A power transmission line defect detection method and system based on a knowledge graph

Abstract

This invention provides a knowledge graph-based method and system for detecting defects in power transmission lines. The method includes: combining the computing power of an inspection terminal with the inspection scenario, performing scenario-based trimming and lightweighting of the full defect knowledge graph, and simultaneously embedding matching lightweight inference operators to generate a defect knowledge graph package deployed offline on the inspection terminal; extracting defect-sensitive deep features from the multimodal inspection data, and semantically anchoring these features to ontology nodes within the defect knowledge graph package to simultaneously generate corresponding inspection feature sets; performing two-stage inference on the inspection feature sets using the lightweight inference operators via the inspection terminal to generate initial detection results; and in offline scenarios, performing ontology constraint consistency verification on the initial detection results based on the defect knowledge graph package, simultaneously eliminating false positives, and outputting the corresponding target detection results. This invention effectively improves detection efficiency.

Description

Technical Field

[0001] This invention relates to the field of knowledge graphs, and in particular to a method and system for detecting defects in power transmission lines. Background Technology

[0002] Transmission lines are the core infrastructure for cross-regional energy transmission in power systems. Over 70% of my country's transmission lines are located in remote mountainous areas, uninhabited plateau regions, and other geographically complex areas, making line inspections difficult, defect and hazard identification challenging, and safety management risks high. With the deepening of the digital transformation of the power grid, drones and intelligent inspection robots have become core equipment for line inspection. Knowledge graph technology, with its core advantages of domain knowledge integration and interpretable logical reasoning, has become a core technological direction for intelligent detection and diagnosis of transmission line defects. The development of related methods and systems has significant engineering value for ensuring the safe and stable operation of the power grid.

[0003] Currently, publicly available knowledge graph-based methods and systems for detecting transmission line defects generally adopt a centralized cloud architecture. Their typical technical approach involves: collecting inspection images and operational data, and conducting initial defect checks via edge inspection terminals; transmitting all collected inspection data back to the cloud management platform via the public network; and utilizing a cloud-based knowledge graph of the entire transmission line domain to perform secondary defect verification, level assessment, root cause analysis, and generate remediation plans. Finally, the diagnostic results are distributed to the maintenance terminal, thus realizing the basic application of knowledge graphs in transmission line defect detection scenarios.

[0004] However, the existing methods and systems mentioned above have inherent technical defects that cannot be avoided: their entire process relies entirely on public network data transmission and cloud-based centralized computing power, making them unsuitable for core inspection scenarios without public network coverage, such as remote mountainous areas and uninhabited areas. They cannot complete the entire closed-loop process of offline real-time defect detection, diagnosis, classification, and handling at the inspection site, which not only leads to cumbersome inspection processes and low operational efficiency, but also causes serious delays in handling critical line defects, posing a significant threat to the safe operation of the power grid. Summary of the Invention

[0005] Therefore, the purpose of this invention is to provide a knowledge graph-based method and system for detecting defects in transmission lines, in order to solve the problem that existing technologies cannot complete offline defect detection at the inspection site, resulting in certain detection limitations and reduced detection efficiency.

[0006] The first aspect of the present invention proposes: A knowledge graph-based method for detecting defects in transmission lines, wherein the method includes: Based on the full business data of the transmission lines, a full knowledge graph of the defects of the transmission lines is constructed. Combining the computing power of the inspection terminal and the inspection scenario, the full knowledge graph of defects is subjected to scenario-based trimming and lightweight processing, and matching lightweight inference operators are embedded simultaneously to generate a defect knowledge graph package that is deployed offline on the inspection terminal. In the offline state of the inspection terminal, multimodal inspection data of the inspection site is collected, and the defect-sensitive depth features contained in the multimodal inspection data are extracted simultaneously. The defect-sensitive depth features are semantically anchored with the ontology nodes in the defect knowledge graph package, and the corresponding inspection feature set is generated simultaneously. Based on the lightweight inference operator, the inspection feature set is subjected to two-stage inference through the inspection terminal to generate an initial detection result; In offline scenarios, the initial detection results are subjected to ontology constraint consistency verification based on the defect knowledge graph package, and false judgment results are eliminated simultaneously to output the corresponding target detection results, which contain specific defect types.

[0007] The beneficial effects of this invention are as follows: This technical solution effectively solves the limitation of existing inspection technologies in offline conditions, which cannot complete defect detection on-site, and significantly improves the efficiency of transmission line inspection. By constructing a full knowledge graph of transmission line defects and performing scenario-based trimming and lightweight processing, coupled with embedded lightweight inference operators, a defect knowledge graph package that can be deployed offline is generated, realizing the implementation of offline detection capabilities of the inspection terminal. In offline scenarios, by collecting multimodal inspection data, extracting defect-sensitive deep features and semantically anchoring them with the ontology nodes of the knowledge graph package, and combining two-order inference and ontology constraint consistency verification to eliminate misjudgments, the accuracy of the target detection results is ensured, and specific defect types are output, providing precise basis for inspection personnel to handle defects on-site, further improving the intelligence and efficiency of transmission line inspection.

[0008] Furthermore, the step of constructing a full-scale defect knowledge graph of the transmission line based on the full-scale business data of the transmission line, and combining the computing power of the inspection terminal with the inspection scenario to perform scenario-based trimming and lightweight processing on the full-scale defect knowledge graph, and simultaneously embedding matching lightweight inference operators to generate a defect knowledge graph package deployed offline on the inspection terminal includes: The full amount of business data of the transmission line is integrated to construct a four-dimensional linkage ontology architecture that includes equipment topology ontology, defect ontology, causal evolution ontology and rule ontology. At the same time, the attributes, associated semantic edges and weight coefficients of each ontology node are defined within the four-dimensional linkage ontology architecture to construct the full knowledge graph of the defect. Based on the computing power, storage and offline inference latency requirements of the inspection terminal, the corresponding computing power constraint boundary is determined. Simultaneously, with the defect causal evolution link as the core constraint, irrelevant nodes and irrelevant semantic edges are removed from the full defect knowledge graph. Knowledge distillation and low-rank quantization compression are performed on the remaining nodes and semantic edges to generate a lightweight knowledge graph that adapts to the target scenario and terminal computing power. The target ontology architecture of the lightweight knowledge graph is detected, and the lightweight inference operator is simultaneously subjected to deep binding of inference logic with the target ontology architecture to generate the defect knowledge graph package accordingly.

[0009] Furthermore, the step of performing deep binding of the lightweight inference operator with the target ontology architecture to generate the defect knowledge graph package includes: Extract the corresponding four-dimensional linkage topological features contained in the target ontology architecture, and simultaneously reconstruct the lightweight inference operator based on the four-dimensional linkage topological features to generate the corresponding operator inference graph. Based on the semantic constraint rules and defect causal evolution chain of the target ontology architecture, the operator reasoning graph and the lightweight knowledge graph are integrated and fused to generate the corresponding graph reasoning executor. The graph reasoning executor is subjected to terminal-adaptive lightweight compression and encryption, and a standardized offline deployment interface and version management mechanism are configured simultaneously to generate the corresponding defect knowledge graph package.

[0010] Furthermore, the step of extracting the defect-sensitive deep features contained in the multimodal inspection data, and semantically anchoring the defect-sensitive deep features with the ontology nodes in the defect knowledge graph package to synchronously generate the corresponding inspection feature set includes: Based on the ontology architecture and defect determination rules of the defect knowledge graph package, a corresponding defect representation sensitive weight matrix is ​​constructed according to the multimodal inspection data. Simultaneously, based on the cross-membrane feature extraction network, the defect representation sensitive weight matrix is ​​converted into the corresponding defect sensitive depth features to generate a multimodal defect feature set. The multimodal defect feature set is subjected to progressive semantic anchoring, wherein coarse-grained matching of the ontology domain is first completed, followed by fine-grained precise anchoring of the ontology nodes within the domain, and cross-validation of the anchoring results is performed simultaneously to match the confidence labels corresponding to the verified features. The verified target features are hierarchically packaged according to the confidence level labels to generate the inspection feature set.

