Multi-uav power inspection method based on real-time attention mechanism and fixed time

Abstract

The application relates to a multi-unmanned aerial vehicle power line inspection method based on real-time attention mechanism and fixed time, wherein the unmanned aerial vehicle collects data of a power transmission line through a visual sensor, detects the power transmission line by using an improved YOLOv5s model, calculates the length of the power transmission line by using the length of a detection frame, and outputs the end position; each unmanned aerial vehicle in the unmanned aerial vehicle formation adopts a neighbor communication rule, and the unmanned aerial vehicle adopts a fixed time distributed control strategy according to the end position coordinates, so that autonomous inspection of the formation is realized. Simulation shows that the application not only improves target recognition accuracy, but also proves stable operation of the control system. Flight experiments in a real environment show that the method has practicability and effectiveness for autonomous power line inspection of multiple unmanned aerial vehicles, and provides solid and highly accurate basic guarantee for a power line detection system of a smart power grid.

Description

Technical Field

[0001] This invention relates to a multi-UAV power line inspection method based on a real-time attention mechanism and fixed time. Background Technology

[0002] As a critical infrastructure of the power system, the operational status of transmission lines directly affects the security of the power grid and the reliability of power supply. To ensure the safety of transmission lines, a periodic inspection system is necessary. The current monitoring system mainly adopts a combination of manual inspection and fixed sensor monitoring. Manual inspection relies on professionals climbing towers or walking along the line, which has problems such as high labor intensity, high safety risks (especially in complex terrains such as high mountains and canyons), and blind spots. Although fixed sensors can achieve local condition monitoring, their deployment is limited by the power corridor environment, and they have inherent drawbacks such as high equipment maintenance costs, limited monitoring range, and susceptibility to electromagnetic interference.

[0003] With the development of drone technology, the use of drones for power line inspection is becoming increasingly common. The use of drones can effectively reduce the labor intensity of maintenance personnel, especially in areas with harsh operating environments, greatly improving inspection efficiency and showing broad application prospects. Currently, drone-based power line target detection and autonomous inspection systems have become a research hotspot in the field of intelligent power system operation and maintenance.

[0004] In early research on power line target detection, most methods were based on traditional image processing algorithms, using image features such as color, shape, and texture. These methods suffered from poor generalization ability and high computational costs. With the development of deep learning, many excellent target detection algorithms have emerged, greatly promoting the development of the field of power line target detection. Existing target detection algorithms are divided into two-stage and one-stage methods. Two-stage target detection algorithms, represented by the R-CNN series, first generate candidate regions and then predict the bounding boxes and categories of target objects within these regions. This method is generally more accurate than single-stage methods but requires more computational resources and time. One-stage target detection algorithms, represented by the YOLO series, do not require pre-defining the regions to be detected; instead, they directly predict the bounding boxes and categories of target objects. This method is simple and efficient, but its accuracy is usually slightly lower than that of two-stage methods. Summary of the Invention

[0005] The purpose of this invention is to provide a multi-UAV power line inspection method that combines target detection with UAV formation based on a real-time attention mechanism and fixed time.

[0006] The present invention adopts the following technical solution:

[0007] A multi-UAV power line inspection method based on real-time attention mechanism and fixed-time is proposed. The UAVs will collect data on the power transmission line through visual sensors, detect the power transmission line using an improved YOLOv5s model, calculate the length of the power transmission line using the length of the detection box, and output its end position. Each UAV in the UAV formation adopts the neighbor communication rule, and the UAVs use a fixed-time distributed control strategy based on the end position coordinates to achieve autonomous inspection of the formation.

[0008] Furthermore, the improved YOLOv5s model is based on the original YOLOv5s, with improvements to the backbone network and Neck, and the use of the loss function EIOU to replace the original loss function CIOU.

[0009] Furthermore, the improved backbone network is achieved by introducing the CBAM attention mechanism after the original backbone network.

[0010] Furthermore, the improved Neck involves performing convolution and upsampling again on the 18th layer of the original network's FPN layer, adding a 160×160 pixel shallow feature extraction network to the structure. The original network's PAN layer is then used for downsampling to obtain a shallow network structure. The feature maps from the backbone network layers that have passed through convolutional layers are fused with the feature maps of the same scale from the PAN layer that have passed through convolutional layers across layers to generate the final feature map input to the detection head.

