Abnormal target detection method for power transmission lines based on UAV-mounted edge computing platform

Abstract

The present invention discloses a method for detecting abnormal targets in power transmission lines using an edge computing platform carried by a drone, and belongs to the field of drone target detection. The method mainly includes four steps: sample preparation, model establishment, target identification, and feedback results. First, various abnormal target image data are collected to prepare samples and divide them; second, a lightweight model is established by improving the YOLOv8n model, and the model is trained, tested, and verified using samples; then, the model is deployed to the edge computing platform carried by the drone to identify abnormal targets in real time, such as abnormal conditions such as broken insulators and bird nests in power transmission lines; finally, the identification results are fed back to safety managers in real time. This method not only improves the detection efficiency, but also improves the detection accuracy, and better solves the problem of safety management of power transmission lines.

Description

Technical Field

[0001] The present invention belongs to the field of unmanned aerial vehicle target detection, and more specifically to a method for detecting abnormal targets in power transmission lines using an unmanned aerial vehicle equipped with an edge computing platform. Background Art

[0002] Drones are widely used in many fields, such as agriculture, construction, geological exploration, etc. Among them, in the power industry, drones are becoming more and more common for inspection of transmission lines. Compared with traditional manual inspection methods, drone inspections are not affected by terrain, can complete line inspection tasks more quickly and comprehensively, and can obtain clear real-time images, greatly improving the efficiency and accuracy of line inspections. However, due to the complex environment of transmission lines and the presence of many interference factors, such as weather and electromagnetic interference, how to quickly and accurately detect abnormal targets in transmission lines has become a challenge.

[0003] With the development of artificial intelligence technology, many studies have begun to try to apply computer vision technology to drone inspections to automatically identify and locate line defects. These methods can improve the automation level of drone inspections to a certain extent, but because image processing and deep learning algorithms require a lot of computing resources, the real-time performance of the algorithm and inspection efficiency are affected.

[0004] In recent years, the introduction of edge computing technology has provided a new idea for solving this problem. Edge computing can realize local processing of data, avoid network delays caused by long-distance data transmission, and improve the real-time performance of the algorithm. Drones equipped with edge computing platforms can realize real-time processing of collected data, greatly improving data processing efficiency.

[0005] However, most of the existing transmission line inspection methods using drones equipped with edge computing platforms use traditional image processing and machine learning algorithms, which are not efficient in target detection and recognition in complex situations. Therefore, a new abnormal target detection method for transmission lines using drones equipped with edge computing platforms is needed to improve inspection efficiency and accuracy. Summary of the invention

[0006] The present invention mainly solves the technical problems that the existing transmission line inspection method using an unmanned aerial vehicle equipped with an edge computing platform cannot achieve fast and accurate identification of target detection in complex environments, and the inspection efficiency and accuracy are insufficiently improved. The present invention aims to provide a transmission line abnormal target detection method using an unmanned aerial vehicle equipped with an edge computing platform, so as to achieve fast and accurate abnormal target detection and improve inspection efficiency and accuracy.

[0007] In order to achieve the above object, the present invention is implemented by adopting the following technical scheme: the method comprises:

[0008] Sample preparation: collect various abnormal target image data, prepare various abnormal target samples in high-voltage transmission lines and their forest channels, and divide them into training samples, test samples and verification samples;

[0009] Model establishment: By improving the YOLOv8n model, a lightweight model architecture is formed and the model is trained, tested and verified through the established samples;

[0010] Target recognition: deploy the model to the edge computing platform carried by the drone to identify broken insulators on power transmission lines and abnormal bird nests in real time;

[0011] Feedback results: Feedback abnormal target identification results to security managers to provide support for security prevention and control.

[0012] In one embodiment, the sample preparation includes:

[0013] Collect image data: Image data comes from images captured