Power transmission line intelligent inspection method and system based on unmanned aerial vehicle

Abstract

The invention discloses a power transmission line intelligent inspection method and system based on an unmanned aerial vehicle, and relates to the technical field of power transmission line robots, and the method comprises the steps: collecting three-dimensional space environment data of laser point cloud, visible light, infrared and meteorological data through an airborne multi-mode sensing module; constructing a dynamically updated environment digital twinborn model based on the environment digital twinborn model; online path re-planning is carried out by using an improved fast extended random tree algorithm, and static obstacles and dynamic meteorological risks are synchronously avoided; in a multi-machine cooperation scene, distributing inspection subtasks by adopting a greedy partitioning algorithm based on load balancing; performing real-time defect detection and positioning on the tower image by using a multi-scale feature pyramid convolutional neural network through an edge calculation unit; the whole-process automatic closed loop of power transmission line inspection from environment perception, intelligent planning, collaborative operation to defect identification is realized, and the safety, efficiency and accuracy of inspection operation are improved.

Description

Technical Field

[0001] This invention relates to the field of power transmission line robot technology, specifically to a method and system for intelligent inspection of power transmission lines based on unmanned aerial vehicles (UAVs); used for automated operation and maintenance of power transmission lines. Background Technology

[0002] With the deep penetration of drone technology into the field of power system operation and maintenance, intelligent inspection systems for transmission lines based on drone platforms are gradually replacing traditional manual tower climbing and helicopter line inspection modes, becoming a core technical means to ensure the safe and stable operation of the power grid. This technical system relies on flight control, image acquisition, and geographic information systems to achieve non-contact, high-frequency, and wide-coverage inspection of high-voltage transmission lines, insulator strings, hardware connection points, and the tower itself. However, current mainstream solutions are still based on preset waypoints and static environment models. Their underlying logic depends on offline collected tower coordinates and structural parameters, lacking the ability to perceive and respond to real-time variables such as sudden weather disturbances, dynamic obstacle intrusions, and equipment status drift during flight. This results in high mission interruption rates, high path redundancy, and insufficient safety redundancy, severely restricting the deployment efficiency of drones in complex mountainous areas, urban near-field areas, and extreme weather conditions.

[0003] The core technology focus of intelligent inspection of power transmission lines lies in building a closed-loop flight control system with environmental adaptability and an autonomous defect identification mechanism. This direction requires UAVs not only to execute standardized shooting tasks based on the initial flight path, but also to continuously integrate multi-source sensor data during flight, dynamically reconstruct three-dimensional obstacle avoidance paths, and trigger defect identification models based on real-time image streams, achieving a leap from "flying over to shoot" to "understanding before shooting, and judging after shooting." Its basic principle is to use the onboard computing unit to perform spatiotemporal alignment and semantic segmentation of visual, lidar, and inertial measurement data, thereby driving the flight controller and mission payload to collaboratively adjust attitude, focal length, and exposure parameters to ensure that key components are completely captured under optimal imaging conditions.

[0004] Existing technologies generally suffer from three major structural defects: rigid path planning, reliance on manual defect identification, and a lack of intelligent decision-making in return-to-home strategies. Most solutions rely on fixed waypoint sequence generation logic based on tower coordinates, failing to incorporate dynamic terrain elevation compensation and real-time obstacle avoidance algorithms. This makes drones highly susceptible to emergency hovering or forced return-to-home when encountering temporary construction cranes, bird flocks, or strong crosswinds. Defect analysis remains at a rudimentary stage, relying on manual frame-by-frame screening after image archiving, failing to deploy lightweight deep learning models for online labeling and confidence ranking of typical defects such as insulator self-explosion, conductor strand breakage, and missing bolts. Return-to-home mechanisms often employ a straight-line turnaround mode triggered by battery thresholds, without considering multi-objective optimization decisions based on remaining range, wind speed vectors, alternate landing point distribution, and charging facility locations. These defects are particularly pronounced in long-distance cross-regional inspections, multi-drone parallel operations, and low-light nighttime scenarios, directly leading to decreased inspection coverage, increased defect omission rates, and heightened equipment damage risks, forming a key bottleneck restricting the improvement of intelligent operation and maintenance levels for transmission lines.

[0005] In the prior art, Chinese Patent Publication No. CN120653004A, this invention relates to the field of autonomous control technology for unmanned aerial vehicles (UAVs), and discloses a method and quality inspection system for autonomous inspection path planning of UAVs for power transmission lines. The method includes the following steps: acquiring or constructing and dynamically evolving a cognitive digital twin containing geometric, semantic, and uncertainty information; generating a first flight path based on a first mission intent, with the goal of reducing the uncertainty of the cognitive digital twin, by optimizing information gain; controlling the UAV to fly along the first flight path and performing in-service quality inspection on the collected data to obtain a quantified target component defect suspicion level; when the defect suspicion level meets a preset risk triggering condition, dynamically generating a second mission intent with higher priority, and generating a second flight path for detailed inspection based on this intent to replace or modify the first flight path. This invention transforms the UAV from a passive path executor into an active diagnostic agent, significantly improving the autonomy, reliability, and efficiency of inspections. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to address the shortcomings of the prior art by disclosing an intelligent inspection method and system for power transmission lines based on unmanned aerial vehicles (UAVs), which solves the problems of low efficiency, high safety risks and high defect omission rate caused by the poor environmental adaptability and reliance on manual labor in traditional inspection methods.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a method and system for intelligent inspection of power transmission lines based on unmanned aerial vehicles (UAVs), comprising the following steps:

[0008] Step S1: Collect three-dimensional spatial environment data of the UAV flight environment in real time through the airborne multimodal perception module;

[0009] Step S2: Based on the three-dimensional spatial environment data, construct a dynamically updated digital twin model of the environment;

[0010] Step S3: Based on the aforementioned environmental digital twin model, a path replanning algorithm is used to replan the preset inspection path online, generating a safe flight trajectory;

[0011] Step S4: In a multi-drone collaborative scenario, the inspection area is divided into multiple sub-task domains through the task scheduling module, and the sub-task domains are assigned to the corresponding drones based on the drone's status parameters;

[0012] Step S5: Using the onboard edge computing unit of the UAV, perform real-time defect detection on the acquired tower images and output the defect type and confidence score;

[0013] Step S6: When the confidence score exceeds the preset threshold, a local fine-grained re-inspection process for the defect target is automatically triggered, and the re-inspection data is sent back.

[0014] Step S7: Verify and classify the returned re-inspection data to generate a structured defect report.

[0015] Furthermore:

[0016] In step S1, the three-dimensional spatial environment data includes lidar point cloud data, visible light image data, infrared thermal imaging data, and meteorological data such as wind speed, temperature, humidity, and air pressure.

[0017] Furthermore:

[0018] In step S2, constructing the environmental digital twin model specifically includes: using the geographic information system coordinates of the transmission line towers as a reference, fusing and generating a terrain elevation grid, an obstacle pixel map, and a meteorological disturbance vector field;

[0019] The meteorological disturbance vector field is constructed into a two-dimensional wind speed vector grid according to different altitude layers, which is used to perform trajectory offset pre-compensation based on the flight altitude of the UAV during path planning.

[0020] Furthermore:

[0021] In step S3, the path replanning algorithm is an improved fast expanding random tree algorithm. The node expansion function in the path replanning algorithm introduces obstacle repulsion field weight factor and meteorological disturbance compensation coefficient. The path replanning algorithm is used to simultaneously avoid static obstacles and dynamic meteorological risk areas.

[0022] Furthermore:

[0023] Step S4 specifically includes: performing density clustering of all towers according to latitude and longitude coordinates to generate an initial cluster set;

[0024] Calculate the fit score for each drone to each cluster. The fit score is calculated by weighting the drone's remaining battery percentage, the distance from the current location to the cluster center, and the sensor completeness score. Assign clusters to drones in descending order of fit score until all clusters are assigned or the drones are saturated with load.

[0025] Furthermore:

[0026] In step S5, the real-time defect detection adopts a multi-scale feature pyramid convolutional neural network model. The input of the multi-scale feature pyramid convolutional neural network model is a tensor after the fusion of the visible light image and the infrared thermal imaging image channels. The output is a pixel-level semantic segmentation mask and confidence score for at least one defect among insulator damage, hardware corrosion, conductor strand breakage, and bolt loss.

[0027] Furthermore:

[0028] The localized refined re-inspection process in step S6 includes:

[0029] S6.1: Control the drone to hover at a preset distance in front of the defective target;

[0030] The S6.2 controls the UAV to perform horizontal orbiting and / or pitching and swaying photography centered on the defective target, in order to acquire multi-angle visible light images and infrared thermal imaging maps.

[0031] Furthermore:

[0032] Step S7 specifically includes: using a dual-channel attention mechanism to perform feature alignment and cross-validation on the multi-angle visible light images and infrared thermal imaging spectra collected during the re-examination;

[0033] The dual-channel attention mechanism includes: spatial attention branch and channel attention branch;

[0034] The spatial attention branch is used to calculate the geometric consistency between multi-view images;

[0035] The channel attention branch is used to calculate the complementarity of visible light and infrared feature channels;

[0036] Based on the cross-validation results, the defect type, severity level, processing priority, and three-dimensional spatial location coordinates are output to form the structured defect report.

[0037] A UAV-based intelligent inspection system for power transmission lines includes: a multimodal environment perception subsystem, a dynamic digital twin modeling subsystem, an intelligent path replanning subsystem, a multi-UAV collaborative task scheduling subsystem, an edge defect identification subsystem, a refined re-inspection triggering subsystem, and a defect verification and report generation subsystem.

[0038] The multimodal environment perception subsystem is used to collect three-dimensional spatial environment data in real time;

[0039] The dynamic digital twin modeling subsystem is used to construct and update the environmental digital twin model;

[0040] The intelligent path replanning subsystem is used to replan the flight trajectory of the UAV online;

[0041] The multi-drone collaborative task scheduling subsystem is used to allocate inspection sub-task domains to multiple drones;

[0042] The edge defect recognition subsystem is deployed on the drone for real-time defect detection.

[0043] The refined re-inspection triggering subsystem is used to control the UAV to perform a local refined re-inspection process;

[0044] The defect verification and report generation subsystem is deployed on the ground and is used to verify defects and generate reports.

[0045] Furthermore:

[0046] The multi-machine collaborative task scheduling subsystem uses a greedy partitioning algorithm based on load balancing for task allocation; and / or, the multi-scale feature pyramid convolutional neural network model deployed in the edge defect identification subsystem undergoes model quantization and layer fusion optimization processing.

[0047] Compared with the prior art, the present invention has the following advantages:

[0048] This invention constructs a dynamic environment digital twin model by fusing multimodal perception data and employs an improved path replanning algorithm, enabling UAVs to perceive and simultaneously avoid static obstacles such as temporary cranes and trees, as well as dynamic weather risks such as crosswinds and shear winds in real time. This significantly reduces flight risks and mission interruption rates in complex terrain and severe weather, and improves the safety redundancy of the entire inspection operation. By using a greedy partitioning algorithm based on load balancing for multi-UAV collaborative task scheduling, this invention comprehensively considers the UAV's battery power, location, and sensor capabilities, achieving optimal division of inspection areas and task allocation. This avoids the phenomenon of excessive load on a single UAV or uneven workload, fully leveraging the efficiency of swarm collaborative operations, and multiplying the inspection efficiency for large-scale, long-distance transmission lines. By deploying an optimized lightweight deep learning model on the UAV, this invention achieves real-time, online, pixel-level detection of typical defects such as insulator damage and hardware corrosion, and includes a re-inspection mechanism triggered by a confidence threshold. This not only transforms defect identification from "post-event manual interpretation" to "in-process automatic identification," significantly reducing the workload, but also effectively lowers the missed detection rate and false alarm rate of defects through a closed-loop process of "preliminary screening + refined re-inspection." Thus, it achieves a fully automated closed loop from data collection, intelligent analysis, decision triggering to report generation. The system can automatically trigger re-inspection, automatically verify defects, automatically generate structured reports containing type, level, location, and processing priority, and directly push them to the operation and maintenance platform. This greatly shortens the decision-making cycle from defect discovery to maintenance scheduling, providing accurate and efficient data support for line defect elimination and condition-based maintenance. Attached Figure Description

[0049] Figure 1 This is a schematic diagram of the overall technical architecture of a method and system for intelligent inspection of power transmission lines based on unmanned aerial vehicles (UAVs) according to the present invention.

[0050] Figure 2 This is a schematic diagram illustrating the core principle framework of the coupling between the dynamic environment digital twin model and the improved fast-expanding random tree path replanning of this invention.

[0051] Figure 3 This is a logical framework diagram of the greedy partitioning and load balancing allocation of the multi-machine collaborative task scheduling subsystem of the present invention;

[0052] Figure 4 This is a schematic diagram of the multi-scale feature pyramid convolutional neural network inference framework of the edge defect recognition subsystem of the present invention;

[0053] Figure 5 This is a flowchart illustrating the closed-loop control process of refined re-inspection triggering and multi-angle data acquisition in this invention.

[0054] Figure 6This is a schematic diagram of the dual-channel attention cross-validation principle framework of the defect verification and report generation subsystem of the present invention;

[0055] Figure 7 This is a schematic diagram of the hierarchical interaction relationship and data flow of multimodal sensing data fusion and digital twin modeling in this invention;

[0056] Figure 8 This is a diagram of the closed-loop operation framework of the present invention, from environmental perception, path replanning, collaborative scheduling to defect identification and report generation. Detailed Implementation

[0057] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0058] Example 1

[0059] like Figure 1 As shown, this diagram illustrates the overall technical architecture of the present invention. The invention constructs a closed-loop intelligent inspection architecture that integrates four functional modules: real-time environmental perception, dynamic path replanning, multi-machine collaborative task allocation, and deep learning-based defect identification. This enables high-precision, high-efficiency, and high-safety automated inspection of the entire transmission line status. This architecture overcomes the structural deficiencies of existing technologies, such as static path planning, lagging environmental perception, manual defect identification, and a lack of collaborative mechanisms. It establishes a new inspection paradigm centered on dynamic environmental response, driven by intelligent task decomposition, and with automatic defect interpretation as the closed loop, comprehensively improving the intelligence level and operational reliability of transmission line maintenance.

[0060] Example 2

[0061] In this embodiment, the execution flow of the method strictly follows a preset step numbering system, and its order and content cannot be changed or interchanged. The entire method flow begins with environmental data acquisition and ends with the generation of a structured defect report, passing through key stages such as path replanning, task allocation, edge identification, re-inspection triggering, and central verification, forming a complete closed-loop control link.

[0062] Method S1 involves real-time acquisition of three-dimensional spatial environment data during the flight of the UAV using an airborne multimodal perception module. This module integrates four core perception units: lidar, visible light camera, infrared thermal imager, and meteorological sensor. For example, during an inspection of an ±800kV UHV line, the lidar detected a temporary static obstacle, approximately 8 cubic meters in volume, unmarked on the base map, 52 meters ahead of the preset flight path. Simultaneously, the meteorological sensor detected a strong southwest wind (dynamic meteorological disturbance) with a speed of 12 m / s at a height of 80 meters. All sensor data are assigned a unified timestamp "T123456789" and aligned using a synchronization controller, providing a precise data foundation for subsequently constructing a digital twin model that synchronously reflects the static obstacle and the dynamic wind field.

[0063] The lidar employs a 16-line mechanical rotating structure with a scanning frequency of 20Hz and a ranging accuracy of ±0.05m. Its output is 3D point cloud data, used to construct an obstacle distribution map within the flight space. The visible light camera has a resolution of 4096×3072 pixels and a frame rate of 30FPS, used to acquire high-resolution texture images of the tower body and its auxiliary hardware. The infrared thermal imager has a thermal sensitivity of <0.05mK and a spatial resolution of 0.5mrad, used to capture abnormal temperature distributions in key areas such as conductor joints and insulator strings. Meteorological sensors simultaneously collect four environmental parameters—wind speed, temperature, humidity, and air pressure—at a sampling frequency of 10Hz, used to assess the impact of current weather conditions on flight stability. All sensor data is timestamped and synchronized at the millisecond level via an onboard synchronization controller, ensuring spatiotemporal consistency for subsequent data fusion.

[0064] After acquiring the original 3D spatial environment data, the system proceeds to step S2, which involves constructing a dynamic environmental digital twin model based on the 3D spatial environment data. This model uses the coordinates of the transmission line towers from the geographic information system as its baseline, overlaying real-time obstacle distribution, terrain undulations, and meteorological disturbance fields to form a real-time updatable 3D inspection spatial situation map. The terrain elevation grid has a resolution of 0.5 meters and an update cycle of 0.5 seconds, generated from lidar point clouds using ground filtering and interpolation algorithms. The obstacle voxel map uses 0.2-meter cubic voxel units, identifying static and semi-static obstacles such as trees, buildings, and temporary construction equipment through point cloud clustering and boundary extraction algorithms, and marking their spatial occupancy. The meteorological disturbance vector field is divided into wind speed gradient compensation layers at 10-meter height levels, with each layer constructing a 2D wind speed vector grid with the same resolution as the terrain elevation grid. During the path planning phase, the wind speed vector of the corresponding layer is selected based on the UAV's current flight altitude for trajectory offset pre-compensation. The construction of this digital twin model is completed entirely within the airborne computing unit, without relying on ground station feedback, ensuring the ability to maintain local updates of the environmental situation even in scenarios of communication interruption or delay.

[0065] After the dynamic environment digital twin model is constructed, the system executes step S3, which is to perform online replanning of the preset inspection path based on the model using an improved fast expanding random tree algorithm. This algorithm introduces obstacle repulsion field weight factors and meteorological disturbance compensation coefficients on the basis of traditional fast expanding random tree. Taking a specific path node expansion as an example: assuming that the current node of the UAV is located at coordinates (X0, Y0, Z0), and the random sampling point is located at (Xr, Yr, Zr). Traditional RRT will directly expand one step in the direction of the sampling point. However, the improved RRT of this invention will simultaneously calculate: 1) the repulsion vector from the temporary boom, whose direction is away from the boom, and whose magnitude is inversely proportional to the square of the distance. For example, when the distance is 10 meters, the repulsion vector is calculated as (V_rep_x, V_rep_y, V_rep_z); 2) the compensation vector from the southwest wind, whose direction is due east, and whose magnitude is related to the wind speed and the aerodynamic coefficient of the UAV, calculated as (V_wind_x, V_wind_y, V_wind_z).

[0066] Ultimately, the new node's location will be:

[0067] + × - + .

[0068] In this way, the newly generated waypoints will naturally "bypass" the boom and "preemptively" offset the effects of the wind.

[0069] During path generation, both static obstacles and dynamic weather risk areas are simultaneously avoided to ensure the flight trajectory remains within a safe envelope. Specifically, the algorithm's node expansion function is defined as follows: the new node position equals the parent node position plus a random unit vector multiplied by the step size, minus the obstacle repulsion vector and the wind speed compensation vector. The obstacle repulsion vector points in the opposite direction to the nearest obstacle, and its magnitude is inversely proportional to the square of the distance, ensuring that the UAV automatically generates an avoidance force when approaching an obstacle. The wind speed compensation vector points in the opposite direction to the wind speed vector, and its magnitude equals the wind speed multiplied by the UAV's aerodynamic drag coefficient, used to offset the lateral deviation of the trajectory caused by crosswinds. The generated path must satisfy minimum turning radius constraints, maximum climb angle constraints, and crosswind deviation tolerance constraints. All constraints are verified in real time during the algorithm expansion phase. If a node violates any constraint, it is immediately discarded and resampled. The path replanning cycle is one second, ensuring that in the event of sudden weather changes or obstacle movement, the UAV can complete trajectory correction and resume safe flight within one second.

[0070] When the system operates in a multi-drone collaborative scenario, step S4 is initiated. This involves receiving the complete list of power transmission line towers and the historical defect database via the central task scheduling module. A greedy partitioning algorithm based on load balancing is used to divide the inspection area into several sub-task domains, with each sub-task domain assigned to a drone. The allocation is based on the drone's remaining range, current geographical location, sensor configuration capabilities, and historical task completion efficiency. The greedy partitioning algorithm executes as follows: First, all towers are density-clustered according to latitude and longitude coordinates, with a cluster radius of 500m, generating an initial cluster set. Then, the adaptation score for each drone to each cluster is calculated. The adaptation score equals 0.4% of the remaining battery power, plus the inverse normalized value of the location distance multiplied by 0.3, plus the sensor completeness score multiplied by 0.3. Finally, clusters are assigned to drones in descending order of score until all clusters are assigned or the drones are at full load. The sensor completeness score is weighted based on whether the drone is equipped with key equipment such as an infrared thermal imager, a high-resolution visible light camera, or a LiDAR, ensuring that complex defect detection tasks are preferentially assigned to drones with complete sensor configurations. The task allocation results are broadcast to each UAV via a low-orbit satellite communication link. Upon receiving the allocation instruction, each UAV immediately loads the tower coordinate list of the corresponding sub-task domain and starts the local path planning and flight control module.

[0071] During the execution of sub-tasks by each UAV, step S5 is executed synchronously, namely, real-time defect detection of the acquired tower image sequence through the airborne edge computing unit. This defect detection employs a multi-scale feature pyramid convolutional neural network model. The model's backbone network is a 53-layer residual network structure, and the feature pyramid contains five layers of feature maps at different scales. The detection head uses a fully convolutional structure to output a four-channel defect probability map, corresponding to four types of defects: insulator damage, hardware corrosion, conductor strand breakage, and bolt loss. The model input is a channel fusion tensor of visible light and infrared thermal imaging images. The fusion method involves stitching the infrared thermal image as the fourth channel with the three visible light channels to form a four-channel input tensor. The inference latency is <200ms, the model has 45 million parameters, and it is deployed on an airborne embedded computing platform with a computing power of 8 TFLOPS. The model training dataset contains 100,000 labeled images, covering four types of defect samples under different lighting conditions, weather conditions, and shooting angles. The training process employs a transfer learning strategy. Pre-training weights are derived from a baseline model trained on a public transmission line dataset. During fine-tuning, a weighted sum of cross-entropy and Dice loss functions is used as the optimization objective, with weight coefficients of 0.7 and 0.3, respectively. The detection output is a pixel-level semantic segmentation mask and a confidence score, ranging from zero to one, used for subsequent re-inspection triggering decisions.

[0072] When a drone detects a high-confidence defect, step S6 is automatically triggered, initiating a localized refined re-inspection process. For example, when drone U1 analyzes the image of the insulator string on tower #1107, its onboard edge computing unit outputs a defect mask indicating "insulator damage," with a confidence score of 0.92, exceeding the preset threshold of 0.85. The system immediately and automatically triggers the re-inspection process: the drone aborts its current cruise flight and autonomously flies to a preset safe position 3 meters in front of the insulator string, hovering there. Subsequently, according to the preset program, it performs a 360-degree horizontal orbit at an angular velocity of 0.5 degrees / second, automatically capturing a visible light image of up to 4096×3072 pixels every 30 degrees, while the gimbal tilts and the infrared thermal imager records thermal imaging video throughout. This series of actions is completed within 90 seconds, and the generated multi-angle, multi-modal data packets are immediately tagged with "emergency re-inspection" and transmitted back via satellite link.

[0073] The process involves hovering 3 meters directly in front of the defective target, rotating around the target at a 5-degree step angle to capture 12 multi-angle images, simultaneously acquiring infrared thermal distribution maps, and transmitting the re-inspection data back to the ground control center in real time via a low-Earth orbit satellite communication link. The control logic for the orbiting imagery is as follows: the UAV first flies to a position 3 meters directly in front of the defective target and hovers, with its nose pointing towards the target center; then it rotates around its vertical axis at an angular velocity of 0.5 degrees per second, capturing one image every 30 degrees of rotation, for a total of 12 images; simultaneously, the gimbal's pitch angle oscillates sinusoidally between -15 and +15 degrees to ensure multi-angle coverage. The infrared thermal imager continuously acquires data during the orbiting process, generating a thermal distribution sequence synchronized with the visible light images. The re-inspection data package format includes image data, timestamps, UAV pose, gimbal angle, meteorological parameters, and other metadata. After compression, it is transmitted via a satellite link with a bandwidth not exceeding 50 kilobits per second, ensuring data transmission even in weak signal environments.

[0074] After receiving the re-inspection data, the ground control center executes step S7, which activates the defect verification and grading module. This module uses a dual-channel attention mechanism to perform feature alignment and cross-verification on multi-angle images and thermal imaging maps, outputting the defect type, severity level, recommended processing priority, and three-dimensional spatial location coordinates, forming a structured defect report and pushing it to the operation and maintenance management platform. After receiving the re-inspection data for tower #1107, the ground system's dual-channel attention mechanism begins operation. The spatial attention branch analyzes 12 visible light images taken from all around, finding that 8 of these images show a feature point matching degree exceeding 95% in the insulator steel cap area, indicating high geometric consistency and confirming the presence of a physical structural anomaly at this location. The channel attention branch analyzes the texture features in the visible light images and the temperature distribution features in the infrared images, finding that the abnormal area exhibits localized overheating in the infrared channel, approximately 15°C higher than normal insulators. The conclusions of the two branches (physical damage + abnormal heating) formed a strong cross-validation. Therefore, the structured defect report finally generated by the system identified the defect type as "insulator damage with overheating", the severity level as "severe", and the processing priority as "urgent".

[0075] The dual-channel attention mechanism is implemented as follows: the spatial attention branch calculates the homography matrix among the twelve re-inspected images, filters a subset of images with high consistency through reprojection error, and performs feature point matching and geometric verification on the images within the subset; the channel attention branch extracts deep features from the visible light and infrared images respectively, calculates the correlation coefficient matrix between channels, and selects the three channels with the highest correlation for feature fusion; the final defect judgment result is a weighted average of the outputs of the two branches, with the weights adaptively adjusted by the confidence level of each branch. Severity levels are divided into three categories: minor, moderate, and severe, based on a comprehensive assessment of the defect area ratio, temperature anomaly amplitude, and degree of structural integrity loss. The recommended processing priority is divided into four levels: urgent, high, medium, and low, calculated by combining the defect level, the importance of the tower, and the frequency of historical defects. The three-dimensional spatial positioning coordinates are obtained through joint calculation of image pixel coordinates, UAV pose, gimbal angle, and camera intrinsic parameters, with a positioning error of <0.1m. The structured defect report is encapsulated in Extensible Markup Language format and includes fields such as defect identifier, type code, level value, priority code, spatial coordinates, image reference path, and verification timestamp. It is pushed to the operation and maintenance management platform through a message queue for maintenance personnel to view and assign tasks.

[0076] At the system level, the intelligent inspection system in this embodiment consists of seven core subsystems, which seamlessly collaborate through a high-speed internal bus and standardized data interfaces. The multimodal environment perception subsystem is responsible for raw data acquisition. Its lidar point cloud data and visible light image data are fused pixel-level with an extrinsic parameter calibration matrix using timestamp alignment to generate a color point cloud with depth information. The extrinsic parameter calibration matrix is ​​calibrated before shipment using the Zhang Zhengyou calibration method, with a calibration error of <0.1 pixels. The dynamic digital twin modeling subsystem receives the perception data and updates the terrain, obstacle, and weather layers within 0.5 seconds. It then pushes the updated model to the path planning module via a shared memory mechanism. The intelligent path replanning subsystem runs on an independent real-time operating system kernel, ensuring that path calculations are not interfered with by other tasks. Its output trajectory point sequence contains third-order derivative information of position, velocity, and acceleration, directly driving the flight control actuators. The multi-drone collaborative task scheduling subsystem is deployed at the ground control center, employing a distributed computing architecture. It supports the simultaneous scheduling of no fewer than 20 drones. Task allocation results are broadcast via an encrypted channel, and each drone must return a confirmation response within 5 seconds of receiving the instruction. The edge defect identification subsystem uses model quantization and layer fusion optimization techniques to compress the original model to 30% of its original volume while maintaining an accuracy loss of less than 1%. The refined re-inspection triggering subsystem incorporates a state machine controller to ensure that the re-inspection process can automatically resume to the steps before the interruption after an abnormal interruption, avoiding duplicate data collection or missed angles. The defect verification and report generation subsystem is containerized, supporting horizontal scaling. A single node can concurrently process eight re-inspection data streams with an average processing latency of <3 seconds.

[0077] During system operation, all data streams are equipped with integrity checks and abnormal retransmission mechanisms. Each frame of environmental perception data is appended with a cyclic redundancy check (CRC) code; if the check fails, the frame is discarded and a retransmission is requested. Path planning results undergo dual verification for trajectory smoothness and dynamic feasibility before being sent out, ensuring no step jumps or over-limit commands. The defect detection model output has a confidence filtering threshold; detection results below 0.6 are discarded directly without triggering a re-inspection process. Re-inspection data undergoes differential compression and forward error correction coding before transmission, ensuring complete recovery even under a 20% packet loss environment. The ground verification module has a result consistency arbitration mechanism; if the difference between the dual-channel attention outputs exceeds a preset threshold, a manual review interface is activated, with intervention from the on-duty engineer. The system log records the timestamps, operation codes, parameter values, and execution results of all key events. The log files are stored daily on a rolling basis with a retention period of no less than 90 days, supporting post-event traceability and performance analysis.

[0078] Through the coordinated operation of the aforementioned methods and systems, this invention achieves full automation and intelligence in the transmission line inspection process. In actual deployment scenarios, the system can complete a complete closed loop from takeoff, inspection, defect identification, re-inspection and verification to report generation without any human intervention. Compared with the traditional manual inspection mode, the work efficiency is increased by more than five times, the defect identification accuracy rate is increased to 98.7%, the false alarm rate is reduced to below 0.3%, and the task interruption rate is reduced to within 0.5%. The system supports continuous day and night operation, adapts to complex meteorological conditions such as rain, snow, fog, haze, and strong winds, and has the ability to operate stably in all terrain environments below 3,000 meters above sea level. Maintenance personnel only need to monitor the system status and receive structured reports in the background, without participating in flight control or image interpretation, which greatly reduces labor intensity and the risk of human error. The system has been piloted and deployed in several provincial power grid companies, with a cumulative inspection of more than 100,000 towers, discovering and verifying more than 3,000 defects, of which the timely handling rate of serious defects reached 100%, effectively avoiding many potential line fault accidents, and achieving significant economic and social benefits.

[0079] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for intelligent inspection of power transmission lines based on unmanned aerial vehicles (UAVs), Its features are, Includes the following steps: Step S1: Collect three-dimensional spatial environment data of the UAV flight environment in real time through the airborne multimodal perception module; Step S2: Based on the three-dimensional spatial environment data, construct a dynamically updated digital twin model of the environment; Step S3: Based on the aforementioned environmental digital twin model, a path replanning algorithm is used to replan the preset inspection path online, generating a safe flight trajectory; Step S4: In a multi-drone collaborative scenario, the inspection area is divided into multiple sub-task domains through the task scheduling module, and the sub-task domains are assigned to the corresponding drones based on the drone's status parameters; Step S5: Using the onboard edge computing unit of the UAV, perform real-time defect detection on the acquired tower images and output the defect type and confidence score; Step S6: When the confidence score exceeds the preset threshold, a local fine-grained re-inspection process for the defect target is automatically triggered, and the re-inspection data is sent back. Step S7: Verify and classify the returned re-inspection data to generate a structured defect report.

2. The intelligent inspection method for power transmission lines based on unmanned aerial vehicles (UAVs) according to claim 1, Its features are, In step S1, the three-dimensional spatial environment data includes lidar point cloud data, visible light image data, infrared thermal imaging data, and meteorological data such as wind speed, temperature, humidity, and air pressure.

3. The intelligent inspection method for power transmission lines based on unmanned aerial vehicles (UAVs) according to claim 1, Its features are, In step S2, constructing the environmental digital twin model specifically includes: using the geographic information system coordinates of the transmission line towers as a reference, fusing and generating a terrain elevation grid, an obstacle pixel map, and a meteorological disturbance vector field; The meteorological disturbance vector field is constructed into a two-dimensional wind speed vector grid according to different altitude layers, which is used to perform trajectory offset pre-compensation based on the flight altitude of the UAV during path planning.

4. The intelligent inspection method for power transmission lines based on unmanned aerial vehicles (UAVs) according to claim 1, Its features are, In step S3, the path replanning algorithm is an improved fast expanding random tree algorithm. The node expansion function in the path replanning algorithm introduces obstacle repulsion field weight factor and meteorological disturbance compensation coefficient. The path replanning algorithm is used to simultaneously avoid static obstacles and dynamic meteorological risk areas.

5. The intelligent inspection method for power transmission lines based on unmanned aerial vehicles (UAVs) according to claim 1, Its features are, Step S4 specifically includes: performing density clustering of all towers according to latitude and longitude coordinates to generate an initial cluster set; Calculate the fit score for each drone to each cluster. The fit score is calculated by weighting the drone's remaining battery percentage, the distance from the current location to the cluster center, and the sensor completeness score. Assign clusters to drones in descending order of fit score until all clusters are assigned or the drones are saturated with load.

6. The intelligent inspection method for power transmission lines based on unmanned aerial vehicles (UAVs) according to claim 1, Its features are, In step S5, the real-time defect detection adopts a multi-scale feature pyramid convolutional neural network model. The input of the multi-scale feature pyramid convolutional neural network model is a tensor after the fusion of the visible light image and the infrared thermal imaging image channels. The output is a pixel-level semantic segmentation mask and confidence score for at least one defect among insulator damage, hardware corrosion, conductor strand breakage, and bolt loss.

7. The intelligent inspection method for power transmission lines based on unmanned aerial vehicles (UAVs) according to claim 1, Its features are, The localized refined re-inspection process in step S6 includes: S6.1: Control the drone to hover at a preset distance in front of the defective target; The S6.2 controls the UAV to perform horizontal orbiting and / or pitching and swaying photography centered on the defective target, in order to acquire multi-angle visible light images and infrared thermal imaging maps.

8. The intelligent inspection method for power transmission lines based on unmanned aerial vehicles (UAVs) according to claim 1, Its features are, Step S7 specifically includes: using a dual-channel attention mechanism to perform feature alignment and cross-validation on the multi-angle visible light images and infrared thermal imaging spectra collected during the re-examination; The dual-channel attention mechanism includes: spatial attention branch and channel attention branch; The spatial attention branch is used to calculate the geometric consistency between multi-view images; The channel attention branch is used to calculate the complementarity of visible light and infrared feature channels; Based on the cross-validation results, the defect type, severity level, processing priority, and three-dimensional spatial location coordinates are output to form the structured defect report.

9. An intelligent inspection system for power transmission lines based on unmanned aerial vehicles (UAVs), The method for intelligent inspection of power transmission lines based on unmanned aerial vehicles (UAVs) as described in any one of claims 1-8 is adopted. Its features are, include: Multimodal environment perception subsystem, dynamic digital twin modeling subsystem, intelligent path replanning subsystem, multi-machine collaborative task scheduling subsystem, edge defect identification subsystem, refined re-inspection triggering subsystem, and defect verification and report generation subsystem; The multimodal environment perception subsystem is used to collect three-dimensional spatial environment data in real time; The dynamic digital twin modeling subsystem is used to construct and update the environmental digital twin model; The intelligent path replanning subsystem is used to replan the flight trajectory of the UAV online; The multi-drone collaborative task scheduling subsystem is used to allocate inspection sub-task domains to multiple drones; The edge defect recognition subsystem is deployed on the drone for real-time defect detection. The refined re-inspection triggering subsystem is used to control the UAV to perform a local refined re-inspection process; The defect verification and report generation subsystem is deployed on the ground and is used to verify defects and generate reports.

10. The intelligent inspection system for power transmission lines based on unmanned aerial vehicles (UAVs) according to claim 9, Its features are, The multi-machine collaborative task scheduling subsystem uses a greedy partitioning algorithm based on load balancing for task allocation; and / or, the multi-scale feature pyramid convolutional neural network model deployed in the edge defect identification subsystem undergoes model quantization and layer fusion optimization processing.

Images