Multi-unmanned aerial vehicle electric power inspection method based on real-time attention mechanism and fixed time

Abstract

The invention relates to a multi-unmanned aerial vehicle electric power inspection method based on a real-time attention mechanism and fixed time. An unmanned aerial vehicle performs data acquisition on 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 of the power transmission line; and each unmanned aerial vehicle in the unmanned aerial vehicle formation adopts a neighbor communication rule, and the unmanned aerial vehicles adopt a fixed time distributed control strategy according to the end position coordinates to realize autonomous routing inspection of the formation. Simulation shows that the target identification precision is improved, and stable operation of the control system is proved. 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 a solid and highly accurate basic guarantee is provided for a power line detection system of an intelligent power grid.

Description

TECHNICAL FIELD

[0001] The application relates to a multi-unmanned aerial vehicle power inspection method based on real-time attention mechanism and fixed time. BACKGROUND

[0002] As a key infrastructure of the power system, the operation state of the transmission line directly affects the safety and power supply reliability of the power grid. In order to ensure the safety of the transmission line, a periodic inspection system needs to be implemented. The current detection system mainly adopts a mode combining manual inspection and fixed sensor monitoring. Manual inspection relies on professional personnel to climb the tower or walk along the line to investigate, and has problems such as high labor intensity, high safety risk (especially in complex terrains such as high mountains and gorges), and blind area of detection. Although the fixed sensor can realize local state monitoring, its layout is limited by the power corridor environment, and has inherent defects such as high equipment maintenance cost, limited monitoring range, and susceptibility to electromagnetic interference.

[0003] With the development of unmanned aerial vehicle technology, it is more and more common to use unmanned aerial vehicles for line inspection. The use of unmanned aerial vehicles can effectively reduce the inspection labor intensity of operation and maintenance personnel, especially in areas with poor line operation environment, which can greatly improve the line inspection efficiency and has broad application prospects. At present, the target detection and autonomous inspection system based on unmanned aerial vehicles for power lines has become a research hotspot in the field of intelligent operation and maintenance of the power system.

[0004] In the early research of power line target detection, most methods are based on traditional image processing algorithms, using image features such as color, shape and texture, which have obvious shortcomings such as poor generalization ability and large computing power. With the development of deep learning, many excellent target detection algorithms have emerged, which have greatly promoted the development of the field of power line target detection. Existing target detection algorithms are divided into two stages and one stage. The two-stage target detection algorithm is represented by the R-CNN series, which first generates candidate regions, and then predicts the bounding box and class of the target object in these regions. This method is usually more accurate than the single-stage method, but requires more computing resources and time. The one-stage target detection algorithm is represented by the YOLO series, which does not need to define the region to be detected in advance, but directly predicts the bounding box and class of the target object. This method is simple and efficient, but the accuracy is usually slightly lower than that of the two-stage method. SUMMARY

[0005] The purpose of the present application is to provide a multi-unmanned aerial vehicle power inspection method based on real-time attention mechanism and fixed time combined with target detection and unmanned aerial vehicle formation.

[0006] The application adopts the following technical solutions:

[0007] A multi-unmanned aerial vehicle power inspection method based on real-time attention mechanism and fixed time, the unmanned aerial vehicle will collect data of the power transmission line through the visual sensor, detect the power transmission line by using the improved YOLOv5s model, calculate the length of the power transmission line by using the length of the detection frame, and output the end position; each unmanned aerial vehicle in the unmanned aerial vehicle formation adopts the rule of neighbor communication, and the unmanned aerial vehicle adopts the fixed time distributed control strategy according to the end position coordinates, so as to realize the autonomous inspection of the formation.

[0008] Further, the improved YOLOv5s model is improved on the basis of the original YOLOv5s backbone network, improved Neck and using loss function EIOU instead of original loss function CIOU.

[0009] Further, the improved backbone network is introduced after the original backbone network CBAM attention mechanism.

[0010] Further, the improved Neck is to perform convolution and up-sampling again at the 18th layer of the original network FPN layer, so as to increase a shallow feature extraction network with 160x160 pixels in the structure and use the original network PAN layer for down-sampling to obtain a shallow network structure, and the feature maps in the backbone network layer after the convolution layer are fused with the feature maps in the same scale in the PAN layer after the convolution layer, so as to generate the final feature map input to the detection head.