[0011] Furthermore, the loss function EIOU is:

[0012]

[0013] in, and The width and height of the detection frame; and The width and height of the actual bounding box; and The width and height of the minimum bounding box that covers the two boxes; IOU is the overlap between the predicted region and the ground truth region; This indicates Euclidean distance.

[0014] Furthermore, the guidance law of the fixed-time distributed control strategy is:

[0015]

[0016] in, It is a constant. and The velocities in the directions are defined as follows: and . These are respectively referred to as drones Linear velocity along the X-axis, linear velocity along the Y-axis, and angular velocity. Indicates drone The X-axis position error of the transmission line, Indicates drone The position error relative to the Y-axis of the transmission line, Indicates drone The error is the heading angle, which is the same as that of the power transmission line.

[0017] Furthermore, tracking error , and Defined as:

[0018] .

[0019] Indicates drone Positional error relative to transmission lines Indicates drone The error is the heading angle, which is different from that of the transmission line. Indicates drone Error with the linear velocity of the transmission line, Indicates drone The angular velocity error relative to the transmission line.

[0020] The beneficial effects of this invention are as follows: This invention proposes a solution combining target detection and UAV swarming for the problem of autonomous power line inspection. By using UAV swarming, the problems of poor image quality from a single UAV and high cost of wide-angle cameras are solved. Through the addition of attention mechanisms, cross-layer fusion, and improved loss functions, feature loss is reduced and detection accuracy is improved. A fixed-time UAV swarming guidance law is designed to solve the problem of UAV swarm control. Comparative simulation studies were conducted to verify the accuracy of the proposed enhancement model and evaluate the stability and robustness of the guidance law. Flight experiments further confirmed the effectiveness of the algorithm. This invention lays the theoretical foundation for autonomous power line inspection and monitoring using UAV swarming. Attached Figure Description

[0021] Figure 1 This is a general framework diagram for autonomous line inspection by drones.

[0022] Figure 2 This is a schematic diagram of a distributed communication method for drone formations.

[0023] Figure 3 This is a structural diagram of the CBAM attention mechanism.

[0024] Figure 4 This is the improved backbone network model.

[0025] Figure 5 The structure diagram for Neck feature extraction.

[0026] Figure 6 This is a dataset of images of a portion of power transmission lines.

[0027] Figure 7 The effect of different loss functions on the convergence speed of the prediction box.

[0028] Figure 8 A visualization of the comparison results of different algorithms.

[0029] Figure 9 To perform target detection and output a visualization of the line length.

[0030] Figure 10 This is a diagram showing the tracking trajectory of a drone under known wind conditions.

[0031] Figure 11 Linear velocity graph for each drone.

[0032] Figure 12 An angular velocity graph for each drone.

[0033] Figure 13 The X-axis error diagram for each drone.

[0034] Figure 14 The Y-axis error diagram for each drone.

[0035] Figure 15 A heading angle error diagram for each UAV.

[0036] Figure 16 A comparison diagram of the trajectories generated by the two control methods.

[0037] Figure 17 This is a comparison graph showing the changes in linear velocity between the two control methods.

[0038] Figure 18 This is a comparison chart showing the changes in angular velocity between the two control methods.

[0039] Figure 19 This is a comparison chart of the system tracking errors of the two control methods.

[0040] Figure 20 The results are a comprehensive inspection of the power transmission line conducted by three drones at different time points (t=1s, t=18s, t=36s) and from different directions.

[0041] Figure 21 shows a photo of the experimental site. Detailed Implementation

[0042] The present invention proposes an overall framework for autonomous power line inspection using a visual sensor, as shown in Figure 1. The data acquisition part of the UAV is shown in the second part of Figure 1. As shown, the UAV will perform target detection on the power transmission line using the visual sensor. The image will display the detected power transmission line, and the detection result will generate a detection box. The length of the detection box can be calculated. The distance between the UAV and the power transmission line is constant. Using the principle of similar triangles, the length of the power transmission line is obtained, and its end position is output.

[0043] The drone swarm employs a distributed communication method, with each drone using a neighbor-to-neighbor communication rule, as shown in Figure 2. Based on the output end coordinates, the drones fly to the next target detection location. Simultaneously, a fixed-time distributed control strategy is used to control the drone swarm, thereby enabling autonomous inspection operations.

