Research on Insulator Defect Detection Based on Improved YOLOv7 and Multi-UAV Cooperative System

Abstract

Insulator self-blasts, cracked insulators, and bird nests often lead to large-scale power outages and safety accidents, while the detection system based on a single UAV and YOLOv7 is difficult to meet the speed and accuracy requirements in actual detection. Therefore, a novel insulator defect detection method based on improved YOLOv7 and a multi-UAV collaborative system is proposed innovatively. Firstly, a complete insulator defects dataset is constructed, and the introduction of insulator self-blasts, cracked insulators, and bird nest images avoids the problem of low reliability for single defect detection. Secondly, a multi-UAV collaborative platform is proposed, which improves the search scope and efficiency. Most critically, an improved YOLOv7-C3C2-GAM is proposed. The introduction of the C3C2 module and the CNeB2 structure improves the efficiency and accuracy of feature extraction, and the introduction of a global attention mechanism (GAM) improved the feature extraction ability to extract key information about small targets or occluded targets and feature in the region of interest. Compared with YOLOv7, the accuracies of YOLOv7-C3C2 and YOLOv7-C3C2-GAM are improved by 1.3% and 0.5%, respectively, the speed of YOLOv7-C3C2 is improved by 0.1 ms, and the lightweight sizes are reduced by 8.2 Mb and 8.1 Mb, respectively. Therefore, the proposed method provides theoretical and technical support for power equipment defect detection.

1. Introduction

The National Energy Administration, State Grid Corporation of China, and Yunnan Power Grid Corporation have all called for high attention to be paid to the safety management of the power transmission system. They emphasized the need to enhance the emergency response capabilities of power system security, improve the intelligence level of power transmission equipment inspection, and strengthen the treatment of safety hazards in power transmission facilities. The insulator is an important insulating device in high voltage transmission line [1], which has the function of support, insulation, and protection [2]. The self-explosion and shedding of glass insulators and the local damage of composite insulators reduce the insulation performance, which lead to electric shock accidents and large-scale blackouts, and the economic losses are immeasurable [3]. The bird nest will not only pollute tower poles and transmission lines but also reduce the safety of maintenance personnel and birds. In addition, the building materials of the bird nest may contain conductive materials such as iron wire and metal chips [4]. Dangerous accidents such as short circuits and flashover discharges occur from time to time during thunderstorms. Therefore, the self-explosion detection of glass insulators, the damage detection of composite insulators, and the detection of bird nests have important research value, which provides strong technical support for transmission line maintenance.

Currently, many electric power companies still rely on manual inspection to detect faults and safety hazards on high-voltage transmission lines. Manual inspections require a significant amount of manpower, financial resources, physical materials, and time, and excessive inspection costs can greatly reduce the production efficiency of enterprises. In high-altitude and mountainous areas, manual inspections under extreme weather conditions face great danger, and safety accidents will reduce the economic and social benefits of power companies. The results of manual inspections not only rely heavily on human experience but also significantly consume the physical energy of inspectors during long inspection tasks, reducing the reliability of the inspection process and results. With the development of unmanned aerial vehicle (UAV) technology, 5G communication technology, and target detection technology, the maintenance of high-voltage transmission lines has gradually shifted from manual inspection to robot and UAV inspections. During robot inspections, small climbing robots are placed on the tower and transmission lines. The robot takes photos of each key component along the tower and transmission line in turn and identifies the fault types of key components such as insulators through detection algorithms. However, robots are costly, and extreme weather conditions can easily lead to robot failures and economic losses. The high-power consumption, low efficiency, and long inspection cycles of robot inspections make it difficult to ensure the real-time inspection requirements of high-voltage transmission lines and may delay the optimal maintenance time. The small camera carried by the robot has a limited viewing angle, significantly reducing the accuracy and reliability of fault detection. UAV inspection is another new inspection method. Small UAVs equipped with high-definition cameras can obtain videos and image information of key components such as insulators on transmission lines. The real-time transmission of images enables target detection algorithms deployed on servers to detect defects of key components such as insulators in real-time. However, UAV inspections face several critical issues, such as: (1) Some UAVs do not have communication modules, and the offline fault detection of key components such as insulators will take more time since only video and image acquisition is implemented during UAV inspections. (2) Most inspection tasks are based on a single UAV. When the inspection scope is large, the efficiency and reliability of a single UAV are low. (3) The recognition accuracy, detection speed, and lightweight level of target detection algorithms need to be further improved. In addition, the release of the “Provisional Regulations on the Management of Unmanned Aerial Vehicle Flights” by the Ministry of Transportation and the drafting of the “Regulations on the Safety Management of Civil Unmanned Aerial Vehicles” by the Civil Aviation Administration of China have raised higher requirements for the flight control and safety management of low-altitude UAVs, which, to some extent, limits their application in dangerous scenarios such as transmission line inspection. Therefore, strict adherence to the regulations on UAV flight management is necessary during the process of transmission line defect detection using UAVs.

