Power distribution network abnormal state intelligent identification method and system based on unmanned aerial vehicle inspection

Abstract

The invention relates to a power distribution network abnormal state intelligent identification method based on unmanned aerial vehicle routing inspection, and the method comprises the following steps: S1, obtaining power distribution network routing inspection related data, including asset main data, historical operation and maintenance data and environment constraints; s2, constructing the digital twinning of the power distribution network based on the power distribution network inspection related data, and obtaining the mapping from the equipment ID to the spatial pose and parameters; s3, generating a risk heat map based on historical defects and environmental risk scoring; s4, constructing an equipment-geography-working condition knowledge graph, and generating inspection targets and priorities according to line sections in combination with mapping from equipment IDs to spatial poses and parameters and a risk heat map to obtain a task package; and S5, performing multi-target route inspection according to the task package, the risk heat map and the mapping from the equipment ID to the spatial pose and the parameter, and intelligently identifying the abnormal state of the power distribution network. The operation reliability of the power distribution network is effectively improved, and the power supply quality is improved.

Description

Technical Field

[0001] This invention relates to the field of drone inspection, and in particular to a method and system for intelligent identification of abnormal states in power distribution networks based on drone inspection. Background Technology

[0002] As the final nerve center of the power system, the power distribution network directly affects the reliability of power supply to end users, and its operation and maintenance face severe challenges. First, my country's overhead power distribution lines exceed 4 million kilometers in length, with hundreds of millions of devices distributed geographically in a highly dispersed manner. Traditional manual inspection methods can only cover 5-8 kilometers per day, resulting in low coverage efficiency and annual inspection costs exceeding 10 billion yuan. Second, abnormal conditions of power distribution equipment are diverse and often hidden, such as early defects like micro-cracks in insulators, loose clamps, and corrosion of hardware. Human visual detection rates are less than 30%, leading to minor issues frequently escalating into major failures. Third, equipment is constantly exposed to the natural environment, facing various meteorological risks such as strong winds, icing, high temperatures, and lightning strikes. Traditional uniform periodic inspections cannot address the varying risks across different regions and seasons. Finally, power distribution network data is severely fragmented. Equipment ledgers, inspection records, defect histories, and meteorological data are scattered across different systems, and valuable expert experience often exists in the form of tacit knowledge, making systematic inheritance and application difficult.

[0003] Existing technological solutions have significant limitations in addressing these pain points. Traditional manual inspections rely on personnel carrying thermometers and ultraviolet imagers for ground patrols, which is inefficient, dangerous, and highly subjective. Simple drone inspections use fixed-route photography followed by manual post-processing of drawings, which improves efficiency but suffers from suboptimal paths, resource waste, and a heavy workload for drawing review. Basic intelligent identification systems are designed for only single types of defects, have weak generalization capabilities, and lack prior support from historical and environmental factors, resulting in a high false alarm rate. Decentralized risk assessment methods are mainly based on static indicators or simple historical statistics, lacking deep integration of environmental factors and equipment characteristics, and cannot form dynamic and accurate risk hotspot predictions. These technological limitations lead to unreasonable allocation of distribution network inspection resources, insufficient coverage of high-risk areas, and waste of resources in low-risk areas, making it impossible to achieve accurate and efficient identification of abnormal states. Summary of the Invention

[0004] To address the aforementioned issues, the present invention aims to provide a method and system for intelligent identification of abnormal states in power distribution networks based on unmanned aerial vehicle (UAV) inspections, which effectively improves the reliability of power distribution network operation and enhances power supply quality.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] A method for intelligent identification of abnormal states in power distribution networks based on unmanned aerial vehicle (UAV) inspection includes the following steps:

[0007] S1: Obtain relevant data for power distribution network inspection, including asset master data, historical operation and maintenance data, and environmental constraints;

[0008] S2: Based on data related to power distribution network inspection, construct a digital twin of the power distribution network and obtain the mapping from device ID to spatial pose and parameters;

[0009] S3: Generate a risk heat map based on historical defects and environmental risk scores;

[0010] S4: Construct a knowledge graph of equipment-geography-operating conditions, and combine it with the mapping of equipment ID to spatial pose and parameters and risk heat map to generate inspection targets and priorities according to line segments to obtain task packages;

[0011] S5: Based on the mapping of task packages, risk heat maps, and equipment IDs to spatial poses and parameters, perform multi-target route inspections and intelligently identify abnormal states of the power distribution network.

[0012] A power distribution network anomaly identification system based on UAV inspection includes a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it specifically performs the steps in the power distribution network anomaly identification method based on UAV inspection as described above.

[0013] The present invention has the following beneficial effects:

[0014] 1. This invention constructs a high-precision digital twin model, providing centimeter-level geometric reference for anomaly identification. Each device ID is mapped to a precise spatial pose, geometric parameters, and ROI region, enabling UAVs to achieve accurate alignment and optimal observation angle selection, significantly improving image acquisition quality and subsequent identification accuracy, and solving the problem of missed detection caused by inaccurate position and poor angle in traditional methods.

[0015] 2. This invention constructs a risk heatmap, integrating historical defect data, environmental factors (wind load, icing, tree obstruction), and topological importance to quantify the risk score for each piece of equipment and line segment. Through refined risk modeling, inspection resources are concentrated in high-risk areas, significantly improving the efficiency of anomaly detection. The risk heatmap employs a spatiotemporal dynamic update mechanism, capable of real-time adjustments based on seasonal changes, meteorological conditions, and newly discovered defects, allowing inspection plans to adaptively optimize with environmental changes. It is particularly sensitive to sudden risk changes under extreme conditions such as strong winds and icing, enabling timely triggering of targeted inspections. The system also provides confidence intervals for risk scores, supporting decision-making based on uncertainty and avoiding misjudgments due to data quality issues.

