain knowledge graph construction system for unmanned aerial vehicle inspection-oriented field according to claim 8, characterized in that: The space-time analysis module comprises a trajectory preprocessing unit, a density clustering unit, a coverage evaluation unit, a blind area identification unit, and a knowledge update unit. The trajectory preprocessing unit is configured to segment continuous flight trajectories in space-time and extract trajectory point features. The density clustering unit is configured to use a density-based spatial clustering algorithm to cluster trajectory points and identify high-frequency inspection areas. The coverage evaluation unit is configured to calculate device coverage, timeliness score, and observation angle diversity. The blind area identification unit is configured to identify spatial and temporal blind areas. The knowledge update unit is configured to add space-time analysis results to the knowledge graph in the form of entities, establish associated relationships, and label risk levels.

Images