[0044] Consider a power transmission line inspection formation system consisting of N drones, the first... The kinematic model of a drone can be represented as:

[0045] (1)

[0046] in, Indicates drone Location, Indicates the heading angle. and Indicates drone linear velocity and angular velocity, It's wind speed; let's assume they have an additive effect on the drone's speed. This indicates the effect of wind speed on the angular velocity of the drone. It is bounded, that is and .

[0047] When performing line inspection missions, each drone moves at a reference speed. Due to physical limitations, the control inputs of each drone are constrained.

[0048] (2)

[0049] in, and These are all assumed, known constants.

[0050] The error between the transmission line and each UAV is expressed as:

[0051] (3)

[0052] in, It is a drone The relative position of the target Indicates the location of the power transmission line. This is the heading angle. Indicates drone Positional error relative to transmission lines This is expressed as heading angle error.

[0053] For any initial state and Find a guidance law. , making

[0054] (4)

[0055] in , and These are constants that tend to zero and are affected by wind disturbances. When the disturbance is known, .

[0056] The following kinematic error model is obtained by deriving the error system (3):

[0057] (5)

[0058] in Represented as drone The relative position with respect to transmission lines It is the relative heading angle. and These are the target's linear velocity and angular velocity, respectively. Assume the target's velocity is affected by... and The restrictions, among which and It is a known constant.

[0059] Lemma 1: Consider the following nonlinear system:

[0060] (6)

[0061] ,in , , and They are all continuous functions, and a unique local solution exists for all initial conditions. For any They all and Therefore, the origin This is the equilibrium point of system (6). This is achieved by ignoring interconnection terms. The system can be decomposed into the following effects: An isolated subsystem:

[0062] (7)

[0063] Consider the equilibrium point of system (6), and assume that there is a positive definite decreasing Lyapunov function. satisfy:

[0064] (8)

[0065] (9)

[0066] (10)

[0067] when hour, and All are constants. Among them, , It is positive definite and continuous. If the matrix... Its elements are:

[0068] (11)

[0069] The matrix is The origin of the matrix (6) is fixed-time stable.

[0070] The settling time of system (6) is:

[0071] (12)

[0072] in, , .

[0073] Assume constants and known wind and To determine the constants, the kinematic error model (5) can be reformulated as:

[0074] (13)

[0075] in, and These represent the linear velocity and angular velocity of the transmission line relative to the wind, respectively.

[0076] I. Improvements to the YOLOv5s Model

[0077] (1) CBAM attention mechanism

[0078] Attention mechanisms are widely used to address the complex challenges posed by drone backgrounds because they allow models to better select relevant features and suppress irrelevant information. By integrating the channel and spatial attention modules (Convolutional Block Attention Module, CBAM) into the backbone network, the model's feature extraction capabilities are enhanced, enabling it to focus on subtle features at different spatial locations. This improves the model's ability to capture key information from the image, thereby increasing the accuracy and robustness of the task.

[0079] CBAM is an attention machine designed to improve the performance of Convolutional Neural Networks (CNNs). CBAM consists of two main components: a channel attention module and a spatial attention module. The final attention-enhanced features are obtained by element-wise multiplication of the output features of the channel and spatial attention modules. These enhanced features are then passed to subsequent layers of the network, where they help suppress noise and irrelevant information while preserving important features. Experimental results show that these modules perform better when the channel dimension is integrated first, followed by the spatial dimension.

[0080] Figure 3 For the CBAM attention mechanism structure diagram, the feature map is... Input the channel attention module and compare the output with... Perform a dot product to get the result. Then Input the data into the spatial attention module and compare the results with... Perform a dot product to obtain the output feature map. The calculation formula is:

[0081] (14)

[0082] (15)

[0083] In the formula For the input feature map, For channel attention weights, For spatial attention weights, The number of channels in the feature map. and These represent the height and width of the feature map, respectively. This indicates a dot product between two vectors.

[0084] By placing the CBAM attention mechanism after the backbone network layer, the backbone network suppresses the influence of complex environment and noise on feature extraction, improves the accuracy of feature information extraction, and gives the model a global perspective. Figure 4This is the improved backbone network model.

[0085] (2) Neck

[0086] The image in the original model undergoes feature extraction via a backbone network, followed by FPN and PAN networks in the Neck module, ultimately yielding feature maps {p3, p4, p5} input to the detection head. Feature map p3 has higher spatial resolution but less semantic information, while feature map p5 possesses higher-level semantic information but lower spatial resolution. After multiple downsampling and convolutional layer processing by the network, the image loses more feature information.