[0011] Further, the loss function EIOU is: wherein, and are the width and height of the detection frame; and are the width and height of the real frame; and are the width and height of the minimum frame covering the two frames; IOU is the overlapping degree of the predicted area and the real area; represents the Euclidean distance.

[0012] Further, the guidance law of the fixed time distributed control strategy is: wherein, is a constant, and the speed in the direction is defined as and . respectively represent the linear speed of the unmanned aerial vehicle in the X-axis direction, the linear speed in the Y-axis direction and the angular speed. represent the X-axis position error of the unmanned aerial vehicle and the power transmission line, representing the unmanned aerial vehicle Y-axis position error with the power transmission line, representing the unmanned aerial vehicle heading angle error with the power transmission line.

[0013] further, tracking error , and defined as: .

[0014] representing the unmanned aerial vehicle position error with the power transmission line, representing the unmanned aerial vehicle heading angle error with the power transmission line, representing the unmanned aerial vehicle linear velocity error with the power transmission line, representing the unmanned aerial vehicle angular velocity error with the power transmission line.

[0015] The beneficial effects of the present application are: the present application aims at the problem of autonomous power line inspection, and proposes a solution combining target detection and unmanned aerial vehicle formation. Through unmanned aerial vehicle formation, the problems of poor image quality of single unmanned aerial vehicle and high cost of wide-angle camera are solved. Through the measures of adding attention mechanism, cross-layer fusion and loss function improvement, the feature loss is reduced and the detection accuracy is improved. The fixed time formation guidance rule of unmanned aerial vehicle is designed to solve the problem of unmanned aerial vehicle formation control. Comparative simulation research is carried out to verify the accuracy of the proposed enhanced model and evaluate the stability and robustness of the guidance law. Flight experiment further confirms the effectiveness of the algorithm. The present application lays a theoretical foundation for autonomous power line inspection and monitoring of unmanned aerial vehicle formation. BRIEF DESCRIPTION OF DRAWINGS

[0016] Figure 1 is the overall framework diagram of unmanned aerial vehicle autonomous line inspection.

[0017] Figure 2 is a schematic diagram of the distributed communication method of unmanned aerial vehicle formation.

[0018] Figure 3 is the structure diagram of CBAM attention mechanism.

[0019] Figure 4 is the improved backbone network model.

[0020] Figure 5 is the structure diagram of Neck feature extraction.

[0021] Figure 6 is part of the power transmission line image data set.

[0022] Figure 7 Convergence speed of the prediction box for different loss functions.

[0023] Figure 8 Visualization results for comparison of different algorithms.

[0024] Figure 9 Visualization of the line length for target detection and output.

[0025] Figure 10 Tracking trajectory of the UAV under known wind conditions.

[0026] Figure 11 Linear velocity of each UAV.

[0027] Figure 12 Angular velocity of each UAV.

[0028] Figure 13 X-axis error of each UAV.

[0029] Figure 14 Y-axis error of each UAV.

[0030] Figure 15 Heading angle error of each UAV.

[0031] Figure 16 Comparison of trajectories generated by two control methods.

[0032] Figure 17 Comparison of linear velocity changes of two control methods.

[0033] Figure 18 Comparison of angular velocity changes of two control methods.

[0034] Figure 19 Comparison of system tracking errors of two control methods.

[0035] Figure 20 Comprehensive inspection results of three UAVs on the transmission line at different time points (t = 1s, t = 18s, t = 36s) and different directions.

[0036] Figure 21 is a photo of the experimental site. DETAILED DESCRIPTION

[0037] The overall framework for autonomous line inspection using a drone using visual sensors, as proposed in this invention, is shown in Figure 1. The drone's data acquisition portion is shown in the second part of Figure 1. As shown in the figure, the drone uses its visual sensors to detect transmission lines. The detected transmission lines are displayed in an image, and the detection results generate a detection frame. The length of the detection frame can be calculated. The distance between the drone and the transmission line is constant. Using the principle of similar triangles, the length of the transmission line is determined, and the position of its end is output.