In order to improve the inspection efficiency of UAVs and realize the accurate detection of insulator self-explosions, local damage, and bird nests on the basis of improving the key feature extraction performance of YOLOv7, an insulator defect detection method based on improved YOLOv7 and a multi-UAV collaborative system is proposed.

The innovation and main contributions of the paper can be summarized as follows:

(1)

We construct a high-performance multi-UAV platform. Object detection and positioning can be completed on the computer, and the performance requirements and energy consumption of the UAV are reduced. In addition, multiple UAVs can be carried through the open design of the platform.

(2)

We propose an insulator defect detection model based on improved YOLOv7, including YOLOv7-C3C2 and YOLOv7-C3C2-GAM. Regarding the improvement strategy, the C3C2 module is introduced to replace the Catconv module of YOLOv7, the two sets of E-ELAN in the backbone network are replaced with lightweight CNeB2 modules, and the speed and accuracy of the feature extraction were improved. In addition, the global attention mechanism (GAM) with superior performance is introduced into the head of YOLOv7, and the key feature information of small targets and occlusion targets is well paid attention to, which improves the detection accuracy of small targets and occlusion targets. Through the combination of a multi-UAV collaborative system and improved YOLOv7, the error detection problems caused by lighting and complex backgrounds have been improved, and the speed and reliability have been improved significantly.

(3)

To improve the reliability of the model, we build an insulator defect dataset, which contains most common insulator defect images, insulator self-blasts, cracked insulators, and bird nest interferences.

If visible light and infrared images can be successfully collected, the proposed method can not only be used to detect defects in components, such as insulators, shock absorbers, nuts, bolts, bird nests, and foreign objects in transmission lines, but also to detect local defects in equipment such as insulators, transformers, high-voltage bushings, and high-voltage direct current devices in substation scenes. Furthermore, the proposed method can be applied not only to the defect detection of power equipment, such as high-voltage transmission lines, substation equipment, wind power generation equipment, and photovoltaic components, but also to various fields such as security, agriculture, transportation, education, natural disasters, medical imaging, construction, food safety, and biodiversity conservation.

The structure of this paper is as follows. The latest related works are introduced in Section 2. The multi-UAV collaboration platform and improved YOLOv7 detection model are described in Section 3. The insulator defect detection experiments are included in Section 4. Finally, the conclusion of the experiment is summarized in Section 5.

2. Related Works

In order to quickly and accurately detect damaged insulators and bird nests in high voltage transmission lines, some recent studies have been carried out at home and abroad. For defect recognition methods based on UAV and object detection algorithms, Jeffrey Kuo [5] used UAVs to capture both infrared and RGB images and enhance image features through CNN and color space transformations, and the accuracy of defect detection in photovoltaic modules was significantly improved. Bruno [6] used UAVs to collect 1593 images of power equipment defects and identified substation equipment failures through YOLOv5x based on ResNet-18. Qiu [7] used UAVs to capture images of flat pedestrian walkways and detected surface cracks in real-time through YOLOv2 and YOLOv4-Tiny, based on ResNet50. Zhang established a UAV road damage database and detected road damage and cracks through YOLOv3, based on a multi-level attention block [8]. Jae Jin [9] captured infrared images of critical equipment at nuclear power plants using UAVs and identified equipment failures through object detection methods, based on CNN. Wang [10] used UAVs to capture thermal images of dam surfaces and introduced an auxiliary input branch into the U-Net framework to reduce the misdiagnosis of dam cracks and damage caused by complex backgrounds.

For insulator defect detection, Deng replaced YOLOv4’s backbone CSPDarknet53 with MobilieNetv3, replaced the ReLU activation function of MobilieNetv3 with the PReLU activation function, and proposed a lightweight insulator defect detection network [11]. Han proposed an insulator defect detection method based on YOLOv4-Tiny by combining a Self-Attention mechanism and ECA-Net, and transplanted it to the Jetson Xavier NX to improve the speed [12]. Li converted the image from the RGB space to the HSV space, fused the dark details and overall brightness through Retinex, and, finally, detected the insulator defects through YOLOv5 [13]. Dai introduced the Gaussian function into the detection head of YOLOX, and improved the bounding box prediction by the estimated uncertainty score, and the insulator’s uncertainty estimation problem and the robustness of YOLOX were improved [14]. Zheng used k-means++ to cluster the insulator target frames and then added the Coordination Attention (CoordAtt) module and the HorBlock module in YOLOv7, and the insulator detection accuracy reached 93.8% [15]. Chen [16] proposed a method that combined attention feedback and a dual spatial pyramid to improve the detection accuracy of insulator defects under occluded backgrounds by enhancing the discriminative features with attention. Yuan [17] constructed a dataset of transmission line defect images containing bird nests, insulator defects, and seismic hammer defects and introduced an attention mechanism and a small object detection layer into YOLOv5, improving the defect detection accuracy. He [18] proposed a multi-layer information fusion and attention mechanism network based on YOLOv4, which fused the shallow features with detailed texture information into the feature pyramid, resulting in a 4.78% improvement in the detection accuracy of insulator explosions. Xing [19] used Gaussian filtering, k-means++ clustering, and mosaic data augmentation to process the data and proposed a lightweight model based on MobileNet-YOLOv4 to detect insulator defects. Liu [20] constructed a dataset of 6000 infrared images of electrical equipment and then used YOLOv4 to locate the areas of thermal faults in power equipment.