[0016] 3. This invention constructs a complete knowledge-driven closed-loop system. The equipment-geography-operating condition knowledge graph established in Section 1 organically integrates the physical characteristics, spatial constraints and dynamic environment of the power distribution network, and can express complex conditional reasoning rules, making the inspection strategy more in line with the failure mechanism and operating condition correlation of power equipment. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0018] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments:

[0019] refer to Figure 1 In this embodiment, a method for intelligent identification of abnormal states in a power distribution network based on unmanned aerial vehicle (UAV) inspection is provided, including the following steps:

[0020] S1: Obtain relevant data for power distribution network inspection, including asset master data, historical operation and maintenance data, and environmental constraints;

[0021] S2: Based on data related to power distribution network inspection, construct a digital twin of the power distribution network and obtain the mapping from device ID to spatial pose and parameters;

[0022] S3: Generate a risk heat map based on historical defects and environmental risk scores;

[0023] S4: Construct a knowledge graph of equipment-geography-operating conditions, and combine it with the mapping of equipment ID to spatial pose and parameters and risk heat map to generate inspection targets and priorities according to line segments to obtain task packages;

[0024] S5: Based on the mapping of task packages, risk heat maps, and equipment IDs to spatial poses and parameters, perform multi-target route inspections and intelligently identify abnormal states of the power distribution network.

[0025] In this embodiment, the asset master data includes equipment list, geographic coordinates (towers / lines / substations), structural parameters, and importance level; historical operation and maintenance data includes defect database, historical alarms, and maintenance records; environmental constraints include airspace restrictions, no-fly zones, weather windows, and terrain data (DEM / DSM).

[0026] Specifically, the equipment list (equipment ledger) fields include: equipment ID (globally unique), equipment type (tower / conductor / switch / insulator / substation equipment, etc.), line / substation to which it belongs, commissioning date, manufacturer / model, operation and maintenance unit, status (in operation / out of operation / under maintenance), and inspection strategy (cycle / frequency).

[0027] Geographic coordinates:

[0028] Tower / Position: WGS84 latitude and longitude, elevation (altitude / relative), tower type, span, adjacent tower ID.

[0029] Route: Route geometry (multi-segment broken line / 3D line), phase direction, conductor and ground wire elevation, corridor width.

[0030] Power distribution room / station: location coordinates, building outline, incoming and outgoing line directions, and perimeter buffer zone.

[0031] Structural parameters:

[0032] Towers: Tower height, material, crossarm parameters, insulator string length / quantity, hardware type, foundation type and burial depth. Conductors: Type, cross-section, sag, allowable current carrying capacity, historical tension range. Nameplate parameters and interface standards for typical equipment such as switches, disconnectors, surge arresters, and instrument transformers.

[0033] Importance level:

[0034] The classification method is based on factors such as the power supply range (important users), the area affected by the fault, the density of historical defects, and the recovery cost, and is classified into A / B / C levels.

[0035] Structured representation: Device ID → Importance score (0-100) and level label, to achieve quantitative sorting.

[0036] Preferred historical operation and maintenance data are as follows:

[0037] Defect database fields: Defect ID, Device ID, Defect type (crack / flashover / rust / foreign matter / icing / discharge, etc.), Severity (S1-S3), Discovery method (manual / AI / alarm), Discovery time, Location information, Handling status (unprocessed / in progress / defect eliminated), Image / video / thermal imaging evidence link.

[0038] Historical alarm fields: alarm ID, device ID, alarm code and description, timestamp, duration, acknowledgment / clearing time, alarm level and reason code.

[0039] Maintenance record fields: Work order ID, Equipment ID, Maintenance type (planned / defect elimination / emergency repair), Start / End time, Work content and materials, Retest results, Responsible team, Power outage impact;

[0040] Preferably, environmental constraints include airspace restrictions and no-fly zones, weather windows, and topographic data:

[0041] Airspace restrictions and no-fly zones include permanent no-fly zones, temporary control zones, buffer zones between airports / power plants / military sensitive areas, and maximum flight altitude restrictions;

[0042] Representation method: Polygon / Polyline / Height Volume (3D volume), including validity period and control level; Usage method: Real-time verification of flight path conflicts, automatic avoidance and generation of alternative routes;

[0043] The weather window includes wind speed / gusts, precipitation, visibility, temperature, humidity, lightning warning, and icing / condensation risk index; Usage: Generate flyable windows, limit gimbal angle and speed, and dynamically adjust mission weight and return-to-home strategy;

[0044] Topographic data (DEM / DSM) includes DEM: bare ground elevation; DSM: including features (buildings, trees); usage methods: line-of-sight analysis (LoS), terrain following, minimum safe passage (MSA) calculation, wind corridor / eddy risk assessment.

[0045] In this embodiment, a digital twin of the distribution network is constructed based on data related to distribution network inspection, and the mapping from device ID to spatial pose and parameters is obtained, as follows:

[0046] The distribution network asset ledger and GIS data are integrated and normalized. On the basis of unified data, a multi-level topology skeleton is constructed. Based on the established topology skeleton, parameterized templates for various equipment are constructed to realize the transformation from simplified skeleton to refined three-dimensional model.

[0047] Register the aerial survey point cloud with the topological skeleton, and use actual aerial survey data to correct and optimize the theoretical model to ensure a geometrically accurate match between the digital twin and the physical entity;

[0048] Based on the registered topology skeleton, the precise attitude angles, assembly vectors, installation heights, and normal parameters of each device are calculated, realizing the conversion from geometric expression to engineering parameters;

[0049] Based on voxelization to generate ROI and sensor reachability domain, a digital twin of the power distribution network is constructed to obtain the mapping from device ID to spatial pose and parameters.