[0011] Furthermore, the step of performing hierarchical encapsulation processing on the verified target features based on the confidence labels to generate the inspection feature set includes: Based on the confidence labels corresponding to the target features, and combined with the topological relationships between the target features, spatiotemporal-semantic dual-dimensional aggregation is performed on the verified target features to generate corresponding spatiotemporal-semantic feature aggregation units; Using the ontology node as the encapsulation root node, the spatiotemporal-semantic aggregation unit is encapsulated in a nested manner from the core to the periphery according to the hierarchical dependency relationship of the anchored defect ontology, rule ontology and causal evolution ontology. Simultaneously, based on each confidence label, the corresponding inference trigger threshold and linkage verification rule are embedded for each encapsulation layer to generate the corresponding feature encapsulation body. Each feature encapsulation body is configured with a corresponding spatiotemporal reference identifier, and the integrated encapsulation is completed synchronously according to the spatial topology order of the inspection path to generate the corresponding inspection feature set.

[0012] Furthermore, the step of performing two-stage inference on the inspection feature set based on the lightweight inference operator through the inspection terminal to generate initial detection results includes: Invoke the first-order deterministic inference link that is compatible with the inspection feature set, and simultaneously perform progressive defect rule matching on the inspection feature set through the lightweight inference operator to output the corresponding first-order inference result; Based on the first-order inference result, matching features and unmatched features are detected. A second-order inference trigger command is generated synchronously for the unmatched features. In response to the second-order inference trigger command, the corresponding second-order fuzzy inference link is started to output the corresponding second-order inference result. The first-order inference result and the second-order inference result are subjected to cross-validation with linkage constraints to generate the initial detection result.

[0013] Furthermore, the step of performing a cross-validation of the first-order inference result and the second-order inference result to generate the initial detection result includes: The defect types anchored by the first-order inference results and the second-order inference results are detected respectively, and corresponding exclusive verification rule sets are constructed respectively. Simultaneously, based on the hierarchical dependency relationship of the defect knowledge graph package, conflict judgment thresholds are constructed to form a linkage verification benchmark. Based on the aforementioned linkage verification benchmark, a hierarchical progressive cross-verification is performed on the first-order inference result and the second-order inference result to output the corresponding valid inference result and the misjudgment result. The false positives are removed, and the valid inference results are simultaneously deeply bound to the corresponding anchored full inference source information to complete the structured integration and encapsulation, and the initial detection results are output accordingly.

[0014] The second aspect of the present invention proposes: A knowledge graph-based transmission line defect detection system, wherein the system comprises: The construction module is used to construct a full knowledge graph of defects of the transmission line based on the full business data of the transmission line. Combined with the computing power of the inspection terminal and the inspection scenario, the full knowledge graph of defects is subjected to scenario-based trimming and lightweight processing, and matching lightweight inference operators are embedded simultaneously to generate a defect knowledge graph package that is deployed offline on the inspection terminal. The extraction module is used to collect multimodal inspection data from the inspection site when the inspection terminal is offline, and simultaneously extract the defect-sensitive depth features contained in the multimodal inspection data, so as to semantically anchor the defect-sensitive depth features with the ontology nodes in the defect knowledge graph package and synchronously generate the corresponding inspection feature set. The inference module is used to perform two-stage inference on the inspection feature set through the inspection terminal based on the lightweight inference operator to generate an initial detection result; The output module is used to perform ontology constraint consistency verification on the initial detection results based on the defect knowledge graph package in offline scenarios, and simultaneously remove false judgment results to output the corresponding target detection results, which contain specific defect types.

[0015] Furthermore, the building module is specifically used for: The full amount of business data of the transmission line is integrated to construct a four-dimensional linkage ontology architecture that includes equipment topology ontology, defect ontology, causal evolution ontology and rule ontology. At the same time, the attributes, associated semantic edges and weight coefficients of each ontology node are defined within the four-dimensional linkage ontology architecture to construct the full knowledge graph of the defect. Based on the computing power, storage and offline inference latency requirements of the inspection terminal, the corresponding computing power constraint boundary is determined. Simultaneously, with the defect causal evolution link as the core constraint, irrelevant nodes and irrelevant semantic edges are removed from the full defect knowledge graph. Knowledge distillation and low-rank quantization compression are performed on the remaining nodes and semantic edges to generate a lightweight knowledge graph that adapts to the target scenario and terminal computing power. The target ontology architecture of the lightweight knowledge graph is detected, and the lightweight inference operator is simultaneously subjected to deep binding of inference logic with the target ontology architecture to generate the defect knowledge graph package accordingly.

[0016] Furthermore, the building module is specifically used for: Extract the corresponding four-dimensional linkage topological features contained in the target ontology architecture, and simultaneously reconstruct the lightweight inference operator based on the four-dimensional linkage topological features to generate the corresponding operator inference graph. Based on the semantic constraint rules and defect causal evolution chain of the target ontology architecture, the operator reasoning graph and the lightweight knowledge graph are integrated and fused to generate the corresponding graph reasoning executor. The graph reasoning executor is subjected to terminal-adaptive lightweight compression and encryption, and a standardized offline deployment interface and version management mechanism are configured simultaneously to generate the corresponding defect knowledge graph package.

[0017] Furthermore, the extraction module is specifically used for: Based on the ontology architecture and defect determination rules of the defect knowledge graph package, a corresponding defect representation sensitive weight matrix is ​​constructed according to the multimodal inspection data. Simultaneously, based on the cross-membrane feature extraction network, the defect representation sensitive weight matrix is ​​converted into the corresponding defect sensitive depth features to generate a multimodal defect feature set. The multimodal defect feature set is subjected to progressive semantic anchoring, wherein coarse-grained matching of the ontology domain is first completed, followed by fine-grained precise anchoring of the ontology nodes within the domain, and cross-validation of the anchoring results is performed simultaneously to match the confidence labels corresponding to the verified features. The verified target features are hierarchically packaged according to the confidence level labels to generate the inspection feature set.

[0018] Furthermore, the extraction module is specifically used for: Based on the confidence labels corresponding to the target features, and combined with the topological relationships between the target features, spatiotemporal-semantic dual-dimensional aggregation is performed on the verified target features to generate corresponding spatiotemporal-semantic feature aggregation units; Using the ontology node as the encapsulation root node, the spatiotemporal-semantic aggregation unit is encapsulated in a nested manner from the core to the periphery according to the hierarchical dependency relationship of the anchored defect ontology, rule ontology and causal evolution ontology. Simultaneously, based on each confidence label, the corresponding inference trigger threshold and linkage verification rule are embedded for each encapsulation layer to generate the corresponding feature encapsulation body. Each feature encapsulation body is configured with a corresponding spatiotemporal reference identifier, and the integrated encapsulation is completed synchronously according to the spatial topology order of the inspection path to generate the corresponding inspection feature set.

[0019] Furthermore, the reasoning module is specifically used for: Invoke the first-order deterministic inference link that is compatible with the inspection feature set, and simultaneously perform progressive defect rule matching on the inspection feature set through the lightweight inference operator to output the corresponding first-order inference result; Based on the first-order inference result, matching features and unmatched features are detected. A second-order inference trigger command is generated synchronously for the unmatched features. In response to the second-order inference trigger command, the corresponding second-order fuzzy inference link is started to output the corresponding second-order inference result. The first-order inference result and the second-order inference result are subjected to cross-validation with linkage constraints to generate the initial detection result.

[0020] Furthermore, the reasoning module is specifically used for: The defect types anchored by the first-order inference results and the second-order inference results are detected respectively, and corresponding exclusive verification rule sets are constructed respectively. Simultaneously, based on the hierarchical dependency relationship of the defect knowledge graph package, conflict judgment thresholds are constructed to form a linkage verification benchmark. Based on the aforementioned linkage verification benchmark, a hierarchical progressive cross-verification is performed on the first-order inference result and the second-order inference result to output the corresponding valid inference result and the misjudgment result. The false positives are removed, and the valid inference results are simultaneously deeply bound to the corresponding anchored full inference source information to complete the structured integration and encapsulation, and the initial detection results are output accordingly.

[0021] The third aspect of the present invention proposes: A computer includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the knowledge graph-based transmission line defect detection method as described above.

[0022] The fourth aspect of the present invention proposes: A readable storage medium having a computer program stored thereon, wherein the program, when executed by a processor, implements the knowledge graph-based transmission line defect detection method as described above.

[0023] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0024] Figure 1 A flowchart of the knowledge graph-based transmission line defect detection method provided in the first embodiment of the present invention; Figure 2The diagram shows the structure of a knowledge graph-based transmission line defect detection system provided in the third embodiment of the present invention.

[0025] The following detailed description, in conjunction with the accompanying drawings, will further illustrate the present invention. Detailed Implementation

[0026] To facilitate understanding of the present invention, a more complete description will be given below with reference to the accompanying drawings. Several embodiments of the invention are illustrated in the drawings. However, the invention can be implemented in many different forms and is not limited to the embodiments described herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete.