For bird nest and foreign body detection, Satheeswari used Visual Geometry Group (VGG16) and EfficientNet as the backbone of a Single Shot Multibox Detector (SSD) to extract features and locate bird nests [21]. Zhang generated the aspect ratio of regional proposals through the region proposal network, and the small target missed detection problem of Faster-RCNN was improved under the transmission line foreign object detection task [22]. Wu improved YOLOX by using atrous spatial pyramid pooling to increase sensitivity to foreign objects, embedding a convolutional block attention module to increase detection accuracy [23]. Qiu proposed a lightweight YOLOv4 with an embedded dual attention mechanism (YOLOv4-EDAM) to detect foreign objects on power lines and used MobileNetV2 to replace CSPDarkNet53 and the depthwise separable convolutions (DSC) to replace the standard convolutions in the SPP and the PANet module, having a mean average precision (mAP) of 96.71% [24]. Zhang [25] proposed a method that inserted a Swin Transformer into the backbone of YOLOv4 and fused attention mechanisms into the neck of the model, achieving a detection accuracy of 88% for bird nests on towers. Sui [26] proposed a bird nest detection algorithm based on dynamic federated learning, which could train a joint model without uploading data from edge nodes, significantly improving the detection accuracy.

However, the current studies on transmission line defect detection still have the following problems. (1) For the detection of objects, these studies either only detect the self-explosion of glass insulators or only detect bird nests or foreign objects. A single detection task cannot meet the multi-task requirements of actual transmission line defect detection. Even in a single detection task, the number of insulator images or bird nest images was small, which made it difficult to train a high-precision model. (2) In addition, insulators or bird nest images are collected by a single UAV in existing studies [27], equipping sensors and carrying detection algorithms, and the UAV system could complete the image acquisition and target recognition tasks [28]. The existing inspection strategy based on a single UAV is prone to miss the defects of key components such as insulators, and the risk of missed and false inspection is significantly increased [29]. (3) Faster-RCNN [30], SSD, YOLOv3, YOLOv4, YOLOv5, and YOLOX have been used for the detection of insulators and foreign objects such as bird nests. However, SSD and Faster-RCNN have slow inference speeds due to the generation of a large number of proposal boxes [31]. The detection performance of YOLO-based algorithms needs to be improved under the conditions of small targets, occlusions, and complex backgrounds. Recently, CenterNet [32], YOLOv7 [33], FCOS [34] without anchor frames, and transform-based methods [35] have been proposed to improve detection performance. Gong replaced the shallow layer with a feature fusion layer and replaced YOLOv5 head with Swin Transformer prediction heads (SPHs), which improved the detection accuracy of small targets in satellite images by 0.071% [36]. Zhu redesigned FCOS by introducing improved focal loss and regression refinement with complete intersection over union (CIoU) loss, improving the missed detection problem of small vessels under scattering interference [37]. Liu inserted the Global Attention Mechanism (GAM) into the Backbone and Head of YOLOv7, and the amplification of the driver’s global key information increased the detection accuracy by 20.26% [38]. Jiang introduced three CBAM modules into the backbone of YOLOv7, which improved the accuracy of Hemp Duck detection [39]. The only two latest improvement studies on YOLOv7 were based on the improvement of key features and attention mechanism, but the effectiveness of the attention needs to be further discussed, and the neck feature fusion needs to be studied.

3. The Proposed Materials and Methods

3.1. Multi-UAV Collaboration Platform

The main components of the multi-UAV system include the multi-UAV module, path planning module, and positioning and detection module. Firstly, we built a multi-UAV module based on an open interface platform. The DJI-branded UAVs were used as data acquisition terminals. A stable DJI Mobile SDK was used to control the UAV flight and coordinated the work of the camera and the pan–tilt. The status of the video and other components could be received in real time. The DJI Mobile SDK is shown in Figure 1, including gimbal, cameral, flight controller, airlink, remote controller, and battery. The detailed structure and characteristics of each submodule have been given in Figure 1.

Secondly, we built a multi-UAV path planning scheme, including two modes, one was the global planning based on the search scope and flight path, and the other was to receive the position of each UAV through map annotation. We implemented path planning through point-of-flight tasks. Firstly, the geographic coordinates that the UAVs needed to reach, in turn, were listed as arrays. Then, we could select specific UAVs and the coordinates of the flight points that the specific UAVs needed to reach; thus, the path of each UAV was clearly planned. The specific steps were as follows.