[0038] The drone formation uses distributed communication, while each drone follows neighbor communication rules, as shown in Figure 2. Based on the output end coordinates, the drones fly to the next target detection location. Furthermore, a fixed-time distributed control strategy is used to control the drone formation, enabling autonomous inspections.

[0039] Consider a transmission line patrol formation system composed of N drones. The kinematic model of a UAV can be expressed as: (1) in, Indicates drone location, represents the heading angle, and Indicates drone The linear and angular velocities, are the wind speeds, assuming they have an additive effect on the drone's speed, Represents the effect of wind speed on the angular velocity of the drone. is bounded, that is and .

[0040] When performing line patrol tasks, each drone moves at a reference speed. Due to the limitations of physical properties, the control input of each drone is constrained. (2) in, and are all assumed known constants.

[0041] The error between the transmission line and each UAV is expressed as: (3) in, It's a drone The relative position of the target, Indicates the location of the transmission line, is the heading angle. Indicates drone Position error of the transmission line, is represented as heading angle error.

[0042] For any initial state and , find a guidance law such that (4) where , and are some constants tending to zero, which are affected by wind disturbance. When the disturbance is known, .

[0043] The error system (3) is derived to obtain the following kinematic error model: (5) where is represented as the relative position between the UAV and the transmission line, is the relative heading angle, and are the linear and angular velocities of the target, respectively. It is assumed that the velocity of the target is limited by and , where and are known constants.

[0044] Lemma 1: Consider the following nonlinear system: (6) where , , and are continuous functions, and there exists a unique local solution for all initial conditions. For any , there exist and . Therefore, the origin is an equilibrium point of system (6). By ignoring the influence of the interconnection term , the system can be decomposed into isolated subsystems: (7) Consider the equilibrium point of system (6) and assume that there exists a positive definite and decreasing Lyapunov function satisfying: (8) (9) (10) When , and are constants. Where, , is positive definite and continuous. If the matrix , whose elements are: (11) The matrix is The origin of the system (6) is fixed-time stable.

[0045] The stability time of the system (6) is: (12) Where, , .

[0046] Assuming constants and known wind and are determined constants. The kinematic error model (5) can be re-expressed as: (13) Where, and represent the linear and angular velocities of the transmission line relative to the wind, respectively.

[0047] I. Improvement of YOLOv5s model (1) CBAM attention mechanism Attention mechanism is widely used to deal with the complex challenges of UAV background, because it can make the model better select relevant features and suppress irrelevant information. By integrating the Convolutional Block Attention Module (CBAM) into the backbone network, the feature extraction capability of the model is enhanced, which enables it to pay attention to subtle features in different spatial positions. This improves the model's ability to capture key information from images, thereby improving the accuracy and robustness of the task.

[0048] CBAM is an attention mechanism designed to improve the performance of Convolutional Neural Network (CNN). CBAM is composed of two main components: channel attention module and spatial attention module. The final attention-enhanced features are obtained by element-wise multiplication of the channel and spatial attention module output features. These enhanced features are then passed to the subsequent layers of the network, where they help suppress noise and irrelevant information while preserving important features. Experimental results show that the effect of these modules is better when integrating the channel dimension first and then the spatial dimension.

[0049] Figure 3 For the CBAM attention mechanism structure diagram, the feature map is input into the channel attention module, and the output result is multiplied by to obtain the result , and then is input into the spatial attention module, and the obtained result is multiplied by to obtain the output feature map . The calculation formula is: (14) (15) In the formula, is the input feature map, is the channel attention weight, is the spatial attention weight, is the number of channels of the feature map, and are the height and width of the feature map, indicates the dot product between two vectors.

[0050] Placing the CBAM attention mechanism after the backbone network layer makes the backbone network suppress the influence of complex environment and noise on feature extraction, improves the accuracy of feature information extraction, and makes the model have a global view. Figure 4 is the improved backbone network model.

[0051] (2) Neck In the original model, the image is extracted by the backbone network, and then the FPN and PAN networks in the Neck are used to obtain the feature map {p3, p4, p5} input to the detection head, where p3 feature map has higher spatial resolution and less semantic information, while p5 feature map has higher-level semantic information but lower spatial resolution. After multiple downsampling and convolution layer processing of the network, the image will lose more feature information.

