A Transmission Line Defect Detection Method Based on YOLOv7 and Multi-UAV Collaboration Platform

,


Introduction
To ensure the safe operation of transmission lines, power departments must regularly inspect transmission lines and power systems [1].Insulators are the key components in transmission lines, and they are mainly used for electrical insulation and mechanical fxation of power systems [2].Afected by mechanical structure damage, material aging, and natural factors, the insulators often sufer from selfblast, damage, pollution, chip shedding, and other failures [3].Wire, iron wire, vine, thatch, cloth, and excrement are commonly used materials for birds to build nests, leading to insulator fashover and short-circuit fault.In extreme weather, insulator defects often lead to instability of transmission lines and even lead to power outages and safety accidents.Terefore, the detection and identifcation of transmission line defects is important.
Te traditional manual inspection methods are low inefcient and high risky and have low accuracy [4].In recent years, unmanned aerial vehicle (UAV) technology has been applied to the inspection tasks of high-voltage transmission lines and power equipment [5].By equipping various sensors and carrying appropriate object detection models, the transmission line monitoring equipment can complete the inspection tasks such as image acquisition, defect detection, and defect location for transmission lines [6].However, the UAV system is limited by the impact of fight costs, endurance, and complex environments, and the detection results rely on human subjective experiences.Since the video and image data of transmission lines can be effectively collected, UAV image acquisition, image processing [7], image fusion [8], and object detection have gradually become commonly used transmission line defect detection methods.Terefore, the main research content of this paper is to propose an efcient multi-UAV coordination system to reduce fight cost and improve data collection efciency.In addition, an accurate and reliable defect detection model is also the main research content of this paper.
Te transmission line defects detection methods based on the object detection algorithm can identify the transmission line defects, which is convenient for power workers to detect power failures early and reduce the occurrence of power accidents.Han improved the semantic feature extraction capability of the U-Net segmentation by embedding efcient channel attention (ECA-Net) module [9].Zhang reconstructed the convolutional layer of the inception network through interleaved group convolution [10], which improves the generalization of insulator defect detection.Although the abovementioned methods can identify the types of insulator fault, the location of insulator defects cannot be accurately located.Many scholars use two-stage detection methods such as the region-based convolutional neural network (R-CNN), Faster R-CNN [11], and Cen-terNet [12] to detect and locate defects and faults in transmission lines.Lei and Sui proposed an insulator and bird's nest identifcation method based on the Faster R-CNN with the regional proposal network [13].Liu et al. used the Faster R-CNN as the detector to achieve the slippage fault diagnosis of dampers in aerial images [14].Xia et al. introduced a spatial and channel attention mechanism convolutional block attention module (CBAM) into CenterNet to improve the prediction accuracy of insulator and replaced the ResNet-50 by the lightweight MobileNet to speed up detection speed [15].In addition, single shot detector (SSD) and you only look once (YOLO) are the most commonly used transmission line defect detection algorithms.Compared with the two-stage method proposed above, the methods do not have the step of extracting candidate regions through a regional proposal network (RPN), so it is fast and less accurate, and the YOLO series is a typical representative of them.Xin et al. proposed a small object detection algorithm named FA-SSD to solve the problem of the umbrella disc shedding occupying only a small proportion of an aerial image, in which deep features and shallow features are combined, and attention mechanisms are introduced [16].Li et al. proposed an SSD based on the improved dual network for the images of insulators and spacers taken by UAVs [17], which realized the high-precision detection of electrical equipment.Aiming at the problem of diferent self-blast areas and complex backgrounds, He et al. proposed the Mina-Net based on YOLOv4, which fused shallow feature maps with more detailed texture information into the feature pyramids, and used improved squeeze-and-excitation networks (SENet) to calibrate the features of diferent levels [18].Deng et al. proposed an improved YOLOv4 object detection algorithm based on the MobileNetv3 backbone and parametric rectifed linear unit (PReLU) activation function and detected insulator defects through this lightweight model [19].Xing and Chen processed images through Gaussian fltering, mosaic data enhancement, and k-means++ clustering and replaced the YOLOv4 backbone with MobileNet, which improved the detection speed and accuracy [20].Li et al. reduced the negative impact of uneven lighting on YOLOv5-based insulator defect detection results through multiscale gradient-domain guided fltering and two-dimensional adaptive gamma transform [21].Souza et al. proposed a hybrid YOLO model based on ResNet-18 and YOLOv5-X, which can detect faulty components in transmission lines by combining with a single UAV system [22].Dai solved the uncertainty problem of insulator defect detection by applying a Gaussian function in front of the inspection head of YOLOX [23].It can be seen that SDD, YOLOv4, YOLOv5, and YOLOX have all been used to detect defects and faults in transmission lines.In actual transmission line defect detection experiments, these methods have high requirements on image quality and cannot detect small-target insulators and defects in complex backgrounds, and the detection accuracy and efciency of these methods need to be further improved.
Terefore, the existing defect detection methods for insulators of high-voltage lines are still difcult to meet the practical application requirements.Te main reasons and research gaps are summarized as follows: (1) Te transmission line passes through plateaus, mountains, valleys, etc., so the power inspection covers a wide range and the terrain and background are particularly complex.However, most of the existing transmission line defect detection methods based on a single UAV only detect insulator defects or bird's nests and the situation where multiple defects occur simultaneously is ignored.(2) Due to the infuence of shooting angles, background noises, and lighting conditions, it is difcult to ensure the reliability of transmission line defect detection.In addition, the YOLOv4-based and YOLOv5-based detection methods cannot meet the requirements of precision and speed.(3) Te existing defect detection methods for transmission lines only detect a single fault or defect, such as insulator self-blast, breakdown, or bird nest.Te reasons for the abovementioned problems depend largely on the lack of image samples and the low quality of small samples.In the actual detection tasks, the same images or lens may contain multiple defects, which brings great challenges to the current transmission line fault detection.
In view of the abovementioned problems, we proposed a transmission line defect detection method based on YOLOv7 and the multi-UAV collaboration platform.Te main contributions and innovations are shown in Figure 1.
(1) To address the problems of low efciency and small search range caused by the existing research using only a single UAV to detect defects in transmission lines, we constructed a multi-UAV cooperation platform.Trough the integration of the multi-UAV module, positioning module, path planning module, and detection module, multi-UAV can not only plan fight paths autonomously through the control of the DJI mobile SDK module but also detect and locate defective components and positions through YOLOv7 and coordinate calibration.Tis platform can not only detect and locate defects well in transmission lines, but also fexibly connect several UAVs.Terefore, the proposed multi-UAV collaboration platform can not only increase the search range per unit time but also improve the defect detection reliability.(2) In order to improve detection accuracy and efciency as much as possible, we proposed a defect detection module based on YOLOv7, and it is used as a detector and integrated into the multi-UAV module through the DJI mobile SDK module.After detecting targets and defects on the transmission line, the actual location of the defects can be located through the corresponding relationship between the calibrated parameters and the coordinate system, and the problems of low accuracy caused by complex backgrounds, single UAV, and shooting lighting have been reduced.(3) To comprehensively consider the types of defects in real transmission line scenarios, we built a transmission line defect detection dataset by combining web crawler datasets, public datasets, and live shooting datasets, which contains three types of most common transmission line defect images, including cracked insulators, insulator self-blast, and bird nests, and the problems of lack of data and low quality of samples have been solved.
Te rest of this paper is arranged as follows.Materials and methods are described in Section 2, including the multi-UAV collaboration platform and the YOLOv7 detection algorithm.Te results of the transmission line defect detection are analysed in Section 3. Some useful experimental analyses and conclusions are summarized in Section 4.