(1)

Division of search area for multi-UAV. If we take a single 4G or 5G signal base station as the reference center, the initial search area can be determined effectively. Centered on the base station of the last communication device, the area with the radius of the maximum moving distance of a person is the initial search area. Each UAV searches outward from the center gradually, and the flight path is prescribed a parallel grid or arc grid. Once the search range is determined, we divide the flight tasks based on the maximum flight distance and camera’s view angle for each UAV. Ultimately, UAVs with the farthest distances can arrange more grid search areas, and UAVs with smaller views can arrange higher missions to match low-altitude UAVs. In this way, the safety and search reliability of UAV are guaranteed.

(2)

High-voltage line inspection path planning. Complex terrain will affect the autonomous flight of the UAV in the high-voltage line inspection. To make the map display contour lines, we load the Digital Elevation Model (DEM) into the Map module. In addition, by a certain height above the contour line (20~30 m) to plan the line waypoint task, the lifting requirements of UAVs have been reduced, and the stability and efficiency of search have been significantly improved.

(3)

Marking of the inspection area. We abstract the camera’s view angle into a convex quad to mark the search area of each UAV, which avoids omissions and duplications of the search area. As the information of the UAV is discrete, each calculated polygon needs to be added to the polygons in the search area.

(4)

Target positioning. To locate the detected target in the video, such as cracked insulators or bird nests, we used a homologous video stream and some parameters from the UAV to obtain a correspondence between the pixel coordinate system and the two sets of points in WGS84. The detailed steps can be found in our paper [40].

(5)

Preliminary Test of multi-UAV Collaborative Platform. To verify the multi-UAVs’ cooperation performance, we build a test platform through two DJI Mavic 2 Pro UAVs, shown in Figure 2.

We tested our multi-UAV collaboration platform at the foot of Dajian Mountain in Kunming through two DJI UAVs. The first orange-marked UAV flew from the starting point to the end point along the contour line at the foot of the mountain in a clockwise direction. The second UAV marked in purple flew from the starting point to the ending point along the contour line at the foot of the mountain in a counterclockwise direction. The video streams collected by the two UAVs were transmitted to the detection module through RTMP and data transmission servers during the inspection process. As shown in “UAV1” and “UAV2” in Figure 2, the target in the video captured by two UAVs could be detected. After reaching the end point, the first orange-marked UAV returned to the starting point counterclockwise along another higher contour line, and the second purple marked UAV returned to the starting point clockwise along another higher contour line.

Through the simultaneous cooperation of two UAVs, the inspection task of the whole foot area of Kunming Dajian Mountain was completed effectively. In fact, more UAVs could be embedded dynamically in our multi-UAV platform, the performance requirements of multi-UAV would be reduced, and the inspection efficiency of a wide area would be improved. However, communication issues among multiple UAVs, path planning in complex terrains, and legal and ethical issues related to UAV flight cannot be ignored in the context of transmission line defect detection, and these issues will be seriously considered in the subsequent work. In addition, in order to further improve the target detection performance of the multi-UAV cooperative platform, the accuracy and speed of the object detection algorithm is a question worth exploring.

3.2. Improved YOLOv7 Detection Model

3.2.1. YOLOv7 Model

YOLOv7 is one of the most advanced single-stage object detectors, and its derivative algorithms also include PPYOLO, YOLOX, and YOLOR. Several versions of YOLOv7 had also been produced by the composite scaling method, including YOLOv7, YOLOv7-e6, YOLOv7-w6, and YOLOv7-x, and the requirements for different reasoning speeds and sizes have been met. The basic framework of YOLOv7 included Input, Backbone, and Head. The Input part enriched the data samples by stitching the data and had less computational overhead. The Backbone part was mainly composed of E-ELAN modules, which extracted image features through CBS-based convolutional layers. The Head aggregated the features of the feature map through SPPCSP and ELAN, and the channel of the output feature can be adjusted by RepConv to, finally, predict and output through 1 × 1 convolution.

3.2.2. Improved YOLOv7 Model

For the defects of YOLOv7, a YOLOv7-C3C2-GAM detection model, as shown in Figure 3, is proposed in this paper. Firstly, the C3C2 module is used to modify the network structure, with the aim of improving the recognition speed and lightweight level of the model. Secondly, the GAM attention mechanism module is introduced into the Head part to extract the key features and important information of the feature map. In addition, some network adjustments are used to further improve the performance of the model.

(1)

Network Structure. The improved YOLOv7 network architecture mainly includes MPConv, SPPCSPC, E-ELAN, BConv, CNeB2, C3C2, and GAM modules, as shown in Figure 3. BConv is composed of convolutional blocks with asynchronous lengths. E-ELAN and Catconv improve network detection accuracy by performing concat operations on other convolutional layers. The local value information and maximum value information of the local area can be extracted by MPConv at the same time, which is a down-sampling module. SPPCSPC is a new and improved spatial pyramid pooling structure (SPP) that combines spatial pyramid pooling and a CSP structure. In order to improve the detection accuracy of YOLOv7 for small targets and occluded targets, the GAM module is inserted into the Head part of the YOLOv7 network structure, and the Catconv module is replaced with the C3C2 module.