[0050] In this embodiment, the distribution network asset ledger and GIS data are integrated and normalized. Based on the unified data, a multi-level topology skeleton is constructed, as follows:

[0051] First, the distribution network asset ledger and GIS data are integrated and normalized. Basic line information is extracted from the power management system, including line number, voltage level, start and end points, tower sequence, and equipment list structured information. All geographic data is converted to a unified coordinate system (usually CGCS2000 or WGS84) to ensure spatial consistency.

[0052] ;

[0053] Where T source→targetP is the coordinate transformation matrix. source and P target These represent the point positions in the source and target coordinate systems, respectively.

[0054] Convert the ellipsoid height to normal height H normal Introducing topographic relief, a elevation datum transformation is performed:

[0055] ;

[0056] Where H ellipsoid For the height of the ellipsoid, N geoid For the geoid difference, ΔH DEM Correction values ​​for the terrain model;

[0057] Based on unified data, construct a multi-level topological graph: define a directed graph G=(V,E);

[0058] The vertex set V contains tower nodes V tower and device node V equip Edge set E contains wire connections to E line Device connection relationship E mount Attribute set A V and A E Includes various parameters for nodes and edges.

[0059] The data is organized according to a four-level hierarchical relationship of lines, towers, equipment, and components, a tree-like inheritance structure is established, and a unique identifier is assigned to each entity.

[0060] Graph theory algorithms are applied to verify the rationality of the topology, ensuring that G is a connected graph with no isolated nodes, and that each tower node v∈V. tower The degree should satisfy:

[0061] deg(v)=N phases ×2+N ground ×2+N equipment .

[0062] In this embodiment, based on the established topology skeleton, parameterized templates for various devices are constructed, as follows:

[0063] Based on the tower model information in the ledger, the towers are divided into angle steel towers and steel pipe towers:

[0064] A parametric model of the steel pipe pole is established, defining the functional relationships of the following key parameters: the cross-sectional radius r(h) as a function of height h.

[0065] ;

[0066] Where H is the rod height, and r base and r topThese are the bottom and top radii, respectively;

[0067] For each crossarm i, define its height h. i Length L i Inclination angle θ i :

[0068] ;

[0069] The parametric model of the angle steel tower uses a truncated quadrangular pyramid to describe the main body, plus the extended crossarms. The equation of the tower's outer contour is as follows:

[0070] ;

[0071] Angle steel units arranged according to rules, angle steel specification D a With spacing S a :

[0072] Define a standardized anchor point coordinate system for each type of tower, with an offset vector relative to the tower center:

[0073] ;

[0074] Among them, R tower Let P be the tower attitude rotation matrix. mount,jlocal This refers to the location of the hanging point in the local coordinate system.

[0075] Insulator string parameterization: A model is established based on the type (suspension / tension) and specifications; suspension string length is calculated.

[0076] ;

[0077] Where, N discs H represents the number of insulator discs. disc For single-chip height, G gap For gaps;

[0078] Insulator string attitude vector calculation:

[0079] ;

[0080] The hardware and connector models are parametrically modeled based on type (tension clamp, suspension clamp, etc.).

[0081] The parametric model of the conductor determines the diameter d, unit mass m, elastic modulus E, and coefficient of linear expansion α based on the conductor type; the catenary equation describes the shape of the conductor between two points.

[0082] ;

[0083] Where parameters T is the horizontal tension, and g is the acceleration due to gravity;

[0084] Establish the relationship between sag variations at different temperatures T:

[0085]

[0086] Where f0 is the sag value at reference temperature T0, L is the span, and α is the coefficient of linear expansion;

[0087] The parametric template described above transforms the model and specification information in the ledger into a precise three-dimensional geometric model, which has the ability to dynamically adjust according to temperature.

[0088] In this embodiment, the aerial survey point cloud is registered with the topological skeleton, and the theoretical model is corrected and optimized using actual aerial survey data to ensure a geometrically accurate match between the digital twin and the physical entity, as detailed below:

[0089] First, aerial survey data from different sensors were merged, including LiDAR point clouds, oblique photography, and video sequence data;

[0090] The raw LiDAR point cloud data is cleaned and standardized, including:

[0091] Noise filtering: Statistical outlier removal, ground point separation

[0092] Point cloud downsampling: Voxel mesh method preserves structural features and reduces redundancy

[0093] Point cloud classification: Preliminary segmentation into categories such as ground, vegetation, buildings, and power facilities.

[0094] Identify corresponding feature points in the ledger skeleton and point cloud, including:

[0095] Center point of tower base and ;

[0096] Obvious ground feature reference points and ;

[0097] Solving rigid body transformations: Find the optimal rotation matrix RR and translation vector tt that minimizes the distance between corresponding point sets.

[0098] ;

[0099] Among them, w i The weights corresponding to the points are proportional to the confidence level.

[0100] Calculate the overall registration error:

[0101] ;

[0102] Ensure Eglobal <τ global (The threshold is usually set to 1-2 meters), τ global The preset threshold;

[0103] Perform independent fine-tuning for each tower and span segment, and adjust the parametric model to best fit the point cloud data:

[0104] Tower model optimization: Adjusting tower position, height, and orientation:

[0105] Tower detection and segmentation: Extracting tower portions P from point clouds tower ;

[0106] Column fitting: Solve for the best fitting parameters and update the tower parameters:

[0107]

[0108] Where θ represents the tower model parameters, and d(p,M) represents the shortest distance from point p to model M;

[0109] Based on point cloud clustering and plane fitting, update the crossarm parameter {h} i ,L i ,θ i};

[0110] For each span of the conductor, fit accurate sag parameters and apply density clustering to separate the conductor point set P from the point cloud. wire ;

[0111] Based on the RANSAC method, fit the sag parameters a and x0:

[0112]

[0113] Where c is the elevation offset constant.