Transmission Line Defect Detection Method
Based on YOLOv7  2.
Ten, we establish a suitable path-planning method for each UAV in the collaboration platform.Te fight path planning method connected to the platform can be divided into two modes; one is to divide a global fight area based on the fight path and search range, and the other is to plan and receive the fight path and area of each UAV through map annotation.In this paper, we divide fight points and missions based on the actual situation and search scope of the transmission line, using map markers.First, the geographic coordinates that the UAVs on the transmission line need to reach are sequentially marked as array coordinates.Ten, we plan specifc fight point coordinates for each specifc UAV, and all UAVs connected to the platform have corresponding paths and fight point coordinates.Under the control of the DJI mobile SDK, all UAVs within the collaboration platform fy according to preplanned fight paths, expanding the search range and improving search efciency.Te steps of the multi-UAV collaboration platform are as follows.
(1) Division of search area on transmission lines.We divide the search area of the 4 G or 5 G signal base station where the transmission line is located as a reference centre.First, we are centred on a 5 G signal base station and preliminarily determined by the UAV's maximum fight distance as the radius.In the divided search area, each UAV gradually shoots video along the transmission line from the inside out.
When the range of each UAV is accurately determined, we can plan the specifc fight path of each UAV based on the perspective of each UAV's camera and maximum fight distance.Generally speaking, those UAVs that can fy longer are allocated more grid search areas, while those UAVs with smaller camera felds are allocated to a higher fight path.In this way, the coordinated operation of multiple UAVs has greater security and reliability.(2) Route planning for transmission line defect detection.Transmission lines often span a variety of diferent terrains, and the autonomous fight of UAVs is often afected by complex terrain.In order to display the high lines in real time in the map of the detection system, we embed a digital elevation model (DEM).We plan the aerial photography path of the UAVs between 20 m and 30 m above the contour line; the UAVs' lifting requirements have been reduced, and the efciency of image collection has been improved.(3) Marking of inspection areas.To avoid duplication and omissions in the inspection area of the transmission line as much as possible, we abstract the perspective of the UAVs' camera into a convex quadrilateral, and the search area of each UAV is visually marked.To solve the problem of discreteness between videos collected by multiple UAVs, we add each calculated polygon area to the polygons in the search area.Trough the abovementioned methods, the same inspection area will neither be repeatedly photographed nor be missed or ignored.
Journal of Electrical and Computer Engineering (4) Target positioning.With the embedded YOLOv7 detection model, device defects in transmission lines can be detected.However, how to locate the detected defects or the true location of the target is a key issue.Te corresponding relationship between the WGS84 coordinate system and the pixel coordinate system can be obtained through the calibrated UAV parameters and the homologous video stream.More information can be found in our publication [24].(5) Collaboration platform testing.To verify the efectiveness of the proposed collaboration platform, we build a simple coordination system by using two DJI Mavic 2 Pro UAVs, and the preliminary test results are shown in Figure 3.
We test the proposed method in Dajianshan, Kunming, and we frst select the lowest contour line of Dajianshan as the fight path.As shown in Figure 3, the orange UAV fies clockwise from the start to the end along the orange-marked aerial route.At the same time, the purple UAV fies counterclockwise from the start to the end along the aerial shooting route marked in purple.In this way, video data from the area around this contour line are collected and saved by UAVs.After both UAVs reach the fnish line, both UAVs fy to a second contour line at a higher level.Te orange UAV then fies counterclockwise from the fnish point to the start along the orange-marked aerial route.At the same time, the purple UAV fies clockwise from the endpoint to the starting point along the aerial photography  As shown in Figure 3, the area near the two contour lines in Kunming's Dajian mountain is completely inspected by two UAVs working together in real time.In fact, more than two UAVs can be embedded into our collaborative platform, the energy consumption requirements of a single UAV are reduced, and the search range and search efciency are signifcantly improved.To sum up, some useful information about the multi-UAV collaboration platform can be summarized as follows.