[0052] In order to enhance the effect of feature extraction, the 18th layer in FPN is again subjected to convolution and up-sampling operation, so that a shallow feature extraction network with 160x160 pixels is added to the structure, i.e. Figure 5 L2 in the PAN layer, and the original model is down-sampled as the input of target detection, and the final feature map input to the detection head is {p2, p3, p4, p5}. Among them, L3, L4, L5 are the FPN layers of the original network, and P3, P4, P5 are the PAN layers of the original network. On the basis of the original network, the shallow network structure of L2 and P2 is added, and L2 and P2 have the same structure as the original network. The detection head adds a detection head with a resolution of 160x160 on the basis of the original model, as shown in the red area of Figure 5 . The location information of the shallow layer is fully utilized, the sensitivity of the model to small targets is enhanced, and the accuracy of small target detection is improved. In order to reduce the loss of feature information of the model, the feature maps in the Backbone network layer after the convolution layer are cross-layer fused with the feature maps in the PAN layer after the convolution layer of the same scale. The fused feature map contains more location information and semantic information, as shown by the blue dashed line in Figure 5 . Cross-layer feature fusion enhances the model's understanding of the image and improves the accuracy of target detection.

[0053] (3) Loss function The function of the loss function is to measure the difference between the predicted box and the real box, update the parameters according to the difference, help the model learn to accurately detect and locate target objects, and improve the precision and accuracy of detection. The loss function used by the YOLOv5s model is CIOU, and the calculation formula is: (16) is the weight function, measures the consistency of the aspect ratio, the intersection over union of the predicted box and the real box, the Euclidean distance between the center points of the predicted box and the real box, the diagonal distance of the minimum box covering the two boxes.

[0054] The CIOU loss function considers the shape information of the target box, the diagonal distance of the target box and the direction, etc., so that the predicted box is more consistent with the real box. However, when the aspect ratio of the predicted box and the real box is the same, the long and wide penalty term is always 0; there is also and cannot be increased or decreased at the same time, which will hinder the effective optimization of the model similarity.

[0055] To solve the problem of CIOU, the aspect ratio is separated on the basis of CIOU, and EIOU Loss is proposed. The penalty term of EIOU is to separate the influence factor of the aspect ratio to calculate the length and width of the target frame and the predicted frame respectively on the basis of the penalty term of CIOU, which solves the problem of CIOU. The calculation formula of the loss function EIOU is: (17) and are the width and height of the detection frame; and are the width and height of the real frame; and are the width and height of the minimum frame covering the two frames.

[0056] Using the loss function EIOU to replace the boundary box regression loss function CIOU in the original model solves the problem of CIOU, accelerates the convergence of the predicted frame, and improves the regression accuracy of the predicted frame.

[0057] II. Fixed time distribution control strategy In order to achieve the goal in (4), the unmanned aerial vehicle The following guidance law is constructed: (18) where is a constant, and the velocity in the direction of and is defined as

[0058] The tracking error , and is defined as: (19) By substituting the guidance law (18) into the error dynamics system (13), a complex closed-loop system can be formed: (20) The generated closed-loop system represents a nonlinear composite system that describes the state of the stratum. According to Lemma 1, the system can be decomposed into two isolated subsystems and two interconnected terms, resulting in the following structure.

[0059] The first isolated subsystem and the first interconnected term: (21) (22).

[0060] The second isolated subsystem and the second interconnected term: (23) (24).

[0061] III. Proof of fixed-time stabilization for nonlinear composite systems Step 1: For the first isolated subsystem, consider the following Lyapunov function. (25) (26) where, , , .

[0062] (27) where, , .

[0063] (28) where, .

[0064] Step 2: For the second isolated subsystem, consider the following Lyapunov function. (29) (30) where, , , .

[0065] (31) where, , .

[0066] (32) where, .

[0067] According to equation (11), the matrix is: (33) The origin of the matrix system (13) can achieve fixed-time stabilization.