[0027] It should be noted that when a component is said to be "fixed to" another component, it can be directly on the other component or there may be an intervening component. When a component is said to be "connected to" another component, it can be directly connected to the other component or there may be an intervening component. The terms "vertical," "horizontal," "left," "right," and similar expressions used in this document are for illustrative purposes only.

[0028] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains. The terminology used herein in the specification of this invention is for the purpose of describing particular embodiments only and is not intended to be limiting of the invention. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0029] Please see Figure 1 The image shows a knowledge graph-based transmission line defect detection method provided in the first embodiment of the present invention. This knowledge graph-based transmission line defect detection method can collect multimodal inspection data, extract defect-sensitive depth features and semantically anchor them with knowledge graph package ontology nodes, and eliminate misjudgments by combining two-order reasoning and ontology constraint consistency verification. This ensures the accuracy of the target detection results and can output specific defect types, providing accurate basis for inspection personnel to handle defects on-site, and further improving the intelligence and efficiency of transmission line inspection.

[0030] Specifically, this embodiment provides: A knowledge graph-based method for detecting defects in transmission lines, wherein the method includes: Step S10: Based on the full business data of the transmission line, a full knowledge graph of the transmission line defects is constructed. Combining the computing power of the inspection terminal and the inspection scenario, the full knowledge graph of defects is pruned and lightweighted according to the scenario. Matching lightweight inference operators are embedded simultaneously to generate a defect knowledge graph package that is deployed offline on the inspection terminal. It should be noted that traditional transmission line defect detection either uses centralized cloud-based detection, which relies entirely on public network transmission and cannot be adapted to inspection scenarios in remote mountainous areas, Gobi Desert, and other areas without network access; or it uses lightweight local terminal detection models, which can only achieve isolated defect identification of single targets and cannot utilize prior knowledge such as the equipment topology of transmission lines, the causal evolution of defects, and industry judgment rules. This can easily lead to problems such as "identifying insulator self-explosion but missing the accompanying loose strands of the lead wire" and "inability to distinguish between early defects and normal operating conditions." At the same time, the defect judgment rules in industry standards cannot be deeply integrated into the detection process, resulting in detection results that do not conform to operation and maintenance specifications. This step first integrates all business data of transmission lines, including equipment ledgers, design drawings, industry defect judgment standards, historical defect cases, defect evolution patterns, operation and maintenance procedures, and other multi-dimensional data, to construct a full-scale defect knowledge graph covering all business scenarios of transmission lines. This transforms scattered business knowledge into a structured, computable semantic network. Then, considering the computing power, storage, and offline inference latency requirements of inspection terminals (drones, handheld inspection instruments, and line online monitoring devices), the full-scale knowledge graph is pruned and lightweighted according to specific scenarios. Knowledge nodes and related edges irrelevant to the target inspection scenario are removed. At the same time, through knowledge distillation and quantization compression, the size of the graph and the computing power required for inference are significantly reduced. Finally, matching lightweight inference operators are embedded into the lightweight graph, and a defect knowledge graph package that can be directly deployed offline on inspection terminals is generated. This allows the knowledge graph inference capability, which originally could only run in the cloud, to run locally on low-computing-power terminals, fundamentally solving the core contradiction of "incompatibility between detection accuracy and terminal computing power" in offline scenarios, and providing a core carrier and inference foundation for the entire offline detection process.

[0031] Step S20: In the offline state of the inspection terminal, multimodal inspection data of the inspection site is collected, and the defect-sensitive depth features contained in the multimodal inspection data are extracted simultaneously. The defect-sensitive depth features are semantically anchored with the ontology nodes in the defect knowledge graph package, and the corresponding inspection feature set is generated simultaneously. It should be noted that the multimodal inspection data collected offline by the inspection terminal includes various types of data such as visible light images, infrared thermal imaging images, partial discharge audio, and tower attitude sensor data. While this data contains the core features of defects, it also contains a large amount of background noise and environmental interference data unrelated to defects. In this step, while the terminal is offline, the collected multimodal inspection data is first preprocessed. Based on the predefined defect judgment rules in the defect knowledge graph package, sensitive deep features related to defects are extracted, and irrelevant background interference is filtered out, significantly reducing the computational cost of subsequent processing. Then, the extracted sensitive deep features are semantically anchored with the ontology nodes in the defect knowledge graph package. This means semantically matching the extracted visual, infrared, and audio features with ontology nodes defined in the graph, such as "insulator self-explosion," "conductor strand breakage," and "fitting corrosion," establishing a connection between the low-level feature data and the high-level business knowledge. Finally, a standardized inspection feature set with semantically anchored labels is generated, providing standardized input for subsequent reasoning and recognition, achieving a leap from "low-level perceptual features" to "high-level business semantics."

[0032] Step S30: Based on the lightweight inference operator, the inspection feature set is subjected to two-stage inference through the inspection terminal to generate an initial detection result; It's important to note that traditional defect detection inference either employs a single rule matching method, which is inadequate for atypical or early-stage defects, or it uses end-to-end deep learning inference, which suffers from poor interpretability, struggles to integrate with industry rules, and fails to balance inference speed and accuracy in scenarios with limited terminal computing power. This step adopts a two-stage inference approach to adapt to the terminal's computing power and the detection requirements of offline scenarios: The first stage is deterministic inference, which quickly outputs inference results through deterministic rule matching for typical defects with high feature-to-graph node anchoring accuracy and conforming to clear defect judgment rules, balancing inference speed and accuracy; the second stage is fuzzy inference, which performs fuzzy inference based on the defect causal evolution links and topological relationships in the graph for atypical and suspected defect features not matched in the first stage, uncovering early, accompanying, and latent defects and avoiding missed detections. This two-stage inference approach ensures both the detection speed of typical defects and the accurate identification of atypical defects, achieving an optimal balance between detection speed and accuracy under limited terminal computing power, ultimately generating initial detection results containing typical and suspected defects.

[0033] Step S40: In an offline scenario, the initial detection results are subjected to ontology constraint consistency verification based on the defect knowledge graph package, and misjudged results are eliminated simultaneously to output the corresponding target detection results, which contain specific defect types.

[0034] It should be noted that the initial detection results may contain inference conflicts and misjudgments. For example, a normal anti-pollution flashover coating on an insulator might be misjudged as corrosion, or a normal anti-vibration hammer on a conductor might be misjudged as loose hardware. These misjudgments do not conform to the ontological constraints and industry judgment standards of transmission lines. In this offline scenario, based on the ontological architecture, semantic constraint rules, and industry judgment standards in the defect knowledge graph package, this step performs an ontological constraint consistency check on the initial detection results. This verifies whether the detection results conform to equipment topology relationships, defect evolution patterns, and industry judgment rules. For example, if the detection result is "broken conductor strand," but the ontological node at that location in the graph is "ground wire," it is judged as a misjudgment and discarded. Only results that pass the consistency check are determined as the final target detection result, and the specific defect type, corresponding equipment, severity level, and judgment basis are clearly marked. This step sets a final safety checkpoint for the detection results, significantly reducing the false detection rate and ensuring that the output detection results fully comply with the industry standards and business logic of transmission line operation and maintenance. This achieves high-precision and high-reliability detection of transmission line defects in offline scenarios.

[0035] Second Embodiment Furthermore, the step of constructing a full-scale defect knowledge graph of the transmission line based on the full-scale business data of the transmission line, and combining the computing power of the inspection terminal with the inspection scenario to perform scenario-based trimming and lightweight processing on the full-scale defect knowledge graph, and simultaneously embedding matching lightweight inference operators to generate a defect knowledge graph package deployed offline on the inspection terminal includes: The full amount of business data of the transmission line is integrated to construct a four-dimensional linkage ontology architecture that includes equipment topology ontology, defect ontology, causal evolution ontology and rule ontology. At the same time, the attributes, associated semantic edges and weight coefficients of each ontology node are defined within the four-dimensional linkage ontology architecture to construct the full knowledge graph of the defect. Based on the computing power, storage and offline inference latency requirements of the inspection terminal, the corresponding computing power constraint boundary is determined. Simultaneously, with the defect causal evolution link as the core constraint, irrelevant nodes and irrelevant semantic edges are removed from the full defect knowledge graph. Knowledge distillation and low-rank quantization compression are performed on the remaining nodes and semantic edges to generate a lightweight knowledge graph that adapts to the target scenario and terminal computing power. The target ontology architecture of the lightweight knowledge graph is detected, and the lightweight inference operator is simultaneously subjected to deep binding of inference logic with the target ontology architecture to generate the defect knowledge graph package accordingly.