(1) Normally, the same inspection range in Figure 3 requires a set of independent UAV to patrol twice along two diferent contour lines.Alternatively, two sets of independent UAVs patrol along two diferent contour lines once.In addition, independent inspection methods face the risk of overlapping inspection scopes.In the method proposed in this paper, two UAVs are integrated into a control system, and the implementation of inspection tasks is completed simultaneously, and there is no overlap in the inspection area.Detection and location of transmission line defects can be completed on a computer with low confguration and energy consumption and hardware requirements of the UAVs are further reduced.(2) If a single UAV is adopted, the path planning method lacks fexibility, and some key points may be ignored or overlapping inspections may occur.
Trough the DEM-based path planning method, the efectiveness of multi-UAV path planning and the accuracy of target positioning have been improved to a certain extent.By capturing video streams from UAVs equipped with monocular cameras, the target positioning system can locate the target from the video stream to the world geographic system 1984.Terefore, through the multi-UAV collaborative system, the problems of omission and overlap in the inspection area, the accuracy of coordinate transformation, and defect localization can be solved to a certain extent.Journal of Electrical and Computer Engineering improving the accuracy of defect detection.In addition, many literatures have proven that YOLOv7 is a fast and accurate detection algorithm, so the experiments in this article achieved the expected detection performance compared to methods such as YOLOv5 under the same parameters and structure.
2.2.YOLOv7-Based Defect Detection Method.Te accuracy of defect detection is a prerequisite for defect location.To meet the accuracy and real-time requirements of defect detection, we choose the newly released YOLOv7, which includes input, backbone, neck, and prediction, and it is currently the fastest detector with the highest accuracy.Te several improvements of YOLOv7 can be summarized as follows.
(1) Extended efcient layer aggregation networks: In YOLOv7, the author proposed an extended-efcient long-range attention network (E-ELAN), which can converge more efectively by controlling the shortest longest gradient path.Te main architecture of E-ELAN is shown in Figure 4. Te E-ELAN enhances the learning ability of the network through expansion, shufe, and merge cardinality without damaging the original gradient path.Trough the group convolution strategy in E-ELAN, the channels of computational blocks are expanded, and the same channel multiplier and group parameter are used for all the computational blocks.(2) Model scaling: For the concatenation-based model, we cannot analyse the efects of diferent scale factors on the amount of parameters and computation separately.In YOLOv7, a corresponding compound model scaling method is proposed.When we scale the depth factor of a computational block, the change in the output channel of the block should also be calculated.Ten, the width factor scaling with the same amount of change on the transition layers can be performed.Te compound scaling up depth and width for the concatenation-based model is shown in Figure 5.When the model is scaled, only the depth the computational block needs to be scaled, and the remaining of the transmission layer performs the corresponding width scaling.(3) Planned reparameterized convolution: In view of the problem that RepConv reduces the detection accuracy, the authors design planned reparameterized convolution in YOLOv7.By analysing the combination of RepConv and diferent structures, we fnd that the RepConv provides more diversity of gradients for diferent feature maps.In YOLOv7, RepConvN is used to design the planned reparameterized convolution; an example is shown in Figure 6.(4) Coarse for auxiliary and fne for lead loss: Deep supervision improves the performance of the model.For the issue of label assignment, researchers use the quality and distribution of the prediction output and consider with the ground truth to generate a reliable soft label.YOLO uses the IOU between the prediction bounding box and the ground truth as the soft label of the object.As shown in Figure 7, the lead head guided label assigner is based on the prediction result of the lead head and the ground truth.Te lead head has a strong learning capability, so the strategy should be more representative of the distribution and correlation between the data and the target.
Coarse-to-fne lead head-guided label assigner generate the coarse label and fne label, where the fne label is the same as the former and the coarse label is generated by allowing more grids to be treated as the positive target.To avoid losing target information, we will focus on optimizing the recall of auxiliary heads in the object detection task.