[0114] Spatial clustering is used to distinguish conductors of different phases and loops, and parameter tuning and consistency constraints are applied to ensure that the adjusted model meets physical constraints.

[0115] The conductor length between the two towers should conform to the catenary theory:

[0116] ;

[0117] Matching degree of the connection point between the conductor and the insulator:

[0118] ;

[0119] Based on the refined registration results, the digital twin skeleton is updated, including tower parameter updates and conductor parameter updates. Through this stage of processing, the theoretical skeleton model is transformed into a geometric model that accurately corresponds to the actual environment, including precise tower location, attitude, and conductor sag, laying a solid foundation for subsequent equipment parameter solutions.

[0120] In this embodiment, a digital twin of the power distribution network is constructed based on voxelization to generate the ROI and the sensor reachability region, and the mapping from device ID to spatial pose and parameters is obtained, as follows:

[0121] Convert the power distribution network scenario into a regular voxel mesh:

[0122] Voxel resolution selection: typically 10-30 cm, adjusted according to the required observation accuracy;

[0123] Using octrees to improve storage efficiency:

[0124]

[0125] Occupied state and semantic annotation:

[0126] Binary occupancy representation: This indicates whether a voxel is occupied.

[0127] Probability Occupation: , representing the probability of occupancy

[0128] Device ID mapping: This is associated with the corresponding device identifier;

[0129] Aggregate point cloud points by their respective voxels:

[0130] Where r is the voxel resolution, (x min ,y min ,z min () is the origin of the coordinate system;

[0131] Mesh voxelization: Calculation of the intersection between triangular meshes and voxels

[0132] Attribute propagation: Passing attributes such as device ID, material, and risk level to the corresponding voxel;

[0133] Key Region of Interest (ROI) Definition: Equipment Type ROI Template: Predefined Region of Interest based on Equipment Type:

[0134] Insulator ROI: porcelain disc edge, hardware connection, anti-flashover coating;

[0135] ROI for conductors: tension clamps, suspension clamps, splicing conduits, vibration dampers;

[0136] ROI of hardware: bolted connections, stress-bearing areas, and anti-corrosion coating;

[0137] Describing geometric ROIs using sets of spheres and cylinders:

[0138] ROI primitive ={Sphere(c1,r1),Cylinder(p1,p2,r2)}

[0139] Surface offset: Voxels within a distance δ that the device surface offsets outward.

[0140]

[0141] And define ROIs in layers according to detection priority;

[0142] Level 1 ROI: Critical parts with high failure rates and severe impacts (such as the middle connection of insulator strings).

[0143] Secondary ROI: Standard areas requiring routine testing (such as the surface of insulator porcelain discs).

[0144] Level 3 ROI: Auxiliary observation area (such as the entire connecting hardware).

[0145] Define a priority-weighted set:

[0146]

[0147] Among them, w r This represents the detection priority weight for the ROI region r;

[0148] For each ROI region, analyze visibility by projecting light rays from potential observation points to the ROI and checking for occlusion:

[0149]

[0150] Field of view constraint: The angle between the line connecting the observation point and the center of the ROI and the sensor orientation should be within the field of view.

[0151]

[0152] in, For the sensor orientation, θ FOV For the field of view;

[0153] Resolution constraint: The observation distance should ensure that the required ground sampling distance (GSD) is reached.

[0154] ;

[0155] Where d is the observation distance, s pixel Where f is the sensor pixel size, and f is the focal length;

[0156] Finally, the various parameters obtained in the preceding steps are integrated into a unified mapping structure to establish a mapping relationship from device ID to complete spatial pose and parameter set.

[0157] In this embodiment, a risk heat map is generated based on historical defects and environmental risk scores, as follows: historical defects and environmental risks (wind, icing, tree obstacles) are integrated and modeled to form a spatiotemporal risk heat map RiskHeatmap(x,y,t) that can be used for target selection and path planning. The output includes both equipment-level risk scores and continuous risk fields along the route and spatial grids.

[0158] In this embodiment, preferably, a set of statistical defect events D is generated for each device e. e For each event d∈D e extract:

[0159] Severity score S(d)∈[0,1] (mapped by alarm level), Outage impact / Topology impact I(d)∈[0,1] (based on load and transfer cost), Repair time T rep (d) Repeated defect marker R(d)∈{0,1}; discovery time t(d) and current time t;

[0160] Timeliness weight: ;

[0161] Repetition weighting (encourage attention to "recurrence points"):

[0162] Overall weighting:

[0163] Historical Defect Score:

[0164] ;

[0165] in, Standardized values ​​for repair time; This reflects the trade-off between severity, impact, and difficulty of remediation; ϵ represents the numerical stability term.

[0166] If the equipment has no historical defects, it can be based on the prior mean μ of similar equipment of the same type and age. peer Smooth:

[0167] ;

[0168] Where ρ∈[0,1] is determined by the sample size.

[0169] Environmental risk modeling, including wind, icing, and tree barrier risk models:

[0170] Risk wind

[0171] Input: Wind speed u(t), wind direction θ u(t) Fext, the annual frequency of extreme wind events;

[0172] Span and orientation are more sensitive to crosswinds; define wind direction sensitivity weights. , where Δ θ The angle between the wind direction and the normal of the guide wire;

[0173] Wind load effect index (normalized):

[0174] ;

[0175] Wind risk score:

[0176] ;

[0177] in, To normalize the frequency of extreme winds, This is the terrain magnification factor (wind gain in mountain passes and canyons).