[0036] It should be noted that traditional transmission line knowledge graphs mostly only cover simple associations between equipment ledgers and defect types, failing to form a complete business logic loop and thus unable to support the entire process of defect detection reasoning. This step integrates all business data of transmission lines, including power grid equipment ledgers, line design drawings, DL / T series transmission line defect judgment industry standards, historical defect operation and maintenance cases, line fault analysis reports, defect evolution mechanism research results, on-site operation and maintenance procedures, and other full-dimensional, full-lifecycle business data. It constructs a four-dimensional interconnected ontology architecture of "equipment topology ontology - defect ontology - causal evolution ontology - rule ontology," which is also the core framework of the entire knowledge graph. The equipment topology ontology defines the full-level spatial topological relationships of the transmission line from towers, foundations, conductors, insulators, hardware, and grounding devices, clarifying the physical connections and hierarchical dependencies of "towers - insulator strings - hardware - conductors," which is the basis for defect spatial localization. The defect ontology defines the types, classifications, and visual / infrared / ... Audio characteristics, hazard levels, and location of occurrence cover all common and rare defect types in line operation and maintenance, serving as the core benchmark for defect identification. The causal evolution ontology defines the causal relationships and evolutionary patterns between defects, such as the evolutionary link of "fitting corrosion → bolt loosening → increased conductor vibration → conductor strand breakage," and the accompanying defect association of "insulator spontaneous explosion → uneven stress on the lead wire → lead wire strand loosening," which is the core basis for discovering early and accompanying defects. The rule ontology defines national and industry defect judgment standards, severity level classification rules, and operation and maintenance handling rules, serving as the compliance benchmark for defect judgment. The four-dimensional ontology is deeply linked through semantic edges. For example, nodes in the defect ontology are simultaneously associated with corresponding device nodes in the equipment topology ontology, upstream and downstream defect nodes in the causal evolution ontology, and corresponding judgment rule nodes in the rule ontology. Detailed attributes, semantic edge association types, and weight coefficients are defined for each ontology node. Ultimately, a comprehensive defect knowledge graph covering all services, links, and rules of transmission lines is constructed, realizing the structuring and computability of dispersed business knowledge and providing complete knowledge support for subsequent defect detection reasoning.

[0037] The full knowledge graph covers all business scenarios of power transmission lines. It has a large node scale and complex relationships, which requires high computing power and storage. It cannot be directly deployed on inspection terminals with limited computing power and storage. At the same time, different inspection scenarios (such as drone line channel inspection, tower fine inspection, and grounding device inspection) require different knowledge nodes, and the full graph contains a lot of irrelevant and redundant information. This step first determines strict computing power constraints based on the hardware parameters of the target inspection terminal, including CPU computing power, memory size, storage capacity, and maximum allowable offline inference latency. Then, using the defect causal evolution chain as the core constraint, it is necessary to retain core nodes and semantic edges related to defect evolution, accompanying defects, and topological associations relevant to the target inspection scenario. Based on this, ontology nodes and semantic edges irrelevant to the target inspection scenario are removed; for example, for the refined inspection scenario of insulators, irrelevant nodes related to grounding devices and tower foundations are removed. Subsequently, knowledge distillation and low-rank quantization compression are performed on the retained core nodes and semantic edges. Knowledge distillation transfers the semantic association capabilities of the entire graph to a small-scale lightweight graph, while low-rank quantization compresses the high-dimensional semantic vectors of the graph into low-dimensional vectors. Without sacrificing core inference capabilities, this significantly reduces the graph's size, computing power overhead, and latency, ultimately generating a lightweight knowledge graph fully adapted to the target inspection scenario and the terminal's computing power, solving the core bottleneck of knowledge graph terminal deployment.

[0038] The lightweight knowledge graph only provides the knowledge base. To achieve offline reasoning on the terminal, corresponding reasoning operators need to be matched. The traditional approach is to deploy the graph and reasoning operators separately, which requires frequent calls to graph data during the reasoning process, resulting in low reasoning efficiency and high resource consumption. This step first detects the target ontology architecture of the lightweight knowledge graph, clarifying the node distribution, relationships, and reasoning logic of the four-dimensional ontology. Then, the pre-optimized lightweight reasoning operators are deeply bound to the target ontology architecture for reasoning logic, ensuring that the reasoning logic of the operators is fully adapted to the ontology structure and semantic relationships of the graph. For example, topology association reasoning operators are matched for device topology ontology, defect evolution reasoning operators are matched for causal evolution ontology, and rule matching reasoning operators are matched for rule ontology, making the graph and reasoning operators an inseparable whole, rather than a simple superposition. Finally, the bound lightweight graph and reasoning operators are packaged to generate a defect knowledge graph package that can be directly deployed offline. At the same time, the adaptation to the terminal operating environment is completed, ensuring that the graph package can run directly on the local system of the inspection terminal without relying on cloud resources.

[0039] Furthermore, the step of performing deep binding of the lightweight inference operator with the target ontology architecture to generate the defect knowledge graph package includes: Extract the corresponding four-dimensional linkage topological features contained in the target ontology architecture, and simultaneously reconstruct the lightweight inference operator based on the four-dimensional linkage topological features to generate the corresponding operator inference graph. Based on the semantic constraint rules and defect causal evolution chain of the target ontology architecture, the operator reasoning graph and the lightweight knowledge graph are integrated and fused to generate the corresponding graph reasoning executor. The graph reasoning executor is subjected to terminal-adaptive lightweight compression and encryption, and a standardized offline deployment interface and version management mechanism are configured simultaneously to generate the corresponding defect knowledge graph package.

[0040] It's important to note that the core of lightweight knowledge graphs is a four-dimensional interconnected graph structure. Traditional general-purpose inference operators are linearly executed logics, which cannot adapt to the graph topology of the knowledge graph. Frequent structural transformations are required during inference, leading to low inference efficiency, high resource consumption, and unsuitability for low-computing-power scenarios. This step first extracts the four-dimensional interconnected topological features from the target ontology architecture of the lightweight knowledge graph, including the hierarchical dependencies of ontology nodes, the association types of semantic edges, the link structure of defect causal evolution, and the progressive judgment logic of the rule ontology. Then, based on these graph topological features, the lightweight inference operators undergo graph structure-aware reconstruction. The originally linearly executed operators are reconstructed into an operator inference graph that perfectly matches the graph topology. Each operator node in the operator inference graph corresponds to an ontology node in the graph, and the connections between operator nodes correspond to semantic edges in the graph. This ensures that the operator's inference logic perfectly fits the graph structure, achieving a one-to-one correspondence between "graph structure" and "inference logic," fundamentally improving inference efficiency and reducing terminal computing power overhead.

[0041] After the operator reasoning graph is reconstructed, it needs to be deeply integrated with the lightweight knowledge graph, rather than simply using interface calls, to achieve the most efficient offline reasoning. This step, based on the semantic constraint rules and defect causal evolution chain of the target ontology architecture, deeply binds each operator node in the operator reasoning graph to the corresponding ontology node in the lightweight knowledge graph. The attributes, semantic edge weights, and constraint rules of the ontology node in the graph are directly embedded into the execution logic of the corresponding operator node. This allows the operator to perform reasoning without needing to call additional stored data from the graph, directly completing the reasoning based on the embedded knowledge. Simultaneously, the defect causal evolution chain in the graph is directly transformed into the execution path of the operator reasoning graph, ensuring that the reasoning process fully follows the defect evolution logic and rule constraints defined in the graph. Ultimately, the operator reasoning graph and the lightweight knowledge graph are fully integrated into a unified graph reasoning execution entity. This entity contains complete business knowledge and reasoning logic, capable of independently completing the entire process from feature input to defect reasoning without relying on external data and resources. This significantly improves reasoning efficiency and is fully adaptable to offline operation scenarios on the terminal.

[0042] After the graph inference executor is generated, it needs to undergo terminal adaptation, security encryption, and deployment configuration to become a deployable package. This step first targets the operating system and hardware architecture of the target inspection terminal, performing targeted lightweight compression and compilation optimization on the graph inference executor to further reduce its size and resource consumption, ensuring smooth operation in the terminal's hardware environment. Simultaneously, the graph inference executor is encrypted using national cryptographic algorithms to prevent the theft or tampering of core transmission line business knowledge and defect judgment rules, ensuring data and model security. Then, standardized offline deployment interfaces are configured for the encrypted executor, including data input interfaces, result output interfaces, and operation and maintenance configuration interfaces, enabling seamless integration with the inspection terminal's data acquisition hardware and display system. A robust version management mechanism is also configured to support offline updates, version rollbacks, and scenario switching of the graph package, adapting to the needs of different inspection scenarios. Finally, the executor, having completed all adaptations and configurations, is packaged into a standardized defect knowledge graph package, which can be directly deployed to the inspection terminal with one click, enabling defect detection and inference in offline scenarios.

