A power transmission line defect differentiation identification method for unmanned aerial vehicle cluster inspection

Abstract

This invention discloses a differential identification method for power line defects in UAV swarm inspection, relating to the fields of power inspection and UAV swarm control technology. The method sequentially completes spatiotemporal index construction and image geometric registration, differential detection and defect identification, distributed collaborative perception modeling, dynamic task allocation, edge collaborative learning, hierarchical trajectory planning, and robust replanning. It combines feature point extraction and feature matching algorithms to achieve accurate image alignment, and uses a two-stage target detection method and differential identification method to achieve defect identification. It integrates federated averaging and the UAV offline decision model to achieve swarm adaptation, employs a hierarchical architecture and deep reinforcement learning to complete trajectory planning, and designs a robust replanning strategy to ensure flight safety. This invention effectively reduces the false alarm rate of minor defects, improves the robustness of UAV swarm communication and the adaptability of trajectory planning, and provides technical support for the digitalization and intelligentization of power inspection.

Description

Technical Field

[0001] This invention relates to the field of power line inspection and UAV swarm control technology, and in particular to a method for differentiated identification of transmission line defects for UAV swarm inspection. Background Technology

[0002] As the scale of power grids continues to expand, the requirements for accuracy and efficiency in identifying minute defects during transmission line inspections are constantly increasing. Unmanned aerial vehicle (UAV) swarms, with their advantages of wide operating range, high collaborative efficiency, and adaptability to complex terrain, have become a core technological means for intelligent transmission line inspections. Their swarm collaborative control, obstacle avoidance in complex environments, and accurate identification of minute defects have become crucial factors determining the success or failure of inspection tasks. Transmission line inspection scenarios are characterized by long distances, multiple obstructions, and strong electromagnetic interference, placing far higher demands on the movement accuracy, communication reliability, and target identification anti-interference capabilities of UAV swarms than on ordinary inspection tasks. Therefore, the implementation of UAV swarm inspection technology urgently needs to address the challenges of collaboration, planning, and identification in complex scenarios.

[0003] Currently, collaborative control of UAV swarms mainly relies on two technologies: swarm bionic intelligence and deep reinforcement learning. While these technologies can achieve good collaborative results under ideal communication conditions, the operation of most collaborative algorithms is highly dependent on frequent and highly reliable state information exchanges between UAVs. In the highly interference environment of power transmission corridors, it is difficult to maintain low-latency and highly stable inter-UAV communication links. Communication interruptions can easily lead to swarm collaboration failures, resulting in inspection mission failures or even UAV collisions. Meanwhile, the technology for identifying minor defects in power transmission lines currently still focuses on single-frame target recognition as its core approach, using deep learning models to detect and locate defects. However, in the complex background of power transmission line inspections, the noise interference from background textures, lighting changes, and seasonal changes often exceeds the intensity of the minor defects themselves, making the model prone to false detections and resulting in a high false alarm rate, which fails to meet the actual needs of automated inspections.

[0004] In terms of UAV trajectory planning and obstacle avoidance, existing technologies largely rely on accurate GPS pose information and stable communication links for planning and control. While this ensures planning accuracy in normal environments, the performance of the planning system drops sharply in scenarios common in power transmission corridors, such as GPS denial and unstable communication. Furthermore, existing trajectory planning schemes often prioritize navigation and obstacle avoidance safety, failing to adequately consider the operational requirements of inspection imaging. This makes it difficult to balance flight safety with optimization of camera observation angles, potentially leading to successful obstacle avoidance but poor inspection image quality.

[0005] In summary, existing drone inspection technologies have not yet formed an integrated solution adapted to the complex scenarios of power transmission lines. They have significant shortcomings in areas such as robustness of cluster collaborative communication, anti-interference ability of minor defect identification, and adaptability of trajectory planning operations, and cannot meet the actual needs of the digital and intelligent transformation of power grid inspection. Summary of the Invention

[0006] The purpose of this invention is to propose a method for differentiated identification of transmission line defects for UAV swarm inspection, which solves the technical problems of high communication dependence of UAV swarms, high false alarm rate of minor defects, and weak anti-interference ability of trajectory planning in transmission line inspection, and achieves high precision, high reliability and strong robustness of unmanned autonomous inspection of transmission lines.

[0007] To achieve the above objectives, this invention proposes a method for differentiated identification of power transmission line defects for UAV swarm inspection, the steps of which are as follows: Step S1: The UAV cluster collects real-time data packets of the power transmission line through airborne equipment, and at the same time uses airborne multi-source sensors to obtain local environmental observations of the inspection area. Step S2: Construct a spatiotemporal index database, perform spatial coarse screening and fine spatiotemporal matching on the real-time data packets collected by the UAV, select the optimal reference image, and perform geometric registration between the real-time image and the optimal reference image. Step S3: Based on the registration results obtained in step S2, extract the features of the registered image pairs and generate a real-time difference feature map; use a two-stage target detection method and a differential identification method to accurately locate defects and identify defect types in the real-time difference feature map, and output the defect identification results. Step S4: The drones exchange minimum necessary state information through a chain or mesh communication network, and model the drone swarm as a dynamic undirected graph; Step S5: Model the inspection task as a task to be assigned. Each drone calculates the task completion rate based on its own status and assigns the task through a distributed consensus auction algorithm. Step S6: Each UAV collects inspection experience data locally, updates the global policy model through the federated averaging algorithm, and distributes the updated model to each UAV; a multi-UAV offline decision model is introduced in the pre-training stage to enable rapid adaptation to new inspection tasks; Step S7: Use a hierarchical architecture to plan the UAV trajectory. Perform global task allocation and coarse-grained waypoint planning at the top layer, build a refined Markov decision process at the bottom layer, and use deep reinforcement learning algorithm to train the flight strategy of each UAV at the local trajectory generation layer. Step S8: Monitor the trajectory tracking error of each UAV, changes in the inspection area environment, and the operating status of teammates in real time; when the trajectory tracking deviation exceeds the set threshold, or when new obstacles or new defect areas are discovered, trigger the local replanning algorithm to regenerate a flight trajectory that is adapted to the current scenario.