[0087] To improve feature extraction efficiency, a convolutional upsampling operation is performed again at layer 18 of the FPN, adding a shallow feature extraction network of 160×160 pixels to the structure. Figure 5 The L2 layer of the original network is used as the input for object detection, and the feature maps input to the detection head are {p2, p3, p4, p5}. Here, L3, L4, and L5 are the original network's FPN layers, and P3, P4, and P5 are the original network's PAN layers. Shallow network structures L2 and P2 are added to the original network, with L2 and P2 having the same structure as the original network. A detection head with a 160×160 resolution is added to the original model, see [link to details]. Figure 5 The red area. This fully utilizes the shallow layer's positional information, enhancing the model's sensitivity to small targets and improving the accuracy of small target detection. To reduce the loss of model feature information, feature maps from the Backbone network layer after convolutional layers are fused with feature maps of the same scale from the PAN layer after convolutional layers. This results in the fused feature map containing more positional and semantic information. See [link to documentation]. Figure 5 The blue dashed line in the image represents cross-level feature fusion, which enhances the model's understanding of the image and improves the accuracy of object detection.

[0088] (3) Loss function

[0089] The purpose of the loss function is to measure the difference between the predicted bounding box and the ground truth bounding box, update parameters based on the difference, and help the model learn to accurately detect and locate target objects, thereby improving detection precision and accuracy. The YOLOv5s model uses CIOU as its loss function, calculated as follows:

[0090] (16)

[0091] It is a weighting function. The consistency of the aspect ratio is measured. The intersection-union ratio (IoU) of predicted bounding boxes and ground truth bounding boxes. The Euclidean distance between the center points of the predicted bounding box and the ground truth bounding box. The diagonal distance of the smallest box that covers two boxes.

[0092] The CIOU loss function considers multiple factors such as the shape information of the target box, the diagonal distance, and the orientation of the target box, making the predicted box more consistent with the ground truth box. However, when the aspect ratios of the predicted and ground truth boxes are the same, there is a problem where the aspect ratio penalty term is always 0; there are also issues... and The inability to simultaneously increase or decrease similarity hinders the effective optimization of the model.

[0093] To address the problems of CIOU, the aspect ratio is separated from the CIOU model, leading to the EIOU Loss. The EIOU penalty term is based on the CIOU penalty term, but the aspect ratio influence factor is calculated separately for the length and width of the target box and the predicted box, thus resolving the issues inherent in CIOU. The formula for calculating the EIOU loss function is as follows:

[0094] (17)

[0095] and The width and height of the detection frame; and The width and height of the actual bounding box; and The width and height of the smallest box that covers both boxes.

[0096] Replacing the original bounding box regression loss function CIOU with the loss function EIOU solves the problems of CIOU, accelerates the convergence of the predicted boxes, and improves the regression accuracy of the predicted boxes.

[0097] II. Fixed-Time Distributed Control Strategy

[0098] To achieve the goal in (4), a drone can be used. Construct the following guidance law:

[0099] (18)

[0100] in It is a constant. and The velocities in the directions are defined as follows: and .

[0101] Tracking error , and Defined as:

[0102] (19)

[0103] By substituting the guidance law (18) into the error dynamics system (13), a complex closed-loop system can be formed:

[0104] (20)

[0105] The resulting closed-loop system represents a nonlinear composite system describing the formation state. According to Lemma 1, the system can be decomposed into two isolated subsystems and two interrelated terms, resulting in the following structure.

[0106] The first isolated subsystem and the first interconnected item:

[0107] (twenty one)

[0108] (twenty two).

[0109] The second isolated subsystem and the second interconnected item:

[0110] (twenty three)

[0111] (twenty four).

[0112] III. Proof of Fixed-Time Stability in Nonlinear Composite Systems

[0113] Step 1: For the first isolated subsystem, consider the following Lyapunov function.

[0114] (25)

[0115] (26)

[0116] in, , , .

[0117] (27)

[0118] in, , .

[0119] (28)

[0120] in, .

[0121] Step 2: For the second isolated subsystem, consider the following Lyapunov function.

[0122] (29)

[0123] (30)

[0124] in, , , .

[0125] (31)

[0126] in, , .

[0127] (32)

[0128] in, .

[0129] According to equation (11), the matrix for:

[0130] (33)

[0131] All The origin of the matrix system (13) can achieve fixed-time stability.