[0043] Furthermore, the step of extracting the defect-sensitive deep features contained in the multimodal inspection data, and semantically anchoring the defect-sensitive deep features with the ontology nodes in the defect knowledge graph package to synchronously generate the corresponding inspection feature set includes: Based on the ontology architecture and defect determination rules of the defect knowledge graph package, a corresponding defect representation sensitive weight matrix is ​​constructed according to the multimodal inspection data. Simultaneously, based on the cross-membrane feature extraction network, the defect representation sensitive weight matrix is ​​converted into the corresponding defect sensitive depth features to generate a multimodal defect feature set. The multimodal defect feature set is subjected to progressive semantic anchoring, wherein coarse-grained matching of the ontology domain is first completed, followed by fine-grained precise anchoring of the ontology nodes within the domain, and cross-validation of the anchoring results is performed simultaneously to match the confidence labels corresponding to the verified features. The verified target features are hierarchically packaged according to the confidence level labels to generate the inspection feature set.

[0044] It's important to note that traditional multimodal feature extraction typically employs generic feature extraction networks to indiscriminately extract features from image, infrared, and audio data. This results in the extraction of numerous background and environmental features irrelevant to defects, consuming limited computing power and interfering with subsequent defect identification. Furthermore, the lack of a unified defect orientation across different modalities prevents the formation of a complementary feature system. This step first identifies the most sensitive feature dimensions for defect identification in different modal data based on the ontology architecture and defect judgment rules in the defect knowledge graph package. Examples include the glass breakage texture of insulators in visible light images, the abnormal temperature rise area of ​​joints in infrared images, and specific frequency bands of partial discharge in audio data. A defect representation sensitivity weight matrix is ​​constructed for these sensitive dimensions, assigning high weights to defect-sensitive feature dimensions and low or even zero weights to irrelevant background dimensions. Then, based on a cross-modal feature extraction network, the defect representation sensitivity weight matrix is ​​embedded into the feature extraction process, guiding the network to focus on extracting high-weight defect-sensitive features and filtering out low-weight irrelevant interference features. This simultaneously achieves feature alignment and complementarity across different modal data such as visible light, infrared, and audio, ultimately generating a multimodal defect feature set. This defect-sensitive feature extraction method significantly improves the defect identification rate of features, reduces the computational cost of irrelevant features, is fully adaptable to low-computing scenarios on terminals, and provides a high-quality feature foundation for subsequent semantic anchoring.

[0045] Traditional feature anchoring to knowledge graphs typically employs direct feature similarity matching without a hierarchical matching logic, making it prone to anchoring errors and low matching accuracy. For example, a defect feature of a conductor might be incorrectly anchored to an ontology node of a hardware component. This step adopts a progressive semantic anchoring logic, completing the matching in two levels: The first level is coarse-grained ontology domain matching. The multimodal defect feature set is first coarsely matched against the four ontology domains in the defect knowledge graph package to determine the ontology domain corresponding to the feature, such as the equipment topology domain, defect ontology domain, rule ontology domain, or causal evolution domain. Then, the corresponding subdomain is further determined, such as the "insulator defect subdomain" or "conductor defect subdomain" under the defect ontology domain, significantly narrowing the matching range and avoiding cross-domain anchoring errors. The second level is fine-grained precise anchoring of ontology nodes within the domain. Within the subdomains determined by coarse-grained matching, the feature is finely matched against specific ontology nodes within the domain to precisely anchor to the corresponding specific defect node or equipment node, such as "insulator spontaneous explosion" or "glass insulator degradation." After anchoring is complete, the anchoring results are cross-validated based on the ontology association rules of the graph. For example, it checks whether the anchored defect node matches the topological relationship of the corresponding device node. If the validation passes, a confidence label is assigned to the feature. The confidence level represents the degree of matching between the feature and the ontology node, providing a triggering basis for subsequent two-stage inference. This progressive anchoring method significantly improves the accuracy of semantic anchoring, avoids cross-domain matching errors, and enables hierarchical management of features through confidence labels.

[0046] After semantic anchoring and confidence matching are completed, the target features need to be standardized and hierarchically encapsulated to provide a standardized and orderly input for subsequent inference. This step involves hierarchically encapsulating the validated target features based on their corresponding confidence labels. For example, features with a confidence level above 90% are classified as high-confidence features, corresponding to deterministic inference of typical defects; features with a confidence level between 60% and 90% are classified as medium-confidence features, corresponding to fuzzy inference of suspected defects; and features with a confidence level below 60% are classified as low-confidence features and filtered out. Simultaneously, during the encapsulation process, corresponding anchoring ontology node information, confidence labels, modality source, spatiotemporal coordinates, and other metadata are embedded into each feature package, ensuring that each feature package possesses complete semantic, spatiotemporal, and confidence information. Finally, all hierarchically encapsulated feature packages are integrated into a standardized inspection feature set, providing a standardized, orderly, and semantically labeled input for subsequent two-stage inference, ensuring the efficiency and accuracy of the inference process.

[0047] Furthermore, the step of performing hierarchical encapsulation processing on the verified target features based on the confidence labels to generate the inspection feature set includes: Based on the confidence labels corresponding to the target features, and combined with the topological relationships between the target features, spatiotemporal-semantic dual-dimensional aggregation is performed on the verified target features to generate corresponding spatiotemporal-semantic feature aggregation units; Using the ontology node as the encapsulation root node, the spatiotemporal-semantic aggregation unit is encapsulated in a nested manner from the core to the periphery according to the hierarchical dependency relationship of the anchored defect ontology, rule ontology and causal evolution ontology. Simultaneously, based on each confidence label, the corresponding inference trigger threshold and linkage verification rule are embedded for each encapsulation layer to generate the corresponding feature encapsulation body. Each feature encapsulation body is configured with a corresponding spatiotemporal reference identifier, and the integrated encapsulation is completed synchronously according to the spatial topology order of the inspection path to generate the corresponding inspection feature set.

[0048] It is important to note that the multimodal data collected during transmission line inspections possesses clear spatiotemporal correlations. For example, inspection data of insulators, hardware, and conductors on the same tower share the same spatial coordinates and timestamps. Furthermore, these features also exhibit semantic topological relationships. Traditional feature encapsulation only encapsulates individual features independently, ignoring the spatiotemporal and semantic relationships between features. This leads to subsequent inference being unable to utilize these relationships, easily resulting in missed detections. This step first aggregates target features at the same time and spatial location based on their corresponding spatiotemporal coordinates, such as aggregating visible light, infrared, and audio features collected from the same tower. Then, based on the topological relationships between the anchored entity nodes, features exhibiting equipment topological correlations and defect causal correlations are semantically aggregated, such as aggregating insulator features with corresponding connected hardware features. Simultaneously, combining feature confidence labels, core and auxiliary features are set for each aggregation unit: high-confidence defect features are core features, and associated equipment features are auxiliary features. The resulting spatiotemporal-semantic feature aggregation unit not only contains the core features of the defect, but also the corresponding spatiotemporal and semantic association information. This allows subsequent reasoning to make full use of the device topology and the associated relationship of the defect, greatly improving the accuracy of defect identification and avoiding missed detections and false detections caused by isolated features.

[0049] Traditional linear feature encapsulation cannot adapt to the hierarchical ontology architecture of knowledge graphs. It requires frequent feature decomposition during inference, resulting in low efficiency. Furthermore, it cannot pre-embed inference trigger rules, necessitating inference type determination during the inference process, increasing unnecessary computational overhead. This step uses the ontology node anchored to the feature as the root node for encapsulation. Following the hierarchical dependencies of the defect ontology, rule ontology, and causal evolution ontology in the graph, it performs nested encapsulation of the spatiotemporal-semantic aggregation unit from the core to the periphery: the innermost core encapsulates the core defect features corresponding to the defect ontology; the middle layer encapsulates the defect judgment rule features corresponding to the rule ontology; and the outermost layer encapsulates the association features and topological features corresponding to the causal evolution ontology, forming a nested encapsulation structure that perfectly matches the graph ontology architecture. Meanwhile, based on the confidence labels corresponding to the features, corresponding inference trigger thresholds and linkage verification rules are embedded for each layer of encapsulation. For example, when the confidence of the core layer features is higher than the trigger threshold, first-order deterministic inference is directly triggered. When the confidence is lower than the threshold, second-order fuzzy inference of the outer layer is triggered. At the same time, the linkage verification rules of the corresponding ontology nodes are embedded, so that the feature encapsulation itself has basic inference triggering and verification capabilities, which greatly improves the efficiency of subsequent inference and eliminates the need for secondary decomposition and judgment of features.