[0008] Preferably, in step S2, the specific steps for acquiring the optimal reference image are as follows: Step S211: Perform rapid spatial coarse screening based on tower or section identifiers to narrow the search scope to historical datasets of the same target; Step S212: Perform refined spatiotemporal matching within the target dataset to calculate the pose of the current UAV's i-th observation. The pose corresponding to the j-th candidate reference image in the historical record The spatial distance is calculated using the following formula: ; in, The spatial distance between the current drone pose and each historical pose. For the current drone i The spatial position vector of the pose at the time of the next observation. The first in history j The spatial position vector of the pose corresponding to each candidate reference image. i It is an integer. j It is an integer; The similarity of the viewpoint between the current drone pose and each historical pose is calculated using the following formula: ; in, The similarity between the current drone pose and the poses in historical records is used to measure the viewpoint similarity. For the current drone i The viewpoint attitude parameters of the pose at the time of the next observation. The first in history j Viewpoint attitude parameters corresponding to the pose of each candidate reference image; Step S213: Select spatial distance Less than the preset threshold and The largest record, and its associated images This is the optimal reference image.

[0009] Preferably, the real-time image and the reference image are geometrically registered using the SuperPoint algorithm and the SuperGlue algorithm, as follows: Step S221: First, use the flight path information to complete the coarse alignment of the images; Step S222: Extract key points and descriptors from the image using the SuperPoint algorithm; Step S223: Output precise matching point pairs using the SuperGlue algorithm; Step S224: Estimate the homography matrix and perform perspective transformation on the real-time image to achieve geometrically accurate alignment with the reference image. The formula is as follows: ; in,( x , y () represents the homogeneous coordinates of a pixel in the real-time image. These are the homogeneous pixel coordinates in the reference image coordinate system after homography matrix transformation. H This is the homography transformation matrix.

[0010] Preferably, in step S3, the same feature encoding network as in the offline stage is used to extract features from the aligned image pairs to obtain a real-time difference feature map, as shown in the following formula: ; in, For real-time images after alignment, This is the real-time feature representation extracted by the feature coding network. For the corresponding reference image, For the parameter set of the feature coding network, For parameters A defined feature encoding mapping function; A two-stage target detection method is used to locate defects, specifically including: Step S311: Process the real-time difference feature map using the YOLOv8 model to output a set of defect candidate boxes, the defect category corresponding to each candidate box, and the classification confidence; for complex defects such as broken strands, corrosion, insulator damage, and irregular contours, output the pixel-level segmentation mask of the defects through the YOLOv8 instance segmentation branch. Step S312: For each candidate box The defective local image patch is cropped out, and fine-grained defect classification and severity regression are performed to achieve accurate defect localization and differential identification. Step S313, the constraints for target detection include: Intersection over Union (IoU) constraint: in, b This represents the actual defect bounding box. For the model's predicted bounding box, The area of ​​the intersection of the two frames. The area of ​​the union of the two frames; Box regression loss constraints: ;in, cThe coordinates of the center of the true bounding box. The coordinates of the center of the prediction box. For Euclidean distance, d This is the diagonal length of the smallest bounding box of the two frames. v This is a measure of the difference in aspect ratio between the two frames. As a weighted term; Classification loss constraints: ;in, This represents the model's predicted probability for the true class. To balance the weights for each category, Focus on parameters for difficult cases.

[0011] The preferred method for identifying defects is as follows: Step S321: Based on the geometrically registered and aligned historical images With current image Each step involves performing defect detection to obtain a defect set. Step S322: For each defect candidate box, crop the local defect region from the corresponding image and extract the appearance embedding vector through the feature extraction network; Step S323: Based on the Hungarian bipartite graph matching algorithm, complete the cross-period pairing of defects between two phases of images; Step S324: Perform metric learning based on triplet loss to optimize the embedding vector space. The loss function formula is as follows: ; in, Anchor point defect characteristics, Positive sample features for the same defect These are negative sample features for different defects. m This is the interval threshold; Step S325: Based on the defect pairing results, output differentiated labels by combining the size, category, and severity variation rules.

[0012] Preferably, in step S4, the drone swarm is in t The dynamic undirected graph at time t is as follows: ; in, For drone swarms at all times t A dynamic undirected communication graph, For a moment t The set of nodes, For a moment t The set of edges; design The graph attention network layer at the head processes neighbor node information, using the following formula: ; in, The unnormalized attention coefficient For attention vectors, For learnable weight matrix, For nodes i The input feature vector, For nodes j The input feature vector, It is a linear rectified activation function with leakage slope; The neighbor attention coefficients are normalized using softmax, as shown in the following formula: ; in, To normalize attention weights, For nodes i The set of neighboring nodes, l This is the index of the neighboring nodes in the summation.