[0132] IV. Comparative Experiment of Improved YOLOv5s Model

[0133] The processor used in this invention is an Intel Core i7-13620H, the graphics card is an NVIDIA GeForce RTX 4050 Laptop GPU, 16GB of RAM, and the operating system is Windows 11, 64-bit. It is based on the PyTorch 2.0 deep learning framework and uses Python 3.8. Training parameters are set as follows: training batch size of 8, training epochs of 300, image input size of 640×640, and initial learning rate of 0.0100. Ablation experiments are conducted on the improved algorithm using a self-made dataset, and comparative experiments are performed between different modules and different algorithms to verify the effectiveness of the improved algorithm.

[0134] To ensure the smooth conduct of the experiment and to fully evaluate the results, we collected a dataset of transmission line images. This high-quality dataset was collected from real-world scenes, covering various backgrounds such as villages, fields, and water surfaces, and contains a total of 3462 images. For the experiment, 2980 images were selected as the training set, 320 images as the validation set, and 162 images as the test set. Part of the dataset is shown in Figure 6.

[0135] Precision (P), recall (R), and mean AP (mAP) are used as metrics to evaluate each model, and their formulas are as follows:

[0136] (34)

[0137] (35)

[0138] (36)

[0139] in, It predicts the number of positive samples from positive samples. It predicts the number of positive samples from negative samples. It predicts the number of negative samples from positive samples. It is a category. This represents the average accuracy for a single class. In addition to the above metrics, the experiment also used mAP50:90, Params, and GLFOPs as performance metrics for the model. mAP50:90 represents the average accuracy calculated at different IOU thresholds between 0.50 and 0.95, comprehensively reflecting the model's performance under different detection difficulties; Params is the number of model parameters, used to measure the consumption of computational memory resources; GLFOPs is 1 billion floating-point operations per second, used to measure the computational complexity during model training.

[0140] (1) Comparison of CBAM attention mechanism modules

[0141] To reduce environmental impact and improve target detection accuracy, a CBAM attention mechanism module is introduced. There are two ways to add the module to the backbone: the first is to place CBAM after each C3 module in the backbone network, which allows the attention mechanism to see local features and perform attention once per layer, thus reducing the learning pressure; the second is to place CBAM at the end of the backbone network, which allows the attention mechanism to see the feature map of the entire backbone, thus having a global perspective. This experiment will verify the effectiveness of both methods.

[0142] Table 1 Comparison Results of CBAM Attention Mechanism Modules

[0143] .

[0144] As shown in Table 1, Method 1 has lower accuracy and mAP compared to the original model. Method 1's attention mechanism overemphasizes input noise and redundant information, neglecting truly important features and incorrectly focusing on irrelevant parts, leading to a decrease in the network's generalization ability and consequently lower accuracy and mAP. Table 1 also shows that Method 2 significantly improves accuracy and mAP compared to the original model. Method 2 allows the model to focus more on extracting the detection target, significantly reducing the impact of surrounding complex environments and noise on the module.

[0145] (2) Comparison of loss functions

[0146] To analyze the ability of the loss function EIOU to improve the accuracy of the model, based on YOLOv5s, CIOU, DIOU, and EIOU were added to the model and experiments were conducted on the dataset. The experimental results are shown in Table 2.

[0147] As shown in Table 2, EIOU has higher accuracy compared to other loss functions, while DIOU has the lowest accuracy. EIOU is an optimization based on the CIOU and DIOU loss functions. It compensates for the shortcomings of other loss functions by splitting the aspect ratio loss term into the difference between the predicted width and height and the width and height of the minimum bounding box, thereby improving the regression accuracy of the predicted box. Therefore, EIOU is used instead of CIOU to improve model performance.

[0148] Table 2 Comparison of Loss Functions

[0149] .

[0150] from Figure 7 It can be seen that compared with the loss function of the original model CIOU, the loss function of the improved model EIOU converges faster and has a lower loss value, indicating that the EIOU loss function accelerates the convergence of the predicted box.

[0151] (3) Comparison of ablation experiments

[0152] To evaluate the effectiveness of the proposed algorithm, an ablation study was conducted based on YOLOv5s, evaluating the performance of different modules under the same experimental conditions. The experimental results are shown in Table 3.