[0050] Transmission line inspections are conducted according to the spatial sequence of the inspection path, such as from tower 1 to tower N. Traditional feature encapsulation does not integrate according to the inspection path, resulting in a disordered sequence of feature packages. This leads to subsequent inference results that cannot correspond to the on-site inspection location, causing inconvenience for operation and maintenance. This step configures a unique spatiotemporal reference identifier for each nested feature package. The identifier contains complete spatiotemporal information such as the inspection time, tower number, line mileage, and spatial GPS coordinates corresponding to the feature, ensuring that each feature package can accurately correspond to the specific location on the inspection site. Then, according to the spatial topological order of the on-site inspection path, that is, the order of tower numbers and line mileage, all feature packages are integrated and encapsulated in one unit, ultimately generating a standardized, ordered inspection feature set with complete spatiotemporal and semantic information. This inspection feature set is fully adapted to the operation sequence on the inspection site, perfectly matches the ontological architecture of the map, and has built-in inference trigger rules, which can be directly input into the two-level inference system for efficient processing, providing high-quality and standardized input for subsequent defect inference.

[0051] Furthermore, the step of performing two-stage inference on the inspection feature set based on the lightweight inference operator through the inspection terminal to generate initial detection results includes: Invoke the first-order deterministic inference link that is compatible with the inspection feature set, and simultaneously perform progressive defect rule matching on the inspection feature set through the lightweight inference operator to output the corresponding first-order inference result; Based on the first-order inference result, matching features and unmatched features are detected. A second-order inference trigger command is generated synchronously for the unmatched features. In response to the second-order inference trigger command, the corresponding second-order fuzzy inference link is started to output the corresponding second-order inference result. The first-order inference result and the second-order inference result are subjected to cross-validation with linkage constraints to generate the initial detection result.

[0052] It's important to note that the core objective of first-order deterministic reasoning is to quickly and accurately identify typical and well-defined defects. These defects have a high degree of matching and confidence with the defect judgment rules in the defect knowledge graph, eliminating the need for complex fuzzy reasoning and enabling rapid output of results, significantly improving detection efficiency. This step first retrieves a first-order deterministic reasoning link adapted to the feature set based on the ontology anchoring results of the inspection feature set. This link is constructed based on the rule ontology in the defect knowledge graph package and fully aligns with the defect judgment standards of the transmission line industry. Then, a lightweight reasoning operator performs progressive defect rule matching on the inspection feature set. First, it matches the major defect category rules, such as "insulator defect" and "conductor defect," then matches specific defect type rules, such as "insulator spontaneous explosion" and "conductor strand breakage," and finally matches the defect severity level rules, such as "general defect," "serious defect," and "critical defect." The entire matching process strictly follows the industry-standard progressive judgment logic. For features with high confidence and a perfect match to the rules, the first-order inference results, including the corresponding defect type, severity level, and judgment criteria, are directly output. For features that do not match the rules or have insufficient confidence, they are marked as unmatched features and proceed to the subsequent second-order inference stage. First-order inference can quickly identify the vast majority of typical defects. Over 80% of routine inspection scenarios can be detected using first-order inference, significantly reducing the computational overhead of the terminal and improving detection efficiency.

[0053] In first-order inference, unmatched features are mostly atypical early defects, latent defects, and accompanying defects. These defects lack clear, complete matching rules and cannot be identified through deterministic rule matching. They are also the parts most easily missed by traditional detection methods, such as early corrosion of hardware, slight wear of conductors, and low-value degradation of insulators. If these early defects are not identified in time, they will gradually evolve into serious line faults. This step first analyzes the first-order inference results, separating the matched features from the unmatched features, and automatically generates a second-order inference trigger command for the unmatched features. In response to the trigger command, the corresponding second-order fuzzy inference link is automatically started. This link is built based on the device topology ontology and causal evolution ontology in the defect knowledge graph package. The core is to use the topological association relationship and causal evolution law of defects for fuzzy inference. For example, by identifying the abnormal corrosion characteristics of mismatched hardware, and combining this with the causal evolution chain of "hardware corrosion → bolt loosening → conductor wear" in the atlas, as well as the corresponding conductor characteristics, it can be deduced that the defect is an early-stage corrosion defect of the hardware, while simultaneously predicting subsequent evolution risks. Another example is the detection of insulator spontaneous explosion characteristics. Through second-order inference, combined with equipment topology relationships, secondary inference is performed on the corresponding drain wire characteristics to uncover accompanying drain wire loosening defects. Second-order fuzzy inference can uncover early defects, latent defects, and accompanying defects that first-order inference cannot identify, significantly reducing the missed detection rate and achieving full-type, full-cycle coverage of defects.

[0054] First-order and second-order inference results may conflict or overlap. For example, a first-order inference might classify a piece of hardware as normal, while a second-order inference might classify it as early corrosion. Cross-validation is necessary to generate a reliable initial detection result. This step performs a linked constraint cross-validation of the first-order and second-order inference results. Based on the ontology constraint rules of the defect knowledge graph, the consistency of the two results is verified. For duplicate identifications of the same defect, they are merged into a single detection result, retaining the judgment with higher confidence. For conflicting results, a secondary validation is performed based on the graph's rule ontology and causal evolution chain to determine the final judgment. Valid results without conflict or overlap are structurally integrated. Finally, all valid defect identification results after cross-validation are integrated into a structured initial detection result, containing complete information such as the type, severity level, corresponding device, spatial location, judgment basis, and inference tracing chain for each defect, providing complete input for subsequent ontology constraint consistency verification.

[0055] Furthermore, the step of performing a cross-validation of the first-order inference result and the second-order inference result to generate the initial detection result includes: The defect types anchored by the first-order inference results and the second-order inference results are detected respectively, and corresponding exclusive verification rule sets are constructed respectively. Simultaneously, based on the hierarchical dependency relationship of the defect knowledge graph package, conflict judgment thresholds are constructed to form a linkage verification benchmark. Based on the aforementioned linkage verification benchmark, a hierarchical progressive cross-verification is performed on the first-order inference result and the second-order inference result to output the corresponding valid inference result and the misjudgment result. The false positives are removed, and the valid inference results are simultaneously deeply bound to the corresponding anchored full inference source information to complete the structured integration and encapsulation, and the initial detection results are output accordingly.

[0056] It should be noted that the defect types and reasoning logic of first-order inference results are different from those of second-order inference results, and the corresponding verification rules are also different. Using a unified verification rule will lead to insufficient verification accuracy. At the same time, a clear conflict judgment threshold is needed to objectively judge whether there is a conflict between the two results and the severity of the conflict. This step first analyzes the defect types anchored by the first-order and second-order inference results respectively. For different defect types, corresponding defect judgment rules, severity level classification rules, and compliance verification rules are extracted from the rule ontology of the defect knowledge graph package. A dedicated set of verification rules is constructed for each defect type. For example, for the "insulator spontaneous explosion" defect, a dedicated set of verification rules is constructed that includes spontaneous explosion area judgment, corresponding equipment matching, and association with accompanying defects. For the "fitting corrosion" defect, a dedicated set of verification rules is constructed that includes corrosion area judgment, corrosion level classification, and evolution risk verification. At the same time, based on the ontology hierarchy dependency relationship of the defect knowledge graph package and combined with the severity level of the defect, a corresponding conflict judgment threshold is constructed. For example, the conflict judgment threshold for critical defects is stricter, while the conflict judgment threshold for general defects is relatively lenient. Finally, a linkage verification benchmark of "dedicated rules + conflict threshold" is formed, which provides a clear and accurate judgment basis for subsequent cross-verification.