[0178] Icing Risk ice Input: icing thickness h ice (t), duration Dice(t), melting cycle frequency N cycle ;

[0179] Load factor:

[0180] Duration factor:

[0181] ''

[0182] Melting Cycle Effects (Fatigue / Dancing Trigger) Normalization

[0183] Icing Risk Score:

[0184]

[0185] Tree barrier risk veg Input: Minimum distance d from the conductor to the vegetation veg (x,y,t), growth rate r grow pruning cycle T trim ; Safety distance threshold d min Define the infringement limit:

[0186] ;

[0187] Expected encroachment limits (considering growth and pruning):

[0188] ;

[0189] Tree barrier risk score:

[0190]

[0191] in, The time-averaged limit within the window

[0192] By combining the risks of wind, icing, and tree obstruction, a final risk score is obtained:

[0193] ;

[0194] Where ωw+ωi+ωv=1 is the weight coefficient, which can be determined by the Analytic Hierarchy Process (AHP) or Bayesian optimization, or by setting the weight according to the regional climate.

[0195] Consider the criticality of the equipment in the distribution network topology (load disconnection, tie switch substitution):

[0196] Key indicator K(e)∈[0,1]: consists of N-1 availability, transfer path length, and proportion of important users.

[0197] Impact Factor:

[0198]

[0199] in To normalize the cost of power outages.

[0200] Considering historical defects, environmental risks, and topological impacts:

[0201]

[0202] Where α+β+γ=1, to reflect timeliness and data freshness, a data freshness weight w can be added. fresh (e)∈[0.8,1] Adjustment:

[0203] ;

[0204] Uncertainty characterization (e.g., based on Monte Carlo or Bootstrap):

[0205] Give σ Risk (e) or confidence interval Used for risk-reward trade-offs in planning;

[0206] For the risks of devices at both ends and within the span s, use length weights or topology weights:

[0207] ;

[0208] Commonly used weights: weighted weights for key equipment and weighted weights for the central area across the span (sensitive to wind-induced movement).

[0209] A regular grid is constructed in the study area to diffuse equipment risks into a continuous field.

[0210] In this embodiment, a knowledge graph of equipment-geography-operating conditions is constructed, and combined with the mapping from equipment ID to spatial pose and parameters and the risk heat map, inspection targets and priorities are generated according to line segments to obtain task packages, as follows:

[0211] A three-dimensional knowledge graph of equipment, geography, and operating conditions is constructed as the semantic foundation for the intelligent generation of inspection tasks. The graph includes five major entity groups: equipment entities (towers, crossarms, conductors, insulators, clamps, etc.), geographical entities (line segments, terrain units, airspace areas), operating condition entities (wind, icing, tree obstruction status), defect entities (cracks, flashover, corrosion, etc.), and resource entities (sensors, drones, pilot teams). Entities are connected by rich relational edges, including topological relationships (equipment - belongs to - line segment), geographical relationships (line segment - crosses - terrain unit), operating condition relationships (operating condition - acts on - equipment), detection relationships (defect - adapts to - sensor), and risk relationships (equipment - has - risk score).

[0212] After the map is constructed, the system performs multi-source data fusion and injection. Geometric mapping information is injected from the S2 digital twin module to bind each device ID with precise spatial pose (coordinates, attitude angle), geometric parameters, ROI region, and sensor accessibility constraints. Risk scores are injected from the S3 risk heat map module to label the risk value, uncertainty, and risk driving factors (defects / wind / icing / tree obstacles / topology) weights for each device node. At the same time, forecast conditions are obtained from meteorological data sources to establish dynamic environmental constraints.

[0213] Based on the fused knowledge graph, the system scores inspection targets at the equipment and line segment levels. For each equipment entity, a four-dimensional comprehensive score is calculated, including the risk dimension (Risk(e) from S3), the criticality dimension (Criticism(e) – the importance of the equipment in the power grid), the accessibility dimension (Access(e) – considering airspace limitations, terrain obstruction, and sufficiency of observation angles), and the freshness dimension (Freshness(e) – the time interval since the last effective inspection). These scores are then fused into a comprehensive equipment-level score using weighted coefficients.

[0214] ;

[0215] Among them, w R wC w A w F These are the weighting coefficients;

[0216] Based on equipment scoring, segment-level features are aggregated and incorporated. For each segment (span), internal equipment scores are first weighted and aggregated, with key components (such as multi-circuit corner towers) receiving higher weights. Then, the overall risk score of the segment (from S3) is merged. Finally, environmental triggers (such as strong wind orange warnings, icing thresholds, and tree obstruction prediction exceeding thresholds) are added. The segment scoring model is as follows: The system scores and sorts all line segments, and selects the highest priority equipment within each high-scoring segment to form a hierarchical inspection target list. For scenario-based inspections (such as post-strong wind special inspections or tree obstacle special inspections), the system activates the corresponding knowledge rule subsets and adjusts the scoring weights to ensure that specific targets receive higher priority.

[0217] Based on target selection and prioritization, abstract targets are visualized as executable task packages. First, the system extracts the observation requirements for each target from the knowledge graph, including applicable sensor type (RGB / IR / UV / LiDAR), resolution requirements (GSD / NETD), suggested number of shots and overlap rate, as well as geometric constraints (distance range, angle sector). For each target, the system analyzes its ROI characteristics and defect-sensor compatibility relationship, intelligently recommending the best observation scheme. For example, high-resolution RGB imaging is recommended for insulator crack detection, and infrared scanning is recommended for poor contact faults. The observation requirements are superimposed with spatial constraints, weather conditions, and operating time periods to form the spatiotemporal execution constraints for each target.