(2)

Some Improvements. To improve computational efficiency and recognition accuracy, 6 Convs and 1 Concat in YOLOv7’s backbone are replaced with CNeB, which stands for CSP ConvNextBlock with 3 convolutions. In YOLOv7-C3C2 and YOLOv7-C3C2-GAM, CNeB2 is added to the backbone network of YOLOv7, CNeB2 indicates that a convolutional layer is added to the CNeB, and the structure diagram of CNeB is shown in Figure 3. In the improved YOLOv7 network, the Catconv module is replaced with the C3C2 module, which represents CSP Bottleneck with 3 convolutions. The above network structure improvements have been added to the YOLOv7-C3C2 model. In order to improve the YOLOv7 model’s attention to key information and regions of interest, the GAM attention mechanism is introduced into the Head part of the YOLOv7-C3C2 model, and the resulting model is called YOLOv7-C3C2-GAM.

(3)

Global attention mechanism. During the detection of small or blocked objects, if the occluded area or area of interest is focused on, more key features that were conducive to recognition in the area will be extracted. GAM adopts the sequential channel–spatial attention in CBAM, but the internal submodules have been redesigned, as shown in Figure 4. Among them, the channel attention sub-module preserved feature information across three dimensions through 3D arrangement. Multi-layer perceptron is used to amplify the cross-dimensional channel–spatial correlation. The spatial attention sub-module achieves spatial information fusion through two convolutional layers. GAM improves the detection performance of the YOLOv7 network by reducing information loss and amplifying global interaction features, effectively balancing speed and accuracy.

4. Insulator Defect Detection Experiments

4.1. Data Preparation and Experiment Setup

To ensure the richness and practical availability of the data, we collected insulator defect datasets through web crawlers and the proposed multi-UAV system. In the actual data collection stage, we collected picture data of transmission line defects by deploying the multi-UAV collaboration platform of two DJI UAVs. Firstly, we selected the area with frequent faults of a high-voltage transmission line in the Yuxi section of the Yunnan Power Grid Company as the data collection area and the staff station in the data collection area to plan the data collection path of two UAVs, according to the actual situation of the transmission line. Then, we set the initialization parameters and path parameters of the multi-UAV collaboration platform, and the two UAVs photographed the defects of the transmission line according to the planned path, among which the data of the normal insulators and the glass insulators’ self-explosions were relatively more. In the process of data collection, the UAV pitch angle, flight altitude, photo mode, and distance from the transmission line and power equipment had been carefully thought and set. The collected data were stored in the storage unit of the UAV, and the data were transmitted to the remote server through the wireless communication network. The obtained data were further filtered, and some low quality and irrelevant images were deleted to obtain aerial images of various defects. Finally, these aerial images were classified and combined with the images obtained by the web crawler to obtain the final experimental data set. There was a total of 4835 images, which were composed of 5 categories, including bird nest disturbances, cracked insulators, normal ceramic insulators, normal glass insulators, and insulator self-blasts. As shown in Figure 5, the images of glass insulators, composite insulators, and bird nests in the dataset were taken under different lighting conditions and backgrounds. Thus, the size and shape of the objects such as the insulators and nests were different, and the lighting and background were extremely complex. The 4835 images of insulators and bird nests were divided into a training set of size 3734 and a test set of size 1101, with a ratio of nearly 8:2.

In order to verify the superiority of the improved YOLOv7, YOLOv7 was compared with various improved versions. To ensure the fairness of algorithm comparison, a unified hardware configuration and operating environment were used for comparison experiments of all versions. The hardware and environment configuration parameters shown in

Table 1. Experimental environment configuration.

Parameter	Configuration
CPU	Intel Core i7-10700F, CPU 2.90 GHz, RAM 32 GB
GPU	Nvidia GeForce GTX 2080Ti (24G)
Accelerated Environment	CUDA 11.1, cuDNN8.0.5
Visual Studio System	Pytorch1.7.1, Python 3.7
Operating System	Ubuntu 18.04

were used in the training and testing process of the detection models.

To compare and analyze the advantages and disadvantages of different detection models from different perspectives, Precision (P), Recall (R), F1 Score (F1), mean Average Precision (mAP), model size, and speed were used to evaluate the detection performances of different versions of YOLOv7. Precision (P), Recall (R), and F1 are shown in the following Equations (1)–(3).

$$\mathrm{Precision}=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FP}}$$

(1)

$$\mathrm{Recall}=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}$$

(2)

$${\mathrm{F}}_1=2\times \frac{\mathrm{Precision}\times \mathrm{Recall}}{\mathrm{Precision}+\mathrm{Recall}}$$

(3)

where False Positive (FP) indicates the sample is negative, and the predicted result is positive. True Positive (TP) represents the sample is positive, and the predicted result is positive. False Negative (FN) indicates the sample is negative, and the predicted result is negative. True Negative (TN) indicates the sample is positive, and the predicted result is negative. F1 is the harmonic mean of precision and recall. Generally, the larger the P and R are the closer F1 is to 1 and the better the detection performance of the model.