Experimental Preparation.
To verify the real time and efectiveness of the proposed defect detection method based on YOLOv7, we evaluated the detection performance of various inspection methods through a series of experiments.
As shown in Table 1, we analyse the experimental results and verify the efectiveness of the method through the hardware and software confgurations in the following.In this paper, two methods are used to collect transmission line defect datasets, crawling insulator images on the network through web crawlers and taking insulator images of transmission lines in the Yuxi section of Yunnan power grid, and then the data enhancement method is used to optimize the datasets.Finally, a comprehensive dataset consisting of fve types of transmission line defect images, including normal glass insulators, normal ceramic insulators, insulator self-blast, cracked insulator, and bird nests, is obtained, which contained a total of 4835 images.Tere are signifcant diferences in the shooting angle, lighting, and image background of defect targets on these transmission lines, and the background and size transformation of defect targets are particularly complex in most images, as shown in Figure 8.It is thus clear that the complex backgrounds and size transformation bring great difculties and challenges to the research of transmission line defect detection.Terefore, we divide 4835 images into training sets and test sets according to the ratio of 8 : 2 and propose a multi-UAV collaboration platform based on YOLOv7 to improve the accuracy and efciency of transmission line defect detection in complex backgrounds.

Evaluation Methods.
We compare and analyse various defect detection algorithms through evaluation indicators such as precision (P), recall (R), F1 score (F1), mean average precision (mAP), model size and speed, and the detection accuracy, generalization performance, computing efciency, and storage size of the algorithms are well compared and analysed.Several key evaluation indicators are listed in the following: 6 Journal of Electrical and Computer Engineering Precision � TP TP + FP , where TP, FP, and FN represent true positive, false positive, and false negative, respectively.Te closer F1 is to 1, the better the detection accuracy and generalization performance of the model.Average precision (AP) is a defnite integral, which can be represented by the closed interval of the precision recall curve.In addition to the above 4 equations, the average recognition precisions of all defect categories in transmission lines can be represented by mAP.In addition, the storage size of the trained model can be expressed as "size," and the sum of preprocessing time, nonmaximum suppression time, and inference time can be expressed as "speed."2. Figure 9(a) clearly shows that the training precision curves of almost all detection models converge to a fxed value after 300 epochs.Specifcally, YOLOv7 converges to the highest training accuracy after 240 epochs.Te training precision curves of YOLOv5-X and YOLOv7 are not only smoother but also have higher training convergence precision and faster convergence speed.At the end of the training, YOLOv7 achieved the highest precision, indicating that YOLOv7 has the best convergence speed and convergence accuracy compared to other methods.As shown in Figure 9(b), the recall curve obtained by YOLOv7 not only has the fastest convergence speed, convergence occurs after 200 epochs, but also has the largest recall value and the highest smoothness.Te experimental comparison results show that the precision curve and the recall curve obtained from YOLOv7 in the training phase have the fastest convergence speed and the highest convergence accuracy, indicating that the trained YOLOv7 model has better performance than the other versions of YOLOv5 and YOLOv7.In addition, the smoothness of the precision and recall curves of the YOLOv7 model is the best, indicating that YOLOv7 has the best robustness to defect samples in transmission lines.