[0013] Preferably, in step S5, each drone calculates its task completion rate based on its own status, using the following formula: ; in, The competitive value of tasks to be assigned. For the drone's current location to the 1st m The distance cost to the location of each task objective. For the first i The capabilities or revenue indicators of operating drones. For the first i The current mission load level of the drone. The distance term weighting coefficient. The weighting coefficient for the ability item. This represents the weighting coefficient for the load penalty term.

[0014] Preferably, in step S6, the global policy model is updated using federated averaging, as shown in the following formula: ; in, This represents the increment of global model parameters obtained after federated aggregation. This represents the number of local samples used by the drone node in this round of training. This represents the total number of local samples from all participating nodes in this round. This refers to the increment of model parameters calculated by the drone node based on local data. N The total number of drone nodes participating in this round of federal average aggregation; In the pre-training phase, a multi-UAV Transformer offline decision-making model is introduced to achieve multi-UAV task allocation, coverage inspection, dynamic replanning, and rapid adaptation to new routes. Specifically, this includes: Step S61: Construct an offline dataset Each trajectory Among them, actions in multi-drone scenarios , For the first i The first drone t Constant inspection actions, Let t be the observed state. To perform the action at time t The immediate reward obtained afterward, where T is the total time step of a single trajectory; Step S62: Calculate the amount of return received. And use it as a conditional input to the Transformer model, the calculation formula is: ; in, As a discount factor, The instantaneous reward at time k; Step S63: Train the Transformer model in an autoregressive manner to maximize the log-likelihood of actions and achieve behavior cloning and reward condition constraints. Step S64: Input historical trajectories and target returns as context sequences, and output adapted multi-machine inspection actions through the Transformer model.

[0015] Preferably, in step S7, a hierarchical architecture is used to plan the drone trajectory, as follows: At the top level, based on the topology of the power transmission line and the inspection mission objectives, global task allocation and coarse-grained waypoint planning are carried out, and the macro-level mission sequence of each UAV is output. At the underlying level, a refined Markov decision process is constructed for each UAV. The state space of the UAV's refined Markov decision process integrates its own dynamic state, local environmental perception information, and task context, while the action space consists of continuous control commands, as shown in the following formula: ; in, The start time for trajectory optimization or control execution. The termination time for trajectory optimization or control execution. To control the energy or amplitude of the input; In the local trajectory generation layer, a deep reinforcement learning algorithm is used to train the flight strategy of each UAV. The objective function formula is as follows: ; in, Let the overall policy objective function be the one to be optimized. For discount factor terms, For instant rewards, Here is the entropy regularization coefficient. For the policy function, For environmental conditions, Let entropy be the policy function. The weights for the physical smoothness loss, The physical smoothness loss function, To generate the expectation operator for the trajectory distribution, The state-action trajectory is obtained by sampling according to the strategy.

[0016] Preferably, in step S8, the local replanning algorithm explicitly considers the positional uncertainty of the UAV and obstacles, and adopts chance-constrained planning or robust model predictive control methods, so that the generated trajectory can still meet the safety constraints under the worst-case error.

[0017] Therefore, this invention proposes a method for differentiated identification of power line defects for UAV swarm inspection, with the following advantages: (1) This invention constructs a spatiotemporal index and geometric registration mechanism for historical benchmark data, and combines a twin feature comparison network with shared weights to realize cross-time period representation alignment and differential feature extraction. It completes defect localization through a two-stage target detection method and identifies defects through a heterogeneous identification method. It can effectively filter interference caused by lighting, season, and equipment texture, significantly reduce the false alarm rate of minor defects in transmission lines under complex backgrounds, and improve the reliability of defect identification.

[0018] (2) The present invention constructs a distributed collaborative perception and dynamic communication graph modeling framework, which only requires the exchange of minimum necessary state information between UAVs, reducing the dependence on inter-UAV communication. It can still maintain the consensus of the cluster situation in the environment of strong interference of power transmission lines and low bandwidth communication. Combined with the distributed consensus auction algorithm, it realizes dynamic task allocation and has good elasticity and fault tolerance.

[0019] (3) This invention introduces federated averaging and UAV offline decision-making model to realize online adaptive and edge collaborative learning of UAV clusters, which not only protects local data privacy, but also realizes the collaborative evolution of cluster knowledge; at the same time, when the cluster is deployed to a new power transmission line, it can quickly optimize the strategy with only a small amount of local interactive data, reducing the need for new scene labeled data and the adaptation time.

[0020] (4) This invention deeply integrates multiple technologies such as computer vision, UAV swarm collaborative control, and trajectory planning to form an integrated unmanned autonomous inspection solution for power transmission lines. It solves the technical defects of single links in the existing technology and provides key technical support for the digital and intelligent transformation of power inspection.

[0021] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0022] Figure 1 This is a flowchart of a method for differentiating defects in power transmission lines for unmanned aerial vehicle (UAV) swarm inspection, according to the present invention. Detailed Implementation

[0023] To make the technical solutions, advantages, and objectives of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below. The described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the described embodiments of the present invention without creative effort are within the protection scope of the present invention.

[0024] Unless otherwise defined, the technical or scientific terms used in this invention shall have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.