[0218] The system generates task packages in a standard format, comprising four main parts: basic metadata (route / segment ID, generation time, and validity period), a target list (equipment targets grouped by route segment, priorities, observation requirements, and geometric constraints), time windows and compliance information (airspace control, weather permit windows, and operational periods), and sensor and platform recommendations (UAV model, payload combination, and battery planning). The system applies a two-tiered "must-reach / tailorable" strategy to mark targets: high-risk and high-critical targets are marked "must-reach," while medium- to low-priority targets are marked "tailorable." Before task package generation, the system performs consistency and safety reviews, verifying feasible pose paths between adjacent targets, checking for overlaps with no-fly zones, and splitting into multiple sub-task packages if necessary. The final output task package includes structured data (for direct consumption by S5 route planning), visual representations (for manual review), and explanatory appendices (decomposition of risk drivers and selection criteria).

[0219] In this embodiment, multi-target route inspection is performed based on the mapping from task packages, risk heat maps, and device IDs to spatial poses and parameters to intelligently identify abnormal states in the power distribution network, as detailed below:

[0220] Based on the task packages, risk heat maps, and device ID-spatial pose mapping provided by S4, a complete set of observation poses is first constructed. Starting from the geometric and sensor constraints of each inspection target, the set of observation nodes that meet the requirements of line-of-sight accessibility, resolution, and safety interval is calculated. For high-risk equipment (such as insulator strings with frequent historical defects or conductor segments that are prone to breakage during icing periods), the system generates redundant observation poses with multiple angles and distances to ensure the detection rate. At the same time, the priorities in the task packages are converted into node reward weights, with higher priority targets receiving higher weights.

[0221] Based on the observed pose set, the inspection task is modeled as an Orienteering problem (or TSP-TW extension) with time windows and profit maximization. Constraints include UAV endurance, airspace restrictions, time windows, and safety intervals. The optimization objective is to maximize risk coverage and information gain within battery capacity constraints. The solution employs a combination of initial greedy construction and local optimization, prioritizing high-risk route segments and ensuring optimal observation angles. For multi-UAV collaborative tasks, the system decomposes tasks based on geographical partitioning and risk balance principles, ensuring that core high-risk objectives are achieved while reserving 10-30% energy redundancy to handle unforeseen circumstances.

[0222] During the flight path execution phase, the system combines precise positioning with a digital twin model to achieve centimeter-level target alignment. At each observation node, the UAV hovers stably and simultaneously collects multimodal data according to the sensor configuration (visible light / infrared / ultraviolet / LiDAR). For each shot, the system evaluates the image quality (sharpness, exposure, contrast) and geometric parameters (GSD, angle of view) in real time. If the image quality falls below the set standard, a retake is immediately triggered or an alternative observation pose is found.

[0223] The system implements a closed-loop adaptive mechanism, continuously monitoring changes in environmental conditions (wind speed, rainfall, sunlight), power consumption deviations, and communication quality, triggering rolling replanning when necessary. Specifically, when the online identification algorithm detects high-confidence anomalies, the system automatically generates additional data acquisition tasks with the best confirmation perspective, selecting supplementary observation locations based on the principle of maximizing information entropy gain. For environmental risks such as tree obstructions, the system dynamically adjusts the observation distance and angle to obtain optimal measurement accuracy. In extreme weather, abnormal power levels, or degraded communication quality, the system automatically adjusts to a conservative strategy according to preset rules to ensure equipment and data safety, prioritizing the completion of high-value targets before safely returning to base.

[0224] After the inspection data is collected, the system performs intelligent multimodal anomaly identification. First, it precisely aligns the image with the device's ROI using camera calibration parameters and pose information, and extracts the region of interest. The aligned data is then processed separately using specialized algorithms: visible light images are used to detect cracks, flashover marks, and mechanical damage; infrared images are used to detect abnormal temperature rise and poor contact; ultraviolet images are used to detect corona discharge; and laser point clouds are used to measure conductor sag and tree obstruction. The system employs a Bayesian fusion framework to integrate multimodal results, effectively reducing the false alarm rate of a single sensor.

[0225] The identification results are organized by device ID, and each record includes defect type, location, severity, confidence level, and supporting evidence. The system automatically prioritizes high-risk findings and presents them intuitively on the digital twin model. More importantly, the system feeds back inspection results to previous stages to form a closed loop: newly discovered defects are updated in the S3 risk heat map, and risk weights are adjusted; geometric corrections and ROI optimization suggestions are provided to the S2 digital twin; and the device-defect-environment association rules are updated in the S4 knowledge graph. Through this continuous learning mechanism, the system gradually improves the accuracy of risk prediction and inspection efficiency, optimizes future task planning and route design, and achieves a complete closed loop of "planning-execution-identification-optimization," enabling distribution network inspection to shift from "periodic coverage" to a "risk-driven, precise identification" intelligent mode.

[0226] A smart power distribution network anomaly identification system based on UAV inspection includes a processor, a memory, and a computer program stored in the memory. When the processor executes the computer program, it specifically performs the steps described above in the smart power distribution network anomaly identification method based on UAV inspection.

[0227] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0228] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0229] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0230] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0231] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any other way. Any person skilled in the art may make changes or modifications to the above-disclosed technical content to create equivalent embodiments. However, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the protection scope of the present invention.