Test Performance Comparison.
We tested these trained models through 1101 real scene images from the proposed transmission line defect datasets, in which the threshold for IOU is 0.7.Te precision (P) and recall (R) of the fve transmission line defect detections, including normal glass insulator, normal ceramic insulator, cracked insulator, insulator self-blast, and bird nest, are shown in Table 3.Compared with normal insulators, there are fewer images and cases of bird nests, insulator self-blast, and cracked insulator.In the three fault cases, the detection of the insulator self-blast is the most difcult, and the detection difculty of the cracked insulator is only second to the insulator self-blast.YOLOv7-X obtained the best precision in bird nest, insulator self-blast, and cracked insulator defect detection, and YOLOv7 achieved the second highest detection accuracy among these types of defect samples.YOLOv7 received the best recall in four types of defect detection except insulator self-blast.In real transmission line defect detection experiments, YOLOv7 obtained the best precision (P) and recall (R) in almost all categories of defect object detection.Compared with the YOLOv5 and YOLOv7 series, YOLOv7 have better performance for insulator defect detection in complex backgrounds.
To compare these transmission line defect detection models more comprehensively and verify the high efciency and low complexity of the YOLOv7-based transmission line defect detection models, we provide the precision (P), recall (R), mAP, F1-scores, speed, and size obtained from these models in the test phase, as shown in Table 4.It should be noted that the test environment and hardware devices of various models are shown in Table 1, and the test set contains 1101 images in real scenarios, of which the threshold value of the IOU is 0.7, and the other common parameters and working conditions of the model are the same.Te reason why diferent models are tested with the same hardware environment, model parameters, and working conditions is to ensure the fairness and efectiveness of model performance comparison.
As shown in Table 4, YOLOv5-X achieved the highest precision 0.893 and the best F1 score 0.877, YOLOv7 achieved the highest recall 0.868 and highest mAP 0.886, and YOLOv7-Tiny has the fastest speed 1.2 ms and smallest size 11.7 Mb, which indicates that YOLOv7 and YOLOv5-X are more suitable for ofine detection or servers that require higher transmission line defect detection accuracy and YOLOv7-Tiny is more suitable for online monitoring or embedded devices that require higher detection speed and smaller model size.In addition to YOLOv7 and YOLOv5-X, YOLOv5-S and YOLOv7-X also achieved good results on various evaluation indicators.Compared with other detection models, the recall, mAP, and F1 indicators obtained from YOLOv7 all achieved the best values, while precision, speed, and size all ranked third and had a small diference from the top two.Tat is, YOLOv7 achieved the best performance in the most critical indicators and also ranked among the top in other indicators, and the accuracy and speed of defect detection meet the actual needs.
According to the actual situation of transmission line defect detection scenarios, YOLOv7 undoubtedly achieved the best performance in comprehensive evaluation and is, therefore, the best candidate to be added as a detector to the multi-UAV collaboration platform.
To demonstrate the performance of various models in real transmission line defect detection more intuitively, the detection results of each type of defects obtained by the fve defect detection methods are shown in Figure 10.For the cracked insulator, the fve models have successfully detected two cracked insulators and other normal ceramic insulators; YOLOv7 received the highest confdence, with YOLOv5-X coming in second.For the normal insulator, YOLOv7 has     5.
For insulator self-blast and cracked insulator faults with few samples and high identifcation difculty, YOLOv7 has made good breakthroughs in the four evaluation indicators, and several improvements of YOLOv7 are signifcant.For "birdNest," which has the smallest sample, the YOLOv7based method obtains the best mAP 0.915 and the best F1 0.912, respectively.In addition, for "selfBlast" with fewer samples, the YOLOv7-based method obtains the best recall 0.926 and other indicators are also very impressive, indicating that the YOLOv7-based detection model has good generalization performance for small samples and unbalanced samples.Overall, the detection model based on YOLOv7 achieved a precision of 0.886 and recall of 0.868 on  According to diferent environments in actual transmission line inspection and insulator defect detection, we divide the test dataset into four most common scenarios, including green land, transformer substation, high-voltage tower, and other complex environment, as shown in Figure 11."Greenland" means that the background of the picture is a green space."Transformer substation" means that the image data are collected from the substation or the picture contains the background of the substation."Tower" refers to the high-pressure tower in the background of the image."Complex environment" refers to images collected from extreme weather such as foggy or rainy days or poor image quality.Te size, illumination, shooting angle, and image background of transmission line defect images in diferent real scenes are quite diferent, and many defect detection methods are difcult to achieve satisfactory results.However, the YOLOv7 model we trained can successfully detect the fve most common transmission line defects in four diferent complex scenarios, and has achieved very good performance in each type of defect detection; the confdence has reached the practical application requirements.In "Greenland," the proposed method accurately detects bird nest, normal glass insulator, normal ceramic insulator, and insulator self-blast, with confdence levels ranging from 0.85 to 0.91.In "Transformer substation" and "Tower," the proposed method accurately detected all defect targets that appeared, with confdence levels ranging from 0.83 to 0.95.Even in the "Complex environment" with low image quality, missed detection, and false detection did not occur, with confdence levels ranging from 0.82 to 0.95.In addition, we also test and analyse the transmission line defect dataset constructed based on real scenarios through the multi-UAV collaborative platform and YOLOv7 model, and the detection time of a single image is 6.5 ms and the detection accuracy is very high, and the results show that the proposed defect detection method can ensure the detection accuracy and detection efciency of multiple video streams.