[0025] Example 1 like Figure 1 The diagram shows a flowchart of a method for differentiated identification of power line defects for UAV swarm inspection according to the present invention. The specific steps are as follows: Step S1: The UAV cluster collects real-time data packets of the power transmission line through airborne equipment, and at the same time uses airborne multi-source sensors to obtain local environmental observations of the inspection area. Step S2: Construct a spatiotemporal index database, where each historical baseline data record can be represented as... Real-time data packets collected by drones The optimal reference image is obtained through spatial coarse screening and fine spatiotemporal matching. Then, the real-time image and the reference image are geometrically registered using the SuperPoint and SuperGlue algorithms. For the first k The target identifier number corresponding to each data item Spatial location vectors recorded for historical reference. Viewpoint / attitude parameters recorded for historical baselines. As historical benchmark images, This serves as the timestamp for the acquisition of historical baseline images. Feature representation vectors extracted from historical reference images For the current real-time data packet number i The spatial location vector of the second observation For the first i The viewpoint / attitude parameters corresponding to this observation. For the first i Real-time images from the second observation. For the first i The timestamp of the data collection for this observation; The specific steps for obtaining the optimal reference image are as follows: Step S211: Perform rapid spatial coarse screening based on tower or section identifiers to narrow the search scope to historical datasets of the same target; Step S212: Perform refined spatiotemporal matching within the target dataset to calculate the pose of the current UAV's i-th observation. The pose corresponding to the j-th candidate reference image in the historical record The spatial distance is calculated using the following formula: ; in, The spatial distance between the current drone pose and each historical pose. For the current drone i The spatial position vector of the pose at the time of the next observation. The first in history j The spatial position vector of the pose corresponding to each candidate reference image. i It is an integer. j It is an integer; The similarity of the viewpoint between the current drone pose and each historical pose is calculated using the following formula: ; in, The similarity between the current drone pose and the poses in historical records is used to measure the viewpoint similarity. For the current drone i The viewpoint attitude parameters of the pose at the time of the next observation. The first in history j Viewpoint attitude parameters corresponding to the pose of each candidate reference image; Step S213: Select spatial distance Less than the preset threshold and The largest record, and its associated images This is the optimal reference image.

[0026] Geometric registration between real-time and reference images is performed using the SuperPoint and SuperGlue algorithms, as detailed below: Step S221: First, use the flight path information to complete the coarse alignment of the images; Step S222: Extract key points and descriptors from the image using the SuperPoint algorithm; Step S223: Output precise matching point pairs using the SuperGlue algorithm; Step S224: Estimate the homography matrix and perform perspective transformation on the real-time image to achieve geometrically accurate alignment with the reference image. The formula is as follows: ; in,( x , y () represents the homogeneous coordinates of a pixel in the real-time image. These are the homogeneous pixel coordinates in the reference image coordinate system after homography matrix transformation. H This is the homography transformation matrix.

[0027] The obtained homography transformation matrix H is used to perform perspective transformation on the real-time image to generate a corrected image that is geometrically aligned with the reference image.

[0028] Step S3: Based on the registration results obtained in step S2, extract the features of the registered image pairs and generate a real-time difference feature map; use a two-stage target detection method and a differential identification method to accurately locate defects and identify defect types in the real-time difference feature map, and output the defect identification results. During online inspection, the system performs a differential process that explicitly compares the real-time acquired data stream with historical data. Upon receiving a real-time data packet, the system first retrieves the optimal reference image from the historical reference database based on its spatiotemporal metadata. Then, the real-time image is spatially aligned using a geometric registration module. The same feature encoding network as in the offline phase is used to extract features from the aligned image pair, as shown in the following formula: ; in, For real-time images after alignment, This is the real-time feature representation extracted by the feature coding network. For the corresponding reference image, For the parameter set of the feature coding network, For parameters A defined feature encoding mapping function; A two-stage target detection method is used to locate defects, specifically including: Step S311: Process the real-time difference feature map using the YOLOv8 model to output a set of defect candidate boxes, the defect category corresponding to each candidate box, and the classification confidence; for complex defects such as broken strands, corrosion, insulator damage, and irregular contours, output the pixel-level segmentation mask of the defects through the YOLOv8 instance segmentation branch. Step S312: For each candidate box The defective local image patch is cropped out, and fine-grained defect classification and severity regression are performed to achieve accurate defect localization and differential identification. Step S313, the constraints for target detection include: Intersection over Union (IoU) constraint: in, b This represents the actual defect bounding box. For the model's predicted bounding box, The area of ​​the intersection of the two frames. The area of ​​the union of the two frames; Box regression loss constraints: ;in, c The coordinates of the center of the true bounding box. The coordinates of the center of the prediction box. For Euclidean distance, d This is the diagonal length of the smallest bounding box of the two frames. v This is a measure of the difference in aspect ratio between the two frames. As a weighted term; Classification loss constraints: ;in, This represents the model's predicted probability for the true class. To balance the weights for each category, Focus on parameters for difficult cases.

[0029] The differential identification method identifies defects, and the steps are as follows: Step S321: Based on the geometrically registered and aligned historical images With current image Each step involves performing defect detection to obtain a defect set. Step S322: For each defect candidate box, crop the local defect region from the corresponding image and extract the appearance embedding vector through the feature extraction network; Step S323: Based on the Hungarian bipartite graph matching algorithm, complete the cross-period pairing of defects between two phases of images; Step S324: Perform metric learning based on triplet loss to optimize the embedding vector space. The loss function formula is as follows: ; in, Anchor point defect characteristics, Positive sample features for the same defect These are negative sample features for different defects. m This is the interval threshold; Step S325: Based on the defect pairing results, output differentiated labels by combining the size, category, and severity variation rules.