Average Precision (AP) indicates the area under the precision recall curve. The area of the enclosed area, i.e., AP, can be calculated by the definite integral of the curve, as shown in Equation (4).

$$\mathrm{AP}={\displaystyle {\int }_0^1\mathrm{P}\left(\mathrm{R}\right)\mathrm{dR}}$$

(4)

mAP refers to the average recognition accuracy of all target categories, model size represents the storage space of the model after training, and speed represents the sum of pre-processing time, inference time, and non-maximum suppression time.

4.2. Performance Comparison with the Existing Methods

In order to verify the superiority of the improved YOLOv7-based insulator defect detection method, we trained seven state-of-the-art object detection algorithms with the same dataset and configuration parameters, including YOLOv5-S, YOLOv5-X, YOLOv7-Tiny, YOLOv7-X, YOLOv7, YOLOv7-C3C2, and YOLOv7-C3C2-GAM. During the training phase, the parameters of these models are uniform and are shown in

Table 2. Training parameter settings.

Batch Size	Learning Rate	Weight Decay	Momentum	Input Image	Epochs	Gamma
8	1 × 10−5	0.001	0.98	640*640	500	1.5

.

In 500 epochs, all models tended to converge at the 300th epochs, as shown in Figure 6. YOLOv7 had good convergence performance and training accuracy, which had been demonstrated in many studies. Compared with the original YOLOv7, the introduction of the C3C2 module reduced the convergence speed of YOLOv7-C3C2, but the training accuracy during convergence was higher than that of the original YOLOv7. Furthermore, the introduction of C3C2 and GAM attention mechanisms made the YOLOv7-C3C2-GAM model converge at the 250th epoch, and the training accuracy of the model at convergence was the highest. It is worth noting that the training accuracy of YOLOv7-C3C2 between the 300th and 400th epoch was higher than that of YOLOv7-C3C2-GAM, and the above results showed that the introduction of C3C2 and GAM attention mechanism could improve the performance of model training.

In the detection and prediction phase, 1101 insulators and bird nest images collected by the multi-UAV cooperative system along the high-voltage transmission line were used to test the effectiveness of each model after training, and the threshold for IOU was 0.7. For various types of insulator defects and bird nest interferences, the precision (P) and recall (R) of seven types of detection models, such as YOLOv7-C3C2-GAM, are listed in

Table 3. The precision (P) and recall (R) of the seven insulator defect detection models.

Class		Bird Nest	Cracked	NormalCeramic	NormalGlass	SelfBlast
Labels		131	467	4841	1803	271
P	YOLOv5-S	0.901	0.864	0.867	0.888	0.851
	YOLOv5-X	0.901	0.876	0.907	0.918	0.864
	YOLOv7-Tiny	0.876	0.818	0.830	0.790	0.864
	YOLOv7-X	0.902	0.891	0.881	0.895	0.872
	YOLOv7	0.903	0.880	0.874	0.908	0.865
	YOLOv7-C3C2	0.908	0.889	0.910	0.919	0.878
	YOLOv7-C3C2-GAM	0.903	0.907	0.887	0.910	0.884
R	YOLOv5-S	0.907	0.833	0.764	0.734	0.889
	YOLOv5-X	0.905	0.859	0.83	0.784	0.934
	YOLOv7-Tiny	0.923	0.794	0.673	0.702	0.889
	YOLOv7-X	0.893	0.844	0.783	0.743	0.952
	YOLOv7	0.924	0.863	0.842	0.788	0.926
	YOLOv7-C3C2	0.903	0.816	0.696	0.707	0.899
	YOLOv7-C3C2-GAM	0.924	0.844	0.703	0.707	0.890

, which represent the recognition performance of the model for different types of samples, including bird nest disturbances, cracked insulators, normal ceramic insulators, normal glass insulators, and insulator self-blasts.

Compared with bird nests and normal insulators, the detection accuracies of the cracked insulators and insulator self-blasts are lower. One reason is that the dataset capacity of cracked insulator and insulator self-blast is smaller, and the other reason is that insulator local damage and self-explosion occupy fewer pixels in the image, which means they are small targets. Compared with mainstream methods, YOLOv7 has superior performance. Compared to YOLOv7, YOLOv7-C3C2 achieves a higher accuracy in the detection of each defect type, especially for targets with large capacity or large size, such as “normalCeramic” and “normalGlass”. Compared with YOLOv7-C3C2, YOLOv7-C3C2-GAM achieves higher detection accuracy on small capacity targets and small size targets, such as “cracked” and “selfBlast”. The results show that the introduction of the C3C2 module improves the detection performance of YOLOv7 for targets with large sample sizes or large sizes, while the introduction of the GAM attention module makes the YOLOv7 model pay more attention to key information and small target features. Unfortunately, the introduction of the C3C2 and GAM attention modules has not made significant progress in recall, as the introduction of these two modules focuses more on small targets and key feature information, and the missed detection rates for large sample volumes have improved. However, the recall rate of the improved YOLOv7 remains at a high level, which can meet the requirements of the insulator defect detection.

To verify the effectiveness of the improved YOLOv7 model, we also comprehensively compared several models in terms of average accuracy, inference speed, and model size, as shown in

Table 4. Experimental results of different networks.