Conclusion
A transmission line defect detection method based on YOLOv7 and the multi-UAV collaboration platform is proposed.Multiple UAVs can fexibly access the collaborative platform through the same control system.Te practical results indicate that the multi-UAV collaboration platform can patrol a wider range within the same time and can effectively avoid path omissions and overlap issues, which improve the efciency and reliability of inspection.Second, several improvements of YOLOv7 have efectively improved the error detection problem caused by the aerial shooting angle, changing lighting, and complex background.Compared with the detection results obtained by YOLOv5-S, the YOLOv7-based defect detection method improves accuracy by 1.2%, recall by 4.3%, and mAP by 4.1%.Compared with other detection models, the recall 0.868, mAP 0.886, and F1 0.876 obtained from YOLOv7 all achieved the best values, while precision 0.886, speed 6.2 ms, and size 71.3Mb all ranked third and had a small diference from the top two.Te proposed defect detection method based on YOLOv7 has undoubtedly achieved the best results in experimental and comprehensive evaluation, meeting the accuracy, speed, and size requirements for various defect detections in transmission lines.Te defect detection dataset we proposed covers almost all the most common defect types such as normal insulators, insulator self-blast, cracked insulator, and bird nests, and more types of transmission line defect images can be added to this dataset.By expanding the dataset, our proposed method can simultaneously detect multiple defects in transmission lines, without being limited to single defects or certain defects, and the detection efciency and reliability are improved.In addition, the defect detection method proposed by us has a high level of automation and intelligence, and the labor intensity of transmission line inspectors is reduced and the safety is further improved.However, the fight path of multiple UAVs is set in advance according to the contour line, and the image data collected by the system in the undulating areas of the Yunnan Plateau have the problem of omission, overlap, or low quality.Terefore, in the near future, we will study how to make the collaborative system have the functions of autonomously planning paths and dynamically adjusting paths.At the same time, the YOLOv7 detection algorithm is difcult to detect low-quality small targets caused by altitude changes, extreme weather, and object occlusion, so we will study more accurate and faster object detection algorithms and embed the algorithm into multi-UAVs through embedded devices to improve the accuracy and speed of defect detection.