[0068] IV. Comparative experiment of improved YOLOv5s model ​

[0069] The processor of the application is Inter Core i7-13620H, the graphics card is NVIDIA GeForce RTX 4050Laptop GPU, 16GB memory, and the operating system is Windows11, 64-bit. Based on the Pytorch2.0 deep learning framework, Python3.8 is used. The training parameter setting: the training batch is 8, the training round is 300, the image input size is set to 640x640, the initial learning rate is 0.0100, the self-made data set is used to carry out the ablation experiment of the improved algorithm, and the comparative experiment between different modules and different algorithms is carried out, so as to verify the effectiveness of the improved algorithm.

[0070] In order to ensure the smooth progress of the experiment and fully consider the experimental results, we collect a set of power transmission line image data set. This set of high-quality data set is collected from real scenes, covering various backgrounds such as villages, fields and water surfaces, containing a total of 3462 images. For the experiment, 2980 images are selected as the training set, 320 images are selected as the validation set, and 162 images are selected as the test set. Part of the data set is shown in Figure 6.

[0071] The precision P, recall R and average precision mAP are used as the indexes for evaluating each model, and their formulas are as follows: (34) (35) (36)

[0072] Among them, is the number of positive samples predicted by the positive sample, is the number of positive samples predicted by the negative sample, is the number of negative samples predicted by the positive sample, is the class, is the average accuracy of a single class. On the basis of the above indexes, the experiment additionally uses mAP50:90, Params, GLFOPs as the performance indexes of the model. mAP50:90 represents the average value of the average precision calculated at different IOU thresholds between 0.50 and 0.95, which fully reflects the performance of the model under different detection difficulties; Params is the parameter quantity of the model, which is used to measure the consumption of computing memory resources; GLFOPs is 10 billion floating point operations per second, which is used to measure the computational complexity when training the model.

[0073] (1) CBAM attention mechanism module comparison In order to reduce the environmental impact and improve the target detection accuracy, the CBAM attention mechanism module is introduced. There are two ways to add the module to the Backbone. The first way is to place CBAM behind each C3 module in the Backbone network, which can make the attention mechanism see the local features and perform attention at each layer to share the learning pressure. The second way is to place CBAM at the end of the Backbone network, which can make the attention mechanism see the entire feature map of the Backbone part and have a global view. This experiment will verify the effectiveness of the two ways.

[0074] Table 1 CBAM attention mechanism module comparison results .

[0075] As can be seen from Table 1, compared with the original model, the accuracy and mAP values of way 1 are lower than those of the original model. The attention mechanism of way 1 pays excessive attention to the noise and redundant information of the input, while ignoring the truly important features and focusing on irrelevant parts, which leads to a decrease in the generalization ability of the network and thus a decrease in the accuracy and mAP values. As can be seen from Table 1, the accuracy and mAP values of way 2 are significantly improved compared with the original model. Way 2 can make the model focus more on the extraction of the detection target, and the influence of the surrounding complex environment and noise on the module is significantly reduced.

[0076] (2) Comparison of loss functions In order to analyze the ability of the loss function EIOU to improve the accuracy of the model, based on YOLOv5s, CIOU, DIOU, and EIOU are added to the model respectively, and experiments are conducted on the dataset. The experimental results are shown in Table 2.

[0077] As can be seen from Table 2, EIOU has higher accuracy than other loss functions, and DIOU has the lowest accuracy. EIOU is optimized based on CIOU and DIOU loss functions, which 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 minimum bounding box width and height, thereby improving the regression accuracy of the prediction box. Therefore, EIOU is used instead of CIOU loss function to improve the performance of the model.

[0078] Table 2 Comparison of loss function results .

[0079] From Figure 7 it can be seen that compared with the original model CIOU loss function, the improved model loss function EIOU has faster convergence speed and lower loss value, which indicates that the EIOU loss function accelerates the convergence of the prediction box.

[0080] (3) Comparison of ablation experiments 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.