[0030] Step S4: The drones exchange minimum necessary state information through a chain or mesh communication network, and model the drone swarm as a dynamic undirected graph; A lightweight and robust collaborative sensing and communication framework is constructed to enable the cluster to maintain situational consensus even under weak communication conditions. Each UAV acquires local environmental observations, including its own state information such as position, velocity, attitude, and battery level, through onboard multi-source sensors such as visible / infrared cameras, lidar, IMU, and GNSS. UAVs exchange minimal necessary state information through a low-bandwidth, intermittent chain or mesh communication network, enabling the cluster to maintain situational consensus. t The time step is modeled as a dynamic undirected graph: ; in, For drone swarms at all times t A dynamic undirected communication graph, For a moment t The set of nodes, For a moment t The set of edges; design The graph attention network layer at the head processes neighbor node information, using the following formula: ; in, The unnormalized attention coefficient For attention vectors, For learnable weight matrix, For nodes i The input feature vector, For nodes j The input feature vector, It is a linear rectified activation function with leakage slope; The neighbor attention coefficients are normalized using softmax, as shown in the following formula: ; in, To normalize attention weights, For nodes i The set of neighboring nodes, l This is the index of the neighboring nodes in the summation.

[0031] Step S5: Model the inspection task as a task to be assigned. Each drone calculates the task completion rate based on its own status and assigns the task through a distributed consensus auction algorithm. Each drone calculates its mission completion rate based on its own status, using the following formula: ; in, The competitive value of tasks to be assigned. For the drone's current location to the 1st m The distance cost to the location of each task objective. For the first i The capabilities or revenue indicators of operating drones. For the first i The current mission load level of the drone. The distance term weighting coefficient. The weighting coefficient for the ability item. This represents the weighting coefficient for the load penalty term.

[0032] Through a distributed consensus auction algorithm, the cluster quickly converges to a conflict-free, near-optimal task allocation scheme with limited communication. When a sudden event occurs, it triggers local or global task reallocation to ensure the continuity of task execution.

[0033] Step S6: Each UAV collects inspection experience data locally, updates the global policy model through the federated averaging algorithm, and distributes the updated model to each UAV; a multi-UAV offline decision model is introduced in the pre-training stage to enable rapid adaptation to new inspection tasks; During execution, to ensure the cluster can continuously adapt to new environments and tasks after deployment and to achieve knowledge sharing, each drone collects new flight experience data locally. This data is then periodically federated and averaged under the coordination of an edge server or a lead drone to update the global policy model before being distributed to all drones. The formula for updating the global policy model through federated averaging is as follows: ; in, This represents the increment of global model parameters obtained after federated aggregation. This represents the number of local samples used by the drone node in this round of training. This represents the total number of local samples from all participating nodes in this round. This refers to the increment of model parameters calculated by the drone node based on local data. N The total number of drone nodes participating in this round of federal average aggregation; In the pre-training phase, a multi-UAV Transformer offline decision-making model is introduced to achieve multi-UAV task allocation, coverage inspection, dynamic replanning, and rapid adaptation to new routes. Specifically, this includes: Step S61: Construct an offline dataset Each trajectory Among them, actions in multi-drone scenarios , For the first i The first drone t Constant inspection actions, Let t be the observed state. To perform the action at time t The immediate reward obtained afterward, where T is the total time step of a single trajectory; Step S62: Calculate the amount of return received. And use it as a conditional input to the Transformer model, the calculation formula is: ; in, As a discount factor, The instantaneous reward at time k; Step S63: Train the Transformer model in an autoregressive manner to maximize the log-likelihood of actions and achieve behavior cloning and reward condition constraints. Step S64: Input historical trajectories and target returns as context sequences, and output adapted multi-machine inspection actions through the Transformer model.

[0034] Step S7: Use a hierarchical architecture to plan the UAV trajectory. Perform global task allocation and coarse-grained waypoint planning at the top layer, build a refined Markov decision process at the bottom layer, and use deep reinforcement learning algorithm to train the flight strategy of each UAV at the local trajectory generation layer. A hierarchical architecture is used to plan the drone trajectory, as detailed below: At the top level, based on the topology of the power transmission line and the inspection mission objectives, global task allocation and coarse-grained waypoint planning are carried out, and the macro-level mission sequence of each UAV is output. At the underlying level, a refined Markov decision process is constructed for each UAV. The state space of the UAV's refined Markov decision process integrates its own dynamic state, local environmental perception information, and task context, while the action space consists of continuous control commands, as shown in the following formula: ; in, The start time for trajectory optimization or control execution. The termination time for trajectory optimization or control execution. To control the energy or amplitude of the input; The constraint formulas are as follows: ; in, For the first i A drone at all times t Spatial location, To represent the time interval [ t 0 ,t f It holds true at any time within the specified period. For a moment t Assigned to the i The feasible flight space or safe jurisdiction area for drones. For a moment t Assigned to the j The feasible flight space or safe jurisdiction area for drones. This indicates that the regions of two different drones do not overlap at any given time; In the local trajectory generation layer, a deep reinforcement learning algorithm is used to train the flight strategies of each UAV. The flight strategy network architecture embeds prior physical information and considers UAV dynamics differential constraints, such as implicitly constraining the acceleration continuity of the generated trajectory through specific connections or loss functions between network layers. Flight strategy training is conducted in a high-fidelity simulation environment containing abundant obstacles, wind disturbances, and different lighting conditions. The objective function formula for training is as follows: ; in, Let the overall policy objective function be the one to be optimized. For discount factor terms, For instant rewards, Here is the entropy regularization coefficient. For the policy function, For environmental conditions, Let entropy be the policy function. The weights for the physical smoothness loss, The physical smoothness loss function, To generate the expectation operator for the trajectory distribution, The state-action trajectory is obtained by sampling according to the strategy.