3. 3 .
Comparison of Parameters in Training.Te datasets obtained by the multi-UAV collaboration platform and data enhancement processing greatly expand the number of datasets, and then we train 5 state-of-the-art YOLO-based defect detection algorithms through the obtained datasets.In order to ensure the comparability and fairness of experimental results, all datasets, parameters, and experimental environments are the same, and the detection performance of YOLOv5-S, YOLOv5-X, YOLOv7-Tiny, YOLOv7-X, and YOLOv7 on defective targets is fully released.Te training parameters are shown in Table

Figure 9 :
Figure 9: Comparison of precision and recall during training.(a) Precision curves.(b) Recall curves.

Figure 10 :
Figure 10: Comparison of detection efects of diferent models.

Figure 11 :
Figure 11: Detection performance testing of proposed methods in diferent real scenarios.

Table 3 :
Te precision (P) and recall (R) for defects detection.

Table 4 :
Transmission line defect detection results obtained by diferent methods.Electrical and Computer Engineering the highest confdence and the next is YOLOv5-X.In the insulator self-blast and bird nest images, YOLOv7 received the highest confdence, and YOLOv5-S and YOLOv5-X have error detection.Te results show that YOLOv7 is more accurate than the existing YOLOv5 and some YOLOv7based models in defect detection under complex backgrounds, and several improvements to YOLOv7 are efective.From four diferent types of defect detection results, YOLOv7 achieved the best confdence, and there are no cases of missed detection or false detection.Te practical results and various evaluation indicators achieved by YOLOv7 demonstrate the high superiority of the transmission line defect detection method based on YOLOv7.Terefore, it is introduced into a multi-UAV collaborative system to detect various defects in transmission lines.3.5.Transmission Line Defect Detection Results.It can be seen from the comparative experiments of training and testing that YOLOv7 has the highest precision and recall in training, YOLOv7 achieved the best recall and mAP in testing, and YOLOv5-X has the highest precision and F1 score.In terms of time, YOLOv7-Tiny got the fastest speed in training and testing.Transmission line defect detection is critical to the safety of transmission lines and personnel, and it places particular demands on precision.Terefore, we use the YOLOv7 model to detect transmission line defects in various scenarios in the real world, and the testing performance is shown in Table

Table 5 :
Transmission line defect detection results.
all samples, which can meet the actual needs of transmission line defect detection.