[0057] Traditional cross-validation is mostly a single-level consistency comparison, which cannot handle different types of conflicts at different levels, and is prone to problems such as "valid results being mistakenly rejected and incorrect results being retained". This step employs a layered, progressive cross-validation process, consisting of three levels: The first level is basic compliance validation, which, based on a dedicated set of validation rules, verifies whether the first-order and second-order inference results meet the basic compliance requirements for defect assessment. This includes checking if the defect type matches the corresponding device and whether it conforms to industry-standard judgment rules. Results that do not meet basic compliance requirements are directly discarded and marked as misjudged. The second level is consistency validation, which verifies the consistency between first-order and second-order results for the same device and location. Results that are completely consistent are directly marked as valid inference results. The third level is conflict resolution validation, which resolves conflicts between first-order and second-order results based on conflict thresholds and the causal evolution chain of the graph. For example, if a first-order result is judged as normal and a second-order result as an early defect, a second judgment is made, combining the causal evolution of the defect, the confidence level of the features, and the complementary information from multimodal data, to determine valid and misjudged results. This layered, progressive cross-validation ensures the complete retention of valid results while accurately eliminating misjudged results, significantly improving the reliability of the initial detection results.

[0058] Traditional defect detection results only output defect type and location, lacking corresponding reasoning and tracing information. This results in poor interpretability, making it impossible for maintenance personnel to understand the criteria for defect judgment or to verify the results. Furthermore, it fails to meet the full-process traceability requirements of power grid operation and maintenance. This step first removes all marked false positives, retaining only validated valid reasoning results. Then, for each valid reasoning result, it binds corresponding anchored full-scale reasoning tracing information, including feature source, semantically anchored ontology nodes, reasoning link, matching rule clauses, confidence changes, and multimodal data support. This ensures that each detection result possesses complete interpretability and traceability. Subsequently, the valid reasoning results bound with tracing information are structured, integrated, and encapsulated according to the spatial order of the inspection path and the severity level of the defect, generating standardized initial detection results. These initial detection results not only have high accuracy and low false positive rates, but each result also has complete tracing information and judgment criteria, fully complying with the compliance requirements of power grid operation and maintenance. This provides high-quality, high-reliability input for subsequent ontology constraint consistency verification and the final target detection result output.

[0059] Please see Figure 2 The third embodiment of the present invention provides: A knowledge graph-based transmission line defect detection system, wherein the system comprises: The construction module is used to construct a full knowledge graph of defects of the transmission line based on the full business data of the transmission line. Combined with the computing power of the inspection terminal and the inspection scenario, the full knowledge graph of defects is subjected to scenario-based trimming and lightweight processing, and matching lightweight inference operators are embedded simultaneously to generate a defect knowledge graph package that is deployed offline on the inspection terminal. The extraction module is used to collect multimodal inspection data from the inspection site when the inspection terminal is offline, and simultaneously extract the defect-sensitive depth features contained in the multimodal inspection data, so as to semantically anchor the defect-sensitive depth features with the ontology nodes in the defect knowledge graph package and synchronously generate the corresponding inspection feature set. The inference module is used to perform two-stage inference on the inspection feature set through the inspection terminal based on the lightweight inference operator to generate an initial detection result; The output module is used to perform ontology constraint consistency verification on the initial detection results based on the defect knowledge graph package in offline scenarios, and simultaneously remove false judgment results to output the corresponding target detection results, which contain specific defect types.

[0060] Furthermore, the building module is specifically used for: The full amount of business data of the transmission line is integrated to construct a four-dimensional linkage ontology architecture that includes equipment topology ontology, defect ontology, causal evolution ontology and rule ontology. At the same time, the attributes, associated semantic edges and weight coefficients of each ontology node are defined within the four-dimensional linkage ontology architecture to construct the full knowledge graph of the defect. Based on the computing power, storage and offline inference latency requirements of the inspection terminal, the corresponding computing power constraint boundary is determined. Simultaneously, with the defect causal evolution link as the core constraint, irrelevant nodes and irrelevant semantic edges are removed from the full defect knowledge graph. Knowledge distillation and low-rank quantization compression are performed on the remaining nodes and semantic edges to generate a lightweight knowledge graph that adapts to the target scenario and terminal computing power. The target ontology architecture of the lightweight knowledge graph is detected, and the lightweight inference operator is simultaneously subjected to deep binding of inference logic with the target ontology architecture to generate the defect knowledge graph package accordingly.

[0061] Furthermore, the building module is specifically used for: Extract the corresponding four-dimensional linkage topological features contained in the target ontology architecture, and simultaneously reconstruct the lightweight inference operator based on the four-dimensional linkage topological features to generate the corresponding operator inference graph. Based on the semantic constraint rules and defect causal evolution chain of the target ontology architecture, the operator reasoning graph and the lightweight knowledge graph are integrated and fused to generate the corresponding graph reasoning executor. The graph reasoning executor is subjected to terminal-adaptive lightweight compression and encryption, and a standardized offline deployment interface and version management mechanism are configured simultaneously to generate the corresponding defect knowledge graph package.

[0062] Furthermore, the extraction module is specifically used for: Based on the ontology architecture and defect determination rules of the defect knowledge graph package, a corresponding defect representation sensitive weight matrix is ​​constructed according to the multimodal inspection data. Simultaneously, based on the cross-membrane feature extraction network, the defect representation sensitive weight matrix is ​​converted into the corresponding defect sensitive depth features to generate a multimodal defect feature set. The multimodal defect feature set is subjected to progressive semantic anchoring, wherein coarse-grained matching of the ontology domain is first completed, followed by fine-grained precise anchoring of the ontology nodes within the domain, and cross-validation of the anchoring results is performed simultaneously to match the confidence labels corresponding to the verified features. The verified target features are hierarchically packaged according to the confidence level labels to generate the inspection feature set.

[0063] Furthermore, the extraction module is specifically used for: Based on the confidence labels corresponding to the target features, and combined with the topological relationships between the target features, spatiotemporal-semantic dual-dimensional aggregation is performed on the verified target features to generate corresponding spatiotemporal-semantic feature aggregation units; Using the ontology node as the encapsulation root node, the spatiotemporal-semantic aggregation unit is encapsulated in a nested manner from the core to the periphery according to the hierarchical dependency relationship of the anchored defect ontology, rule ontology and causal evolution ontology. Simultaneously, based on each confidence label, the corresponding inference trigger threshold and linkage verification rule are embedded for each encapsulation layer to generate the corresponding feature encapsulation body. Each feature encapsulation body is configured with a corresponding spatiotemporal reference identifier, and the integrated encapsulation is completed synchronously according to the spatial topology order of the inspection path to generate the corresponding inspection feature set.

[0064] Furthermore, the reasoning module is specifically used for: Invoke the first-order deterministic inference link that is compatible with the inspection feature set, and simultaneously perform progressive defect rule matching on the inspection feature set through the lightweight inference operator to output the corresponding first-order inference result; Based on the first-order inference result, matching features and unmatched features are detected. A second-order inference trigger command is generated synchronously for the unmatched features. In response to the second-order inference trigger command, the corresponding second-order fuzzy inference link is started to output the corresponding second-order inference result. The first-order inference result and the second-order inference result are subjected to cross-validation with linkage constraints to generate the initial detection result.

[0065] Furthermore, the reasoning module is specifically used for: The defect types anchored by the first-order inference results and the second-order inference results are detected respectively, and corresponding exclusive verification rule sets are constructed respectively. Simultaneously, based on the hierarchical dependency relationship of the defect knowledge graph package, conflict judgment thresholds are constructed to form a linkage verification benchmark. Based on the aforementioned linkage verification benchmark, a hierarchical progressive cross-verification is performed on the first-order inference result and the second-order inference result to output the corresponding valid inference result and the misjudgment result. The false positives are removed, and the valid inference results are simultaneously deeply bound to the corresponding anchored full inference source information to complete the structured integration and encapsulation, and the initial detection results are output accordingly.

[0066] The fourth embodiment of the present invention provides a computer, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the knowledge graph-based transmission line defect detection method as described above.

[0067] The fifth embodiment of the present invention provides a readable storage medium having a computer program stored thereon, wherein the program, when executed by a processor, implements the knowledge graph-based transmission line defect detection method as described above.

[0068] In summary, the knowledge graph-based transmission line defect detection method and system provided by the above embodiments of the present invention can collect multimodal inspection data, extract defect-sensitive depth features and semantically anchor them with knowledge graph package ontology nodes, and eliminate misjudgments by combining two-order reasoning and ontology constraint consistency verification. This ensures the accuracy of target detection results and can output specific defect types, providing precise basis for inspection personnel to handle defects on-site, and further improving the intelligence and efficiency of transmission line inspection.

[0069] It should be noted that the above modules can be functional modules or program modules, and can be implemented through software or hardware. For modules implemented through hardware, the above modules can reside in the same processor; or the above modules can be located in different processors in any combination.