Model	P	R	mAP	Speed (ms)	Size (Mb)
YOLOv5-S	0.874	0.825	0.845	1.9	13.7
YOLOv5-X	0.893	0.862	0.878	10.3	166
YOLOv7-Tiny	0.836	0.796	0.821	1.2	11.7
YOLOv7-X	0.889	0.843	0.883	6.3	135
YOLOv7	0.886	0.868	0.886	6.2	71.3
YOLOv7-C3C2	0.899	0.853	0.888	6.1	63.1
YOLOv7-C3C2-GAM	0.891	0.847	0.879	6.3	63.2

. YOLOv7-C3C2 obtained the highest accuracy of 0.899 and the highest mAP of 0.868, YOLOv7 obtained the highest recall of 0.868, and the recall of YOLOv7-C3C2 and YOLOv7-C3C2-GAM were close to the highest recall. From the perspective of lightweight and model deployment, in addition to YOLOv5-S and YOLOv7-Tiny, YOLOv7-C3C2 and YOLOv7-C3C2-GAM had the fastest speeds of 6.1 ms and 6.3 ms. In addition, the size of the YOLOv7-C3C2 and YOLOv7-C3C2-GAM models was 63.1 Mb, which was smaller than the size of YOLOv7. Therefore, the improved YOLOv7 algorithm not only had high accuracy and speed but also had a small size that was convenient for lightweight deployment. In the actual scenario data we presented, the detection performance of YOLOv7-C3C2 was slightly better than that of YOLOv7-C3C2-GAM.

In order to further analyze the detection performance of the proposed method in insulator defect detection, Figure 7 provides the detection details of seven detectors. For the cracked insulator, all seven detectors successfully detected two cracked insulators and other normal insulators, and YOLOv7-C3C2 obtained the highest confidences of 0.81 and 0.83. YOLOv7-C3C2-GAM also obtained the next highest confidence of 0.82. For normal insulators, YOLOv7-C3C2 had the highest confidence of 0.94. Although the confidence of 0.93 obtained by YOLOv7-C3C2-GAM was less than that of YOLOv5X, the speed and size of the model were far superior to those of YOLOv5X. For bird nest detection, only YOLOv7-C3C2, YOLOv7-C3C2-GAM, and YOLOv5X had detected the composite insulator hidden behind the tower, which was difficult to detect due to the shielding. The results showed that the introduction of the C3C2 and GAM attention mechanisms could improve the detection performance of the detectors for occluded objects and hidden objects, and the key features of the occluded objects could be fully extracted.

4.3. Insulator Defect Detection Results Based on Multi-UAV Aerial Images

According to the above experimental results, compared with YOLOv7 and other detectors, YOLOv7-C3C2 and YOLOv7-C3C2-GAM had higher detection accuracy, and they were the fastest and easiest to deploy models except YOLOv5-S and YOLOv7-Tiny. Therefore, we used YOLOv7-C3C2, YOLOv7-C3C2-GAM, and a multi-UAV collaborative system to detect insulator defects in different scenarios. In order to analyze the convergence characteristics of the three object detection algorithms,

Table 5. Model loss function analysis during the training and validation phases.

Model	Train Box Loss	Train Obj Loss	Train Cls Loss	Val Box Loss	Val Obj_Loss	Val Cls_Loss
YOLOv7	0.0444	0.0924	0.0078	0.0502	0.0822	0.0120
YOLOv7-C3C2	0.0400	0.0835	0.0011	0.0470	0.0854	0.0052
YOLOv7-C3C2-GAM	0.0407	0.0896	0.0012	0.0468	0.0848	0.0047

presents the location loss (box loss), objectness loss (Obj loss), and classes loss (Cls loss) of the models during the training and validation phases. During the training phase, the introduction of the C3C2 module and the GAM attention mechanism effectively reduced the box loss, Obj loss, and Cls loss. During the validation phase, compared to YOLOv7 with only the C3C2 module, the simultaneous introduction of the C3C2 module and GAM attention mechanism further reduced the box loss, Obj loss, and Cls loss. The experimental results showed that the introduction of both the C3C2 module and GAM attention mechanism could reduce the box loss, Obj loss, and Cls loss, which increased the confidence level of the transmission line defect detection model, improved the accuracy of classification, and enhanced the reliability of object localization.

To further analyze the performance of the model in detecting transmission line defects in practical situations, the detection results are shown in

Table 6. Experimental results of different networks.

Class	YOLOv7		YOLOv7-C3C2		YOLOv7-C3C2 -GAM
	P	R	P	R	P	R
BirdNest	0.903	0.923	0.908	0.903	0.903	0.924
Cracked	0.880	0.863	0.889	0.814	0.907	0.807
NormalCeramic	0.874	0.842	0.9	0.696	0.895	0.676
NormalGlass	0.908	0.788	0.919	0.707	0.895	0.677
SelfBlast	0.865	0.926	0.878	0.879	0.891	0.863
All	0.886	0.868	0.899	0.832	0.891	0.797

.