[0153] As shown in Table 3, the proposed improved algorithm enhances object detection accuracy. After integrating various modules, each module contributes differently to model performance. The improvement to the Neck layer yields the greatest performance boost, increasing accuracy by 4.1% and mAP by 2.9%. Furthermore, replacing the CIOU loss function with EIOU improves accuracy, mAP, and FPS. This indicates that EIOU not only improves detection accuracy but also slightly increases detection speed, meaning it accelerates the convergence of predicted bounding boxes and improves regression accuracy. Compared to the original model, the improved YOLOv5s model achieves a 6.1% increase in accuracy, a 0.6% increase in recall, a 4.7% increase in average precision (mAP), and a 2.5 increase in GFLOPs per second. Although the detection speed per image decreases slightly, the final frames per second (FPS) remains at 98 frames / second. While slightly lower than the original algorithm, this still meets real-time performance requirements. In summary, the results show that the proposed algorithm enhances the performance of the original model while maintaining real-time capabilities.

[0154] Table 3 Ablation Experiment

[0155] .

[0156] (4) Performance test comparison

[0157] To evaluate the performance of the improved YOLOv5s model, we compared it with several object detection algorithms on the same dataset, including YOLOv3, YOLOv5s, YOLOv7-tiny, and YOLOv8s. To ensure a fair comparison, all models were trained and tested under the same experimental conditions using the same parameters. As shown in Table 4, the algorithm of this invention outperforms the compared detectors in terms of recall, mAP, and mAP: 0.5–0.95. The algorithm of this invention achieves higher accuracy with fewer parameters and less floating-point computation, and is superior to deeper and more complex networks. With only 5.99 million parameters, our model is small enough to be deployed on embedded devices while still maintaining good real-time performance. The results show that the improved YOLOv5s model performs excellently in both detection accuracy and speed, making it ideal for real-time detection of transmission lines in video images.

[0158] Table 4 Comparative Experiments

[0159] .

[0160] (5) Comparison of target detection experiments

[0161] Object detection experiments were conducted on three different image types: images where the target is easily identifiable, images where the target and background are difficult to blend, and images where the target and background are cluttered. These experiments aimed to evaluate the performance of different models under various conditions.

[0162] In the first set of experiments, we compared the performance of different lightweight network models. Compared to other lightweight networks, models with greater network depth (such as YOLOv7-tiny and YOLOv8s) were able to extract more complex features. However, these complex features can sometimes lead to misleading recognition results, thus affecting the accuracy of the model.

[0163] In the second set of experiments, due to the better fusion between the target and the background, some models failed to detect the target successfully. For example, YOLOv3 and YOLOv7-tiny failed to detect the target, while YOLOv5s incorrectly identified a line insulator as a transmission line, and YOLOv8s incorrectly identified a small path as a transmission line. This indicates that when the target and background are highly fused, the detection performance of some models is significantly affected.

[0164] In the third set of experiments, we performed object detection on images with relatively cluttered backgrounds. The results showed that all four models—YOLOv3, YOLOv5s, YOLOv7-tiny, and YOLOv8s—experienced varying degrees of false positives or false negatives. This indicates that all models face certain recognition difficulties in complex backgrounds.

[0165] In the fourth set of experiments, models with greater network depth (such as YOLOv7-tiny and YOLOv8s) were still able to extract complex features, but this could also lead to recognition errors. In contrast, YOLOv3 and YOLOv5s models exhibited missed detections in this experiment. This experiment further validates the differences in model performance across different scenarios.

[0166] The results of the comparative experiments show that the improved model exhibits strong adaptability under various background conditions and can effectively complete the target detection task. Although there are some false positives and false negatives, the overall performance still demonstrates the effectiveness and robustness of the model.

[0167] V. Unmanned Aerial Vehicle (UAV) Formation Simulation Experiment

[0168] Consider a power line tracking problem involving a formation of four drones. The kinematics of each drone and the target can be described as (5). In the simulation, it is assumed that the target moves in a straight line and the wind in the environment is irregular white noise. Velocity constraints are set to... and Provided. By default, it is assumed that all variables use... unit.

[0169] For this simulation, wind disturbance is known. The states of all drones and the tracked target at the initial moment are listed. The target's motion state can be described as follows: and Wind is defined as , , It conforms to a Gaussian distribution. , The control parameters were determined through repeated experiments. , , , , , , , , , The simulation duration was 200 s, and the time step was 0.02 s.

[0170] Table 5 Initial State

[0171] .