[0070] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0071] More specific examples of computer-readable media (a non-exhaustive list) include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which the program can be printed, because the program can be obtained electronically, for example, by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0072] It should be understood that various parts of the present invention can be implemented in hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented in software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0073] In the description of this specification, references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0074] The embodiments described above are merely illustrative of several implementations of the present invention, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of the invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of the present invention, and these modifications and improvements all fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.

Claims

1. A method for detecting defects in transmission lines based on knowledge graphs, characterized in that, The method includes: Based on the full business data of the transmission lines, a full knowledge graph of the defects of the transmission lines is constructed. Combining the computing power of the inspection terminal and the inspection scenario, the full knowledge graph of defects is subjected to scenario-based trimming and lightweight processing, and matching lightweight inference operators are embedded simultaneously to generate a defect knowledge graph package that is deployed offline on the inspection terminal. In the offline state of the inspection terminal, multimodal inspection data of the inspection site is collected, and the defect-sensitive depth features contained in the multimodal inspection data are extracted simultaneously. The defect-sensitive depth features are semantically anchored with the ontology nodes in the defect knowledge graph package, and the corresponding inspection feature set is generated simultaneously. Based on the lightweight inference operator, the inspection feature set is subjected to two-stage inference through the inspection terminal to generate an initial detection result; In offline scenarios, the initial detection results are subjected to ontology constraint consistency verification based on the defect knowledge graph package, and false judgment results are eliminated simultaneously to output the corresponding target detection results, which contain specific defect types.

2. The knowledge graph-based transmission line defect detection method according to claim 1, characterized in that, The steps of constructing a full-scale defect knowledge graph of the transmission line based on the full-scale business data of the transmission line, and combining the computing power of the inspection terminal with the inspection scenario to perform scenario-based trimming and lightweight processing on the full-scale defect knowledge graph, and simultaneously embedding matching lightweight inference operators to generate a defect knowledge graph package deployed offline on the inspection terminal include: The full amount of business data of the transmission line is integrated to construct a four-dimensional linkage ontology architecture that includes equipment topology ontology, defect ontology, causal evolution ontology and rule ontology. At the same time, the attributes, associated semantic edges and weight coefficients of each ontology node are defined within the four-dimensional linkage ontology architecture to construct the full knowledge graph of the defect. Based on the computing power, storage and offline inference latency requirements of the inspection terminal, the corresponding computing power constraint boundary is determined. Simultaneously, with the defect causal evolution link as the core constraint, irrelevant nodes and irrelevant semantic edges are removed from the full defect knowledge graph. Knowledge distillation and low-rank quantization compression are performed on the remaining nodes and semantic edges to generate a lightweight knowledge graph that adapts to the target scenario and terminal computing power. The target ontology architecture of the lightweight knowledge graph is detected, and the lightweight inference operator is simultaneously subjected to deep binding of inference logic with the target ontology architecture to generate the defect knowledge graph package accordingly.

3. The knowledge graph-based transmission line defect detection method according to claim 2, characterized in that, The step of performing deep binding of the lightweight inference operator with the target ontology architecture to generate the defect knowledge graph package includes: Extract the corresponding four-dimensional linkage topological features contained in the target ontology architecture, and simultaneously reconstruct the lightweight inference operator based on the four-dimensional linkage topological features to generate the corresponding operator inference graph. Based on the semantic constraint rules and defect causal evolution chain of the target ontology architecture, the operator reasoning graph and the lightweight knowledge graph are integrated and fused to generate the corresponding graph reasoning executor. The graph reasoning executor is subjected to terminal-adaptive lightweight compression and encryption, and a standardized offline deployment interface and version management mechanism are configured simultaneously to generate the corresponding defect knowledge graph package.

4. The knowledge graph-based transmission line defect detection method according to claim 1, characterized in that, The step of extracting the defect-sensitive deep features contained in the multimodal inspection data, semantically anchoring the defect-sensitive deep features to the ontology nodes in the defect knowledge graph package, and synchronously generating the corresponding inspection feature set includes: Based on the ontology architecture and defect determination rules of the defect knowledge graph package, a corresponding defect representation sensitive weight matrix is ​​constructed according to the multimodal inspection data. Simultaneously, based on the cross-membrane feature extraction network, the defect representation sensitive weight matrix is ​​converted into the corresponding defect sensitive depth features to generate a multimodal defect feature set. The multimodal defect feature set is subjected to progressive semantic anchoring, wherein coarse-grained matching of the ontology domain is first completed, followed by fine-grained precise anchoring of the ontology nodes within the domain, and cross-validation of the anchoring results is performed simultaneously to match the confidence labels corresponding to the verified features. The verified target features are hierarchically packaged according to the confidence level labels to generate the inspection feature set.

5. The knowledge graph-based transmission line defect detection method according to claim 4, characterized in that, The step of performing hierarchical encapsulation processing on the verified target features according to the confidence labels to generate the inspection feature set includes: Based on the confidence labels corresponding to the target features, and combined with the topological relationships between the target features, spatiotemporal-semantic dual-dimensional aggregation is performed on the verified target features to generate corresponding spatiotemporal-semantic feature aggregation units; Using the ontology node as the encapsulation root node, the spatiotemporal-semantic aggregation unit is encapsulated in a nested manner from the core to the periphery according to the hierarchical dependency relationship of the anchored defect ontology, rule ontology and causal evolution ontology. Simultaneously, based on each confidence label, the corresponding inference trigger threshold and linkage verification rule are embedded for each encapsulation layer to generate the corresponding feature encapsulation body. Each feature encapsulation body is configured with a corresponding spatiotemporal reference identifier, and the integrated encapsulation is completed synchronously according to the spatial topology order of the inspection path to generate the corresponding inspection feature set.

6. The knowledge graph-based transmission line defect detection method according to claim 1, characterized in that, The step of performing two-stage inference on the inspection feature set based on the lightweight inference operator through the inspection terminal to generate initial detection results includes: Invoke the first-order deterministic inference link that is compatible with the inspection feature set, and simultaneously perform progressive defect rule matching on the inspection feature set through the lightweight inference operator to output the corresponding first-order inference result; Based on the first-order inference result, matching features and unmatched features are detected. A second-order inference trigger command is generated synchronously for the unmatched features. In response to the second-order inference trigger command, the corresponding second-order fuzzy inference link is started to output the corresponding second-order inference result. The first-order inference result and the second-order inference result are subjected to cross-validation with linkage constraints to generate the initial detection result.

7. The knowledge graph-based transmission line defect detection method according to claim 6, characterized in that, The step of performing a cross-validation of the first-order inference result and the second-order inference result to generate the initial detection result includes: The defect types anchored by the first-order inference results and the second-order inference results are detected respectively, and corresponding exclusive verification rule sets are constructed respectively. Simultaneously, based on the hierarchical dependency relationship of the defect knowledge graph package, conflict judgment thresholds are constructed to form a linkage verification benchmark. Based on the aforementioned linkage verification benchmark, a hierarchical progressive cross-verification is performed on the first-order inference result and the second-order inference result to output the corresponding valid inference result and the misjudgment result. The false positives are removed, and the valid inference results are simultaneously deeply bound to the corresponding anchored full inference source information to complete the structured integration and encapsulation, and the initial detection results are output accordingly.

8. A knowledge graph-based transmission line defect detection system, characterized in that, The system includes: The construction module is used to construct a full knowledge graph of defects of the transmission line based on the full business data of the transmission line. Combined with the computing power of the inspection terminal and the inspection scenario, the full knowledge graph of defects is subjected to scenario-based trimming and lightweight processing, and matching lightweight inference operators are embedded simultaneously to generate a defect knowledge graph package that is deployed offline on the inspection terminal. The extraction module is used to collect multimodal inspection data from the inspection site when the inspection terminal is offline, and simultaneously extract the defect-sensitive depth features contained in the multimodal inspection data, so as to semantically anchor the defect-sensitive depth features with the ontology nodes in the defect knowledge graph package and synchronously generate the corresponding inspection feature set. The inference module is used to perform two-stage inference on the inspection feature set through the inspection terminal based on the lightweight inference operator to generate an initial detection result; The output module is used to perform ontology constraint consistency verification on the initial detection results based on the defect knowledge graph package in offline scenarios, and simultaneously remove false judgment results to output the corresponding target detection results, which contain specific defect types.

9. A computer comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the knowledge graph-based transmission line defect detection method as described in any one of claims 1 to 7.

10. A readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements the knowledge graph-based transmission line defect detection method as described in any one of claims 1 to 7.

Images