[0081] As can be seen from Table 3, the proposed improved algorithm improves the accuracy of object detection. After integrating various modules, each module contributes differently to the performance of the model. Among them, the improvement of the Neck layer brings the greatest performance improvement, with an accuracy increase of 4.1% and an mAP increase of 2.9%. In addition, replacing the CIOU loss function with EIOU improves the accuracy, mAP, and FPS. This indicates that EIOU not only improves the detection accuracy but also slightly improves the detection speed, which means that EIOU speeds up the convergence of the prediction box and improves the regression accuracy. Compared with the original model, the improved YOLOv5s model has an accuracy increase of 6.1%, a recall rate increase of 0.6%, an average precision (mAP) increase of 4.7%, and a floating-point operation per second (GFLOPs) increase of 2.5. Although the detection speed per image has slightly decreased, the final frame rate per second (FPS) remains at 98 frames per second. Although it has slightly decreased compared with the original algorithm, it still meets the real-time performance requirements. Overall, the results show that the proposed algorithm enhances the performance of the original model while maintaining real-time capabilities.

[0082] Table 3 Ablation experiment .

[0083] (4) Performance experiment comparison To evaluate the performance of the improved YOLOv5s model, we compared it with several target 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 the present application outperforms the compared detectors in terms of recall rate, mAP, and mAP: 0.5~0.95. The algorithm of the present application achieves higher accuracy with fewer parameters and floating-point operations, and is more advantageous than 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 well in terms of detection accuracy and speed, making it very suitable for real-time detection of transmission lines in video images.

[0084] Table 4 Comparison experiment .

[0085] (5) Target detection experiment comparison Target detection experiments were conducted on three different types of images, including images where the target is easily recognizable, images where the target and background are well integrated, and images where the target and background are cluttered. Through these experiments, the performance of different models under various conditions was evaluated.

[0086] In the first set of experiments, the performance of different lightweight network models was compared. Compared with other lightweight networks, models with greater network depth (such as YOLOv7-tiny and YOLOv8s) can extract more complex features. However, these complex features sometimes lead to misleading recognition results, affecting the accuracy of the model.

[0087] In the second set of experiments, due to the good integration of the target and background, some models failed to successfully detect the target. For example, YOLOv3 and YOLOv7-tiny failed to detect the target, while YOLOv5s incorrectly identified the line insulator as a power line, and YOLOv8s incorrectly identified the small road as a power line. This indicates that when the target and background are highly integrated, the detection performance of some models is greatly affected.

[0088] In the third set of experiments, we conducted target detection on images with relatively cluttered backgrounds. The results showed that YOLOv3, YOLOv5s, YOLOv7-tiny, and YOLOv8s all had different degrees of false positives or false negatives. This indicates that in complex backgrounds, all models face certain recognition difficulties.

[0089] In the fourth set of experiments, models with greater network depth (such as YOLOv7-tiny and YOLOv8s) still managed to extract complex features, but this could also lead to recognition errors. In contrast, YOLOv3 and YOLOv5s models had false negatives in this experiment. This experiment further verifies the differences in model performance under different scenarios.

[0090] Through the comparison of experimental results, it can be seen that the improved model shows strong adaptability under various background conditions and can effectively complete the target detection task. Although there are certain false positives and false negatives, the overall performance still indicates the effectiveness and robustness of the model.

[0091] Five, UAV formation simulation experiment Consider a problem of four UAVs forming a formation to track a power line. The kinematics of each UAV 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. The velocity constraint is and By default, all variables are assumed to use unit.

[0092] For this simulation, the wind disturbance is known. The states of all UAVs and tracked targets at the initial moment are listed. The target's motion state can be described as and . Wind is defined as , , In line with the Gaussian distribution, , The control parameters were determined through repeated experiments and were , , , , , , , , , , the simulation duration is 200 s and the time step is 0.02 s.

[0093] Table 5 Initial state .

[0094] Figure 10 shows the tracking trajectory of the UAVs under known wind conditions. The trajectory shows smooth convergence and rapid stabilization. Figures 11 and 12 depict the linear velocity and angular velocity of each UAV, respectively, both of which show a trend of approaching the target and maintaining stability. Figures 13, 14, and 15 are specifically used to show the tracking errors of the guidance law stability parameters, which tend to zero after a period of fluctuation. The simulation results show that the proposed control strategy (18) can effectively enable the UAV formation to complete the tracking of the target.