[0172] Figure 10 shows the tracking trajectory of the UAVs under known wind conditions. The trajectory exhibits smooth convergence and rapid stabilization. Figures 11 and 12 depict the linear and angular velocities of each UAV, respectively, both showing a trend of approaching the target and maintaining stability. Figures 13, 14, and 15 are specifically used to illustrate the tracking errors of the guidance law stability parameters, which tend to zero after fluctuating for a period of time. Simulation results show that the proposed control strategy (18) can effectively enable the UAV formation to track the target.

[0173] VI. Verification of the Reliability and Effectiveness of UAV Formation Control Strategies

[0174] To verify the reliability and effectiveness of the control strategy proposed in this invention, it is compared with the method in the literature [J. Jia, X. Chen, W. Wang, K. Wu, M. Xie, “Distributed observer-based finite-time control of moving target tracking for UAV formation,” in ISA Transactions, vol. 140, PP.1-17, 2023] (hereinafter referred to as literature [1]).

[0175] The initial state, position, and wind speed of the drone are the same as those in the simulation experiment described above. For ease of explanation, let's assume the target's motion is curvilinear and define it as: , . Figure 16 The comparison of the trajectories generated by the two control methods is shown, and the results show that both control methods successfully enable the UAV formation to track a target moving in a straight line. However, it is clear from the guidance law (18) that the target tracking speed is faster. Figure 17 shows the variation of linear velocity, from which it can be seen that the convergence speed of the algorithm is slightly faster than other algorithms, and the jitter generated under the influence of wind is minimal. Figure 18 shows the variation of angular velocity, including both convergence speed and jitter amplitude. The guidance law proposed in this invention is superior to the algorithm in reference [1]. Figure 19 shows the comparison of system tracking errors, providing further evidence for the effectiveness and superiority of the proposed method.

[0176] To verify the effectiveness of the proposed algorithm, we conducted on-site data collection and algorithm verification on a 220 kV transmission line, with good results. Figure 21 shows a photograph of the experimental site. The three sets of images shown in Figure 20 correspond to the comprehensive inspection of the transmission line conducted by three drones at different time points (t=1s, t=18s, t=36s) and from different directions.

Claims

1. A multi-UAV power line inspection method based on real-time attention mechanism and fixed-time, characterized in that, The drones will collect data on the power transmission lines through visual sensors, detect the power transmission lines using an improved YOLOv5s model, calculate the length of the power transmission lines using the length of the detection box, and output the position of its end. Each drone in the drone formation adopts the neighbor communication rule, and the drones use a fixed-time distributed control strategy based on the coordinates of their end positions to achieve autonomous inspection of the formation. The improved YOLOv5s model is based on the original YOLOv5s, with improvements to the backbone network and Neck, and the use of the loss function EIOU to replace the original loss function CIOU. The improved backbone network is achieved by introducing the CBAM attention mechanism after the original backbone network. The improved Neck involves re-convolution and upsampling at the 18th layer of the original network's FPN layer, adding a 160×160 pixel shallow feature extraction network to the structure. The original network's PAN layer is then used for downsampling to obtain a shallow network structure. The feature maps from the backbone network layers that have passed through convolutional layers are fused with the feature maps of the same scale from the PAN layer that have passed through convolutional layers across layers to generate the final feature map input to the detection head.

2. The multi-UAV power line inspection method based on real-time attention mechanism and fixed-time as described in claim 1, characterized in that, The loss function EIOU is: in, and The width and height of the detection frame; and The width and height of the actual bounding box; and The width and height of the minimum bounding box that covers the two boxes; IOU is the overlap between the predicted region and the ground truth region; This indicates Euclidean distance.

3. The multi-UAV power line inspection method based on real-time attention mechanism and fixed-time according to claim 2, characterized in that, The guidance law of the fixed-time distributed control strategy is: in, It is a constant. and The velocities in the directions are defined as follows: and ; These are respectively referred to as drones Linear velocity along the X-axis, linear velocity along the Y-axis, and angular velocity; Indicates drone The X-axis position error of the transmission line, Indicates drone The position error relative to the Y-axis of the transmission line, Indicates drone The error is the heading angle, which is the same as that of the power transmission line.

4. The multi-UAV power line inspection method based on real-time attention mechanism and fixed-time according to claim 3, characterized in that, Tracking error , and Defined as: Indicates drone Positional error relative to transmission lines Indicates drone The heading angle error of the transmission line, Indicates drone Error with the linear velocity of the transmission line, Indicates drone The angular velocity error relative to the transmission line.

Images