Claims

1. A power distribution network abnormal state intelligent identification method based on unmanned aerial vehicle inspection, characterized in that, The method comprises the following steps: S1: Obtain power distribution network inspection related data, including asset master data, historical operation and maintenance data, and environmental constraints; S2: Based on the power distribution network inspection related data, construct a power distribution network digital twin, and obtain the mapping of device ID to spatial pose and parameters; S3: Based on historical defects and environmental risk scoring, generate a risk heat map; S4: Construct a knowledge graph of device-geography-working condition, and combine the mapping of device ID to spatial pose and parameters and the risk heat map to generate inspection targets and priorities by line section, and obtain a task package; S5: According to the task package, the risk heat map and the mapping of device ID to spatial pose and parameters, multi-target route inspection is performed, and the abnormal state of the power distribution network is intelligently identified. 2.The power distribution network abnormal state intelligent identification method based on UAV inspection according to claim 1, characterized in that, The asset master data includes device list, geographic coordinates, structural parameters, importance level, etc.; the historical operation and maintenance data includes defect library, historical alarm, maintenance record; the environmental constraints include airspace restrictions, no-fly zones, weather windows, and terrain data. 3.The power distribution network abnormal state intelligent identification method based on UAV inspection according to claim 2, characterized in that, The power distribution network digital twin is constructed based on the power distribution network inspection related data, and the mapping of device ID to spatial pose and parameters is obtained, which is as follows: Integrate and normalize the power distribution network asset account and GIS data, and on the basis of unified data, construct a multi-level topological skeleton, and based on the established topological skeleton, construct parameterized templates for various devices; Register the aerial survey point cloud with the topological skeleton, correct and optimize the theoretical model using actual aerial survey data, and ensure geometric accurate matching of the digital twin and the physical entity; According to the registered topological skeleton, calculate the accurate attitude angle, assembly vector, installation height and normal parameter of each device, and realize the conversion from geometric surface to engineering parameter; Based on voxelization, generate ROI and sensor reachable domain, construct a power distribution network digital twin, and obtain the mapping of device ID to spatial pose and parameters.

4. The power distribution network abnormal state intelligent identification method based on unmanned aerial vehicle inspection according to claim 3, characterized in that, The power distribution network asset account and GIS data are integrated and normalized, and on the basis of unified data, a multi-level topological skeleton is constructed, which is as follows: Firstly, the power distribution network asset account and GIS data are integrated and normalized, and the basic information of the line is extracted from the power management system, including line number, voltage level, start and end points, tower sequence, and device list structured information, and all geographic data is converted to a unified coordinate system; ； where T source→target is a coordinate transformation matrix, P source and P target are the positions of points in the source and target coordinate systems, respectively. Convert ellipsoidal height to normal height H normal , introduce terrain undulations, and perform height datum conversion: ； where H ellipsoid is the ellipsoidal height, N geoid is the geoid separation, ΔH DEM is the terrain model correction; On the basis of unified data, a multi-level topological graph is constructed: define a directed graph G=(V,E); Wherein, the vertex set V contains tower nodes V tower and device nodes V equip ; the edge set E contains conductor connections E line and device hanging relations E mount ; the attribute set A V and A E contains various parameters of nodes and edges According to the four-level relationship of line, tower, device and component, organize the data, establish a tree-like inheritance structure, and assign a unique identifier to each entity; And the application of graph theory algorithm to verify the rationality of the topology structure, ensure G is connected graph, no isolated node, each tower node v ∈ V tower The degree should meet: deg(v) = N phases x 2 + N ground x 2 + N equipment .

5. The power distribution network abnormal state intelligent identification method based on unmanned aerial vehicle inspection according to claim 3, characterized in that, Based on the established topological skeleton, parameterized templates for various devices are constructed, which are as follows: According to the tower model information in the account, the tower is divided into angle steel tower and steel pipe pole type: The parameterized model of steel pipe pole establishes the function relationship of the following key parameters: the change function of the cross-sectional radius r(h) with the height h: ； where H is the height of the stem, r base and r top are the bottom and top radii, respectively; For each cross arm i, define its height h i , length L i , inclination angle θ i : ； The parameterized model of angle steel tower uses a truncated four-cornered pyramid to describe the main body, and adds a cross arm extension part, and the tower body contour equation: ； Angle steel unit based on rule arrangement, angle steel specification D a With interval S a : Define a standardized hanging point coordinate system for each type of tower, which is the offset vector relative to the tower body center: ； wherein R tower is the tower attitude rotation matrix, P mount,jlocal is the local coordinate system hanging point position; Parameterization of insulator string: establish a model according to the type and specification, and calculate the length of the suspension string: ； wherein N discs is the number of insulating pieces, H disc is the height of a single piece, G gap is the gap; Insulator string posture vector calculation: ； Hardware and connector model, parameterized model based on type; Conductor parameterized model, determine diameter d, unit mass m, elastic modulus E, linear expansion coefficient a according to model; ； where the parameters T is the horizontal tension and g is the acceleration due to gravity. Catenary equation: describe the conductor shape between two points: ； Establish the relationship between sag and temperature T: Where f0 is the sag value at reference temperature T0, L is the span, and a is the linear expansion coefficient.

6. The power distribution network abnormal state intelligent identification method based on unmanned aerial vehicle inspection according to claim 3, characterized in that, Through the above parameterized template, the model specification information in the account is converted into accurate three-dimensional geometric models, and has the ability to dynamically adjust according to temperature. The registration of the aerial survey point cloud and the topological skeleton uses actual aerial survey data to correct and optimize the theoretical model, ensuring the geometric accurate matching of the digital twin and the physical entity, as follows: First, merge aerial survey data from different sensors, including LiDAR point cloud, oblique photography, and video sequence data; Clean and normalize the original LiDAR point cloud data, including: Noise filtering: statistical outlier filtering, ground point separation Point cloud downsampling: voxel grid method to retain structural features and reduce redundancy Point cloud classification: preliminary segmentation of ground, vegetation, buildings, power facilities, etc. tower base center point and ; apparent object reference point and ; Identify corresponding feature points in the account skeleton and point cloud, including: ； wherein w i is a weight corresponding to the point, which is proportional to the confidence level; Rigid body transformation solution: find the best rotation matrix RR and translation vector tt to minimize the distance between corresponding point sets: ； Ensure E global <τ global ,τ global is a preset threshold value; Calculate the overall registration error: Perform independent fine registration for each pole tower and span section, and adjust the parameterized model to best fit the point cloud data: Tower detection and segmentation: Extract tower part P from point cloud tower ; Pole tower model optimization: adjust the position, height, and posture of the pole tower: ； Cylinder fitting: solve the best fitting parameters and update the pole tower parameters: Update the cross arm parameters {h, L,,} based on point cloud clustering and plane fitting i ,L i ,θ i} Fitting precise sag parameters to the wire of each span and applying density clustering to separate the wire point set P from the point cloud wire ; Where θ is the pole tower model parameter, and d(p, M) is the shortest distance from point p to model M; ； Based on the RANSAC method, fit the sag parameters a and x0: Where c is the elevation offset constant; Use spatial clustering to distinguish conductors of different phases and loops, and through parameter tuning and consistency constraints: ensure that the adjusted model meets the physical constraints: ； The length of the conductor between the two end towers should meet the catenary theory: ； The conductor and insulator connection point position should be consistent:

7. The power distribution network abnormal state intelligent identification method based on unmanned aerial vehicle inspection according to claim 3, characterized in that, Based on the fine registration results, update the digital twin skeleton, including pole tower parameter update and conductor parameter update. Based on the voxelization, generate ROI and sensor accessible domain, construct the digital twin of the power distribution network, and obtain the mapping from device ID to spatial pose and parameters, as follows: Convert the power distribution network scene to a regular voxel grid: ； Use octree to improve storage efficiency: Binary occupancy representation: , indicating whether a voxel is occupied Probability occupancy: , representing the occupancy probability Device ID mapping: , associated with the corresponding device identification; Occupancy state and semantic labeling: where r is the voxel resolution, (x min ,y min ,z min ) is the coordinate origin; Cluster point cloud points by voxel: Key area (ROI) definition: device type ROI template: define the area of interest based on device type: Insulator ROI: porcelain disc edge, hardware connection, anti-fouling coating; Conductor ROI: strain clamp, suspension clamp, joint tube, anti-vibration hammer; Hardware ROI: bolt connection, stress position, anti-corrosion layer; ROI primitive = { Sphere(c1, r1), Cylinder(p1, p2, r2)} Use a set of sphere and cylinder geometric primitives to describe the geometric ROI: ； Surface offset: offset the voxels within a distance δ range outward from the device surface And define the ROI according to the detection priority; Define a priority weighted set: ； wherein w r represents the detection priority weight of the ROI region r; For each ROI region, analyze the visibility, cast rays from potential observation points to the ROI, check if it is occluded: ； Field of view angle constraint: the angle between the line connecting the observation point and the center of the ROI and the sensor orientation should be within the field of view angle: ； wherein, is the sensor orientation, θ FOV is the field of view angle; Resolution constraint: the observation distance should ensure the required Ground Sampling Distance (GSD) is reached: ； where d is the observed distance, s pixel is the sensor pixel size, and f is the focal length. Finally, integrate the various parameters obtained in the previous steps into a unified mapping structure, establishing a mapping relationship from the device ID to the complete spatial pose and parameter set. 8.The power distribution network abnormal state intelligent identification method based on UAV inspection according to claim 1, characterized in that, The risk heat map is generated based on the historical defects and environmental risks, and the specific steps are as follows: the historical defects and environmental risks are fused and modeled to form a spatio-temporal risk heat map RiskHeatmap(x, y, t) that can be used for target screening and path planning, which outputs both device-level risk scores and continuous risk fields along the route segment and spatial grid. 9.The power distribution network abnormal state intelligent identification method based on UAV inspection according to claim 1, characterized in that, The device-geography-operation knowledge graph is constructed, and combined with the mapping of device ID to spatial pose and parameters and the risk heat map, the inspection targets and priorities are generated along the route segment to obtain the task package, and the specific steps are as follows: A device-geography-operation three-dimensional knowledge graph is constructed as the semantic basis for intelligent generation of inspection tasks; after the graph is constructed, the system performs multi-source data fusion injection, injecting geometric mapping information from the S2 digital twin module to bind each device ID with accurate spatial pose, geometric parameters, ROI region, and sensor accessibility constraints; injecting risk scores from the S3 risk heat map module to label risk values, uncertainties, and risk driving factor weights for device nodes; At the same time, forecast the working conditions from the meteorological data source to establish dynamic environmental constraints; Based on the fused knowledge graph, the system performs device-level and route segment-level inspection target scoring, and for each device entity, calculates a four-dimensional comprehensive score: including the risk dimension Risk(e), the criticality dimension Criticality(e), the accessibility dimension Access(e), and the timeliness dimension Freshness(e), which are fused into a device-level comprehensive score through weight coefficients: ； wherein w R , w C , w A , w F are weight coefficients; Based on the device scoring, the line segment-level features are aggregated and integrated, and for each line segment, the internal device scores are first weighted and aggregated, with key components receiving higher weights; then the line segment overall risk score is integrated; finally, the environmental trigger item is added, and the line segment scoring model is: The system scores and sorts all line segments, and selects the highest priority device in each high-scoring segment to form a hierarchical inspection target list; Based on target screening and priority sorting, the abstract target is concretized into an executable task package, and first, the system extracts the observation requirements of each target from the knowledge graph, including the applicable sensor type, resolution requirement, recommended shooting times and overlap rate, and geometric constraints. For each target, the system analyzes the ROI characteristics and defect-sensor adaptation relationship to intelligently recommend the best observation scheme.

10. An intelligent identification system for abnormal state of power distribution network based on unmanned aerial vehicle inspection, characterized in that, The computer program is stored in the memory, and when the processor executes the computer program, the steps of the power distribution network abnormal state intelligent identification method based on unmanned aerial vehicle inspection in any one of claims 1-9 are specifically executed.

Images