[0095] VI. Reliability and Effectiveness Verification of UAV Formation Control Strategy In order to verify the reliability and effectiveness of the control strategy proposed in this paper, it is compared with the method in the literature [J. Jia, X.Chen, W. Wang, K. Wu, M. Xie, “Distributed observer-based finite-time controlof moving target tracking for UAV formation,” in ISA Transactions, vol. 140,PP.1-17, 2023] (the literature is denoted as literature [1]).

[0096] The initial state, position, and wind speed of the UAVs, and other simulation conditions are the same as those described above. For ease of exposition, let us assume that the target motion is a curved motion, and define its motion as: , . Figure 16 The comparison of trajectories generated by the two control methods is shown in Figure 17, which shows that both control methods successfully track a linearly moving target. However, it is clear from the guidance law (18) that the tracking speed of the target will be faster. Figure 17 shows the variation of linear velocity, from which it can be seen that the convergence speed of this algorithm is slightly faster than that of the other algorithm, and the jitter produced under the influence of wind is minimal. Figure 18 shows the variation of angular velocity, including both convergence speed and jitter amplitude, and the guidance law proposed in this application is superior to the algorithm in [1]. Figure 19 shows the comparison of system tracking errors, which provides further evidence of the effectiveness and superiority of the proposed method.

[0097] To verify the effectiveness of the proposed algorithm, we conducted field data collection and algorithm verification on a 220 kV power transmission line, with good results. Figure 21 shows a photo of the experimental site. The three groups of images shown in Figure 20 correspond to the comprehensive inspection of the power transmission line by three UAVs at different time points (t=1s, t=18s, t=36s) and different directions.

Claims

1. A multi-UAV power inspection method based on real-time attention mechanism and fixed time, characterized in that: The drones will collect data on the transmission lines through visual sensors, detect the transmission lines using the improved YOLOv5s model, calculate the length of the transmission lines using the length of the detection frame, and output their end positions; each drone in the drone formation adopts the rules of neighbor communication, and the drones use a fixed-time distributed control strategy based on the end position coordinates to achieve autonomous inspections of the formation.

2. The multi-UAV power inspection method based on real-time attention mechanism and fixed time according to claim 1 is characterized in that: The improved YOLOv5s model is based on the original YOLOv5s by improving the backbone network, improving Neck, and replacing the original loss function CIOU with the loss function EIOU.

3. The multi-UAV power inspection method based on real-time attention mechanism and fixed time according to claim 2 is characterized in that: The improved backbone network introduces the CBAM attention mechanism after the original backbone network.

4. The multi-UAV power inspection method based on real-time attention mechanism and fixed time according to claim 3 is characterized in that: The improved Neck is to perform convolution and upsampling again on the 18th layer of the original network FPN layer, so that a 160×160 pixel shallow feature extraction network is added to the structure and the original network PAN layer is used for downsampling to obtain a shallow network structure. The feature map of the backbone network layer that has passed the convolution layer is fused with the feature map of the same scale that has passed the convolution layer in the PAN layer across layers to generate the feature map that is finally input to the detection head.

5. The multi-UAV power inspection method based on real-time attention mechanism and fixed time according to claim 4 is characterized in that: The loss function EIOU is: in, and is the width and height of the detection box; and is the width and height of the real frame; and is the width and height of the minimum box covering the two boxes; IOU is the overlap between the predicted area and the real area; represents the Euclidean distance.

6. The multi-UAV power inspection method based on real-time attention mechanism and fixed time according to claim 5 is characterized in that: The guidance law of the fixed-time distributed control strategy is: in, is a constant, and The velocities in the directions are defined as and ; Represented as drones Linear velocity in the X-axis, linear velocity in the Y-axis, and angular velocity; Indicates drone The X-axis position error with the transmission line, Indicates drone Y-axis position error with the transmission line, Indicates drone The error with the transmission line is the heading angle.

7. The multi-UAV power inspection method based on real-time attention mechanism and fixed time according to claim 6 is characterized in that: Tracking Error 、 and Defined as: Indicates drone Position error with the transmission line, Indicates drone The heading angle error with the transmission line, Indicates drone Linear velocity error with the transmission line, Indicates drone Angular velocity error with the transmission line.

Images