Insulator self-blasts and cracked insulators had the fewest sample labels, and they were small targets; thus, these two types of defects were the most difficult to detect. For the bird nests, normal ceramic insulators, and normal glass insulators, YOLOv7-C3C2 achieved the highest test accuracies of 0.908, 0.9, and 0.919, respectively. For the cracked insulators and insulator self-blasts, YOLOv7-C3C2-GAM achieved the highest test accuracies of 0.907 and 0.891, respectively. For all defect categories, YOLOv7-C3C2 had the highest test accuracy of 0.899, and YOLOv7-C3C2-GAM achieved the test accuracy of 0.891, both of which were higher than the test accuracies of YOLOv7. The results showed that the proposed improved YOLOv7 model effectively improved the detection accuracy and had a fast inference speed. At the same time, the recall of the improved method was close to the original method.

In order to visually analyze the detection performance of the model under different scenarios in actual transmission lines, we presented two sets of random test results for four scenarios, as shown in Figure 8, including green land, tower, and foggy weather. As shown in the above images of different scenes, the sizes of defect targets of different images were different, the backgrounds were complex, the lights were changing at any time, and the shooting angled of the aerial images were also very different, which brought great difficulties to the insulator defect detection. However, YOLOv7 and its improved algorithms could detect insulator defects in these four scenarios, but there were some differences between them in terms of confidence and missed detection rate, which were caused by the improvement of the algorithms. In the first group of four scenarios, the confidence levels of YOLOv7-C3C2 and YOLOv7-C3C2-GAM were higher than that of YOLOv7, and the confidence level of YOLOv7-C3C2-GAM was higher than that of YOLOv7-C3C2. The results showed that the introduction of the C3C2 module improved the feature extraction ability of the model, and the introduction of the GAM attention mechanism improved the model’s ability to process key information. In the second group of four scenarios, the confidence levels of YOLOv7-C3C2 and YOLOv7-C3C2-GAM were still greater than that of YOLOv7. However, in “Green land”, “Tower”, and “Foggy weather”, the confidence level of YOLOv7-C3C2 was higher than that of YOLOv7-C3C2-GAM, which meant that the introduction of the GAM attention mechanism seemed useless. In fact, sometimes, the attention mechanism focused on a small range of key information and led to the loss of features. However, this did not affect the superiority of the proposed improved YOLOv7 algorithm. In addition, we also tested the insulator defect detection method based on YOLOv7 and the multi-UAV collaborative system through the constructed dataset, and the test speed of each image was only 6.5 ms, which showed that our proposed method could save and detect multiple different video streams in real time. As UAV data collection was extremely difficult and dangerous in rainy and snowy scenarios, we did not test the effectiveness of the proposed insulator defect detection method in rainy and snowy scenarios. However, in the near future, we will collect data on power equipment in extreme weather such as rain and snow, and defect detection under low illumination and low visibility conditions will be a research direction for us in the future.

5. Conclusions

In order to detect insulator self-blasts, cracked insulators, and bird nests quickly and accurately, an insulator defect detection method based on improved YOLOv7 and a multi-UAV cooperative system was proposed. The experimental results showed that the proposed three innovative points could effectively improve the accuracy and reliability of transmission line defect detection. Firstly, the flexibility and detection speed of the proposed method were better than a single UAV system. Multiple UAVs could join the platform dynamically, which improved the detection range, reliability, and flexibility. Secondly, the proposed dataset based on bird nest disturbances, cracked insulators, and insulator self-blast images covered the most common defect types in insulator inspection, which enabled the model to detect multiple types of insulator defects simultaneously, rather than just a single defect. Most innovatively, we proposed a new improved YOLOv7-C3C2-GAM model by introducing C3C2 module and GAM attention mechanism. Six convolutional layers and a concat layer in YOLOv7’s Backbone were replaced by a CNeB layer, and Catconv in the Head was replaced by C3C2. The experimental results of the insulator defect detection in multiple scenes showed that the improvement of C3C2 and CNeB modules improved the feature extraction ability and lightweight level, and the number of network layers, parameters, and volume of the model were reduced. In addition, the insulator defect detection results showed that the introduction of GAM attention mechanism improved the detection accuracy of the model for small target insulators and occluded insulators. The proposed method provided an automatic detection framework for insulator defect detection in a wide range, which could effectively reduce the labor intensity of transmission line inspectors and meet the intelligent requirements of high-voltage line inspection. The fly in the ointment was that there was still some room for improvement in the real-time communication between multiple UAVs in the multi-UAV cooperative system, there was still redundancy and overlap in the path planning of multiple UAVs, and the proposed object detection algorithm still had room to be improved in the transmission line defect detection performance under the background of extreme weather and small targets. With the improvement of camera performance, the proposed method was expected to achieve better defect detection performance in the background of extreme weather and small targets. In the near future, this method will be used in mobile embedded devices, and the fusion of infrared and visible images will make the framework have a wider range of applications, including security, agriculture, transportation, education, natural disasters, medical imaging, construction, food safety, and biodiversity conservation.