[0035] Step S8: Monitor the trajectory tracking error of each UAV, changes in the inspection area environment, and the operating status of teammates in real time; when the trajectory tracking deviation exceeds the set threshold, or when new obstacles or new defect areas are discovered, trigger the local replanning algorithm to regenerate a flight trajectory that is adapted to the current scenario.

[0036] The local replanning algorithm explicitly considers the positional uncertainties of the UAV and obstacles, and adopts chance-constrained programming or robust model predictive control methods. The generated trajectory can still meet safety constraints even under worst-case error.

[0037] The invention will be further illustrated below with specific examples.

[0038] 1. Low-disturbance inspection of ecologically sensitive areas in cross-regional power transmission corridors.

[0039] In power transmission lines traversing ecologically sensitive areas such as nature reserves and water source protection areas, the terrain is complex and the vegetation is dense, making operational disturbances strictly limited. Traditional manual inspections are prone to damaging surface vegetation and have numerous blind spots. Using the method of this invention, a swarm of drones can autonomously complete low-altitude, full-coverage inspections along the transmission line with almost zero disturbance to surface vegetation. It can accurately identify hidden defects such as broken conductor strands and spontaneous insulator explosions, and automatically differentiate the severity and trends of these defects, truly achieving interference-free, routine, and intelligent inspections of power lines in ecologically sensitive areas.

[0040] 2. Corrosion resistance assessment of transmission lines in coastal salt spray corrosion areas.

[0041] In coastal mudflats, ports, and island areas, power transmission lines are constantly exposed to high salt spray, high humidity, and typhoons, resulting in extremely rapid corrosion rates of fittings and towers. Manual inspections are hampered by limited visibility and high safety risks. This invention utilizes a drone swarm that can maintain stable flight under salt spray and strong wind conditions, automatically acquiring multi-angle images and comparing them with historical data. This allows for the accurate identification of early signs of corrosion defects such as rust on fittings and contamination of insulators, and the differentiated classification of corrosion severity, providing maintenance departments with precise decisions for corrosion-resistant maintenance.

[0042] 3. Monitoring the deformation of power transmission lines in mining subsidence areas and areas prone to geological disasters.

[0043] In areas prone to geological disasters such as coal mine goafs, landslide areas, and permafrost degradation areas, pole foundations are highly susceptible to uneven settlement or tilting, and the distance between the power line and the ground changes dynamically. Traditional monitoring methods are costly and lack real-time performance. The method of this invention allows for the periodic deployment of drone swarms to autonomously inspect risk sections. Through precise comparison of multi-temporal images, it automatically identifies potential deformation hazards such as pole tilting, foundation settlement, and abnormal conductor sag, quantifies the magnitude of changes, and marks high-risk locations, effectively supporting disaster early warning and power line relocation decisions.

[0044] 4. Parallel and coordinated inspection of multiple lines in densely populated ultra-high voltage transmission corridors.

[0045] In the "power highways" where multiple UHV AC / DC lines are densely arranged along corridors, the corridors are narrow, tower types are diverse, and the electromagnetic environment is extremely complex. Single-machine inspections are prone to omissions and pose a risk of collision. Using the method of this invention, a swarm of drones can autonomously divide tasks according to multi-machine collaborative formation, and conduct layered and zoned parallel inspections of each line within the corridor. Even under strong electromagnetic interference, it can maintain inter-machine communication and safe spacing, and simultaneously complete the detection and differentiated labeling of defects on multiple lines, significantly improving the full inspection efficiency and operational safety of dense corridors.

[0046] 5. Early warning and inspection of icing on railway lines in high-altitude, cold, and snow-covered areas.

[0047] In winter, power transmission lines in high-altitude and cold mountainous areas are frequently affected by ice rain and rime ice disasters. The thickness of ice accretion on the lines changes extremely rapidly, the coverage of manual observation points is very limited, and the risk to personnel traveling in ice and snow is extremely high. Using the method of this invention, a swarm of drones can autonomously patrol multiple times in low-temperature and icy environments. By comparing historical ice-free baseline images with current icy images, it can automatically identify the type of icing and estimate the thickness and distribution range of icing, providing timely early warning of uneven icing development trends and real-time data support for the switching of de-icing devices and load adjustment.

[0048] 6. Emergency inspection of transmission lines at night or under low visibility conditions.

[0049] When a power outage occurs suddenly at night, or under low visibility conditions such as fog, haze, or dust storms, it is necessary to quickly locate the fault. However, manual night patrols are fraught with difficulties and limited visibility, rendering traditional visible light inspection methods largely ineffective. Using the method of this invention, a drone swarm equipped with multi-source sensors can perform autonomous nighttime inspections. Even under low visibility and weak communication conditions, it maintains formation coordination, automatically identifies and locates abnormal phenomena such as thermal defects and corona discharges, and compares historical images to determine the degree of fault evolution. This provides repair teams with precise fault location navigation and defect classification guidance.

[0050] It is worth noting that all contents not described in detail in this invention are existing technologies and are well known to those skilled in the art.

[0051] Therefore, this invention provides a method for differentiated identification of transmission line defects for UAV swarm inspection. By deeply integrating technologies such as multi-source data spatiotemporal registration, prior-constrained defect differential detection, lightweight distributed swarm collaboration, and hierarchical trajectory planning with physical prior embedding, it effectively reduces the false alarm rate of minor defects in transmission lines under complex backgrounds, and significantly improves the communication robustness, trajectory planning adaptability, and flight safety of UAV swarms in highly interference transmission corridor environments. This provides key technical support for the digital and intelligent transformation of power inspection.

[0052] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A method for differentiated identification of power transmission line defects for UAV swarm inspection, characterized in that, The steps are as follows: Step S1: The UAV cluster collects real-time data packets of the power transmission line through airborne equipment, and at the same time uses airborne multi-source sensors to obtain local environmental observations of the inspection area. Step S2: Construct a spatiotemporal index database, perform spatial coarse screening and fine spatiotemporal matching on the real-time data packets collected by the UAV, select the optimal reference image, and perform geometric registration between the real-time image and the optimal reference image. Step S3: Based on the registration results obtained in step S2, extract the features of the registered image pairs and generate a real-time difference feature map; use a two-stage target detection method and a differential identification method to accurately locate defects and identify defect types in the real-time difference feature map, and output the defect identification results. Step S4: The drones exchange minimum necessary state information through a chain or mesh communication network, and model the drone swarm as a dynamic undirected graph; Step S5: Model the inspection task as a task to be assigned. Each drone calculates the task completion rate based on its own status and assigns the task through a distributed consensus auction algorithm. Step S6: Each drone collects inspection experience data locally, updates the global policy model through the federated averaging algorithm, and distributes the updated model to each drone. In the pre-training phase, a multi-UAV offline decision-making model is introduced to enable rapid adaptation to new inspection tasks; Step S7: Use a hierarchical architecture to plan the UAV trajectory. Perform global task allocation and coarse-grained waypoint planning at the top layer, build a refined Markov decision process at the bottom layer, and use deep reinforcement learning algorithm to train the flight strategy of each UAV at the local trajectory generation layer. Step S8: Monitor the trajectory tracking error of each UAV, changes in the inspection area environment, and the operating status of teammates in real time; when the trajectory tracking deviation exceeds the set threshold, or when new obstacles or new defect areas are discovered, trigger the local replanning algorithm to regenerate a flight trajectory that is adapted to the current scenario.

2. The method for differentiated identification of power line defects for UAV swarm inspection according to claim 1, characterized in that, In step S2, the specific steps for acquiring the optimal reference image are as follows: Step S211: Perform rapid spatial coarse screening based on tower or section identifiers to narrow the search scope to historical datasets of the same target; Step S212: Perform refined spatiotemporal matching within the target dataset to calculate the pose of the current UAV's i-th observation. The pose corresponding to the j-th candidate reference image in the historical record The spatial distance is calculated using the following formula: ； in, The spatial distance between the current drone pose and each historical pose. For the current drone i The spatial position vector of the pose at the time of the next observation. The first in history j The spatial position vector of the pose corresponding to each candidate reference image. i It is an integer. j It is an integer; The similarity of the viewpoint between the current drone pose and each historical pose is calculated using the following formula: ； in, The similarity between the current drone pose and the poses in historical records is used to measure the viewpoint similarity. For the current drone i The viewpoint attitude parameters of the pose at the time of the next observation. The first in history j Viewpoint attitude parameters corresponding to the pose of each candidate reference image; Step S213: Select spatial distance Less than the preset threshold and The largest record, and its associated images This is the optimal reference image.

3. The method for differentiated identification of power line defects for UAV swarm inspection according to claim 2, characterized in that, Geometric registration between real-time and reference images is performed using the SuperPoint and SuperGlue algorithms, as detailed below: Step S221: First, use the flight path information to perform coarse image alignment; Step S222: Extract key points and descriptors from the image using the SuperPoint algorithm; Step S223: Output precise matching point pairs using the SuperGlue algorithm; Step S224: Estimate the homography matrix and perform perspective transformation on the real-time image to achieve geometrically accurate alignment with the reference image. The formula is as follows: ； in,( x , y () represents the homogeneous coordinates of a pixel in the real-time image. These are the homogeneous pixel coordinates in the reference image coordinate system after homography matrix transformation. H This is the homography transformation matrix.

4. The method for differentiated identification of power line defects for UAV swarm inspection according to claim 3, characterized in that, In step S3, the same feature encoding network as in the offline stage is used to extract features from the aligned image pairs to obtain the real-time difference feature map, as shown in the following formula: ； in, For real-time images after alignment, This is the real-time feature representation extracted by the feature coding network. For the corresponding reference image, For the parameter set of the feature coding network, For parameters A defined feature encoding mapping function; A two-stage target detection method is used to locate defects, specifically including: Step S311: Process the real-time difference feature map using the YOLOv8 model to output a set of defect candidate boxes, the defect category corresponding to each candidate box, and the classification confidence; for complex defects such as broken strands, corrosion, insulator damage, and irregular contours, output the pixel-level segmentation mask of the defects through the YOLOv8 instance segmentation branch. Step S312: For each candidate box The defective local image patch is cropped out, and fine-grained defect classification and severity regression are performed to achieve accurate defect localization and differential identification. Step S313, the constraints for target detection include: Intersection over Union (IoU) constraint: in, b This represents the actual defect bounding box. For the model's predicted bounding box, The area of ​​the intersection of the two frames. The area of ​​the union of the two frames; Box regression loss constraints: ;in, c The coordinates of the center of the true bounding box. The coordinates of the center of the prediction box. For Euclidean distance, d This is the diagonal length of the smallest bounding box of the two frames. v This is a measure of the difference in aspect ratio between the two frames. As a weighted term; Classification loss constraints: ;in, This represents the model's predicted probability of the true class. To balance the weights for each category, Focus on parameters for difficult cases.

5. The method for differentiated identification of power line defects for UAV swarm inspection according to claim 4, characterized in that, The differential identification method identifies defects, and the steps are as follows: Step S321: Based on the geometrically registered and aligned historical images With current image Each step involves performing defect detection to obtain a defect set. Step S322: For each defect candidate box, crop the local defect region from the corresponding image and extract the appearance embedding vector through the feature extraction network; Step S323: Based on the Hungarian bipartite graph matching algorithm, complete the cross-period pairing of defects between two phases of images; Step S324: Perform metric learning based on triplet loss to optimize the embedding vector space. The loss function formula is as follows: ； in, Anchor point defect characteristics, Positive sample features for the same defect These are negative sample features for different defects. m This is the interval threshold; Step S325: Based on the defect pairing results, output differentiated labels by combining the size, category, and severity variation rules.

6. The method for differentiated identification of power line defects for UAV swarm inspection according to claim 5, characterized in that, In step S4, the drone swarm is in t The dynamic undirected graph at time t is as follows: ； in, For drone swarms at all times t A dynamic undirected communication graph. For a moment t The set of nodes, For a moment t The set of edges; design The graph attention network layer at the head processes neighbor node information, using the following formula: ； in, The unnormalized attention coefficient For attention vectors, For learnable weight matrix, For nodes i The input feature vector, For nodes j The input feature vector, It is a linear rectified activation function with leakage slope; The neighbor attention coefficients are normalized using softmax, as shown in the following formula: ； in, To normalize attention weights, For nodes i The set of neighboring nodes, l This is the index of the neighboring nodes in the summation.

7. A method for differentiated identification of power line defects for UAV swarm inspection according to claim 6, characterized in that, In step S5, each drone calculates its mission completion rate based on its own status, using the following formula: ； in, The competitive value of tasks to be assigned. For the drone's current location to the 1st The distance cost to the location of each task objective. For the first i The capability or revenue indicators of operating drones. For the first i The current mission load level of the drone. The distance term weighting coefficient. This refers to the weighting coefficient for the ability item. This represents the weighting coefficient for the load penalty term.

8. A method for differentiated identification of power line defects for UAV swarm inspection according to claim 7, characterized in that, In step S6, the global policy model is updated using federated averaging, as shown in the following formula: ； in, This represents the increment of global model parameters obtained after federated aggregation. This represents the number of local samples used by the drone node in this round of training. This represents the total number of local samples from all participating nodes in this round. This refers to the increment of model parameters calculated by the drone node based on local data. N The total number of drone nodes participating in this round of federal average aggregation; In the pre-training phase, a multi-UAV Transformer offline decision-making model is introduced to achieve multi-UAV task allocation, coverage inspection, dynamic replanning, and rapid adaptation to new routes. Specifically, this includes: Step S61: Construct an offline dataset Each trajectory Among them, actions in multi-drone scenarios , For the first i The first drone t Constant inspection actions, Let t be the observed state. To perform the action at time t The immediate reward obtained afterward, where T is the total time step of a single trajectory; Step S62: Calculate the amount of return received. And use it as a conditional input to the Transformer model, the calculation formula is: ； in, As a discount factor, The instantaneous reward at time k; Step S63: Train the Transformer model in an autoregressive manner to maximize the log-likelihood of actions and achieve behavior cloning and reward condition constraints. Step S64: Input historical trajectories and target returns as context sequences, and output adapted multi-machine inspection actions through the Transformer model.

9. A method for differentiated identification of power line defects for UAV swarm inspection according to claim 8, characterized in that, In step S7, a hierarchical architecture is used to plan the drone trajectory, as detailed below: At the top level, based on the topology of the power transmission line and the inspection mission objectives, global task allocation and coarse-grained waypoint planning are carried out, and the macro-level mission sequence of each UAV is output. At the underlying level, a sophisticated Markov decision process is built for each drone. The state space of the refined Markov decision process for UAVs integrates its own dynamic state, local environmental perception information, and task context, while the action space consists of continuous control commands, as shown in the following formula: ； in, The start time for trajectory optimization or control execution. The termination time for trajectory optimization or control execution. To control the energy or amplitude of the input; In the local trajectory generation layer, a deep reinforcement learning algorithm is used to train the flight strategy of each UAV. The objective function formula is as follows: ； in, Let the overall policy objective function be the one to be optimized. For discount factor terms, For instant rewards, Here is the entropy regularization coefficient. For the policy function, For environmental conditions, Let entropy be the policy function. The weights for the physical smoothness loss, The physical smoothness loss function, To generate the expectation operator for the trajectory distribution, The state-action trajectory is obtained by sampling according to the strategy.

10. A method for differentiated identification of power line defects for UAV swarm inspection according to claim 9, characterized in that, In step S8, the local replanning algorithm explicitly considers the positional uncertainty of the UAV and obstacles, and adopts chance-constrained planning or robust model predictive control methods. The generated trajectory can still meet the safety constraints even under the worst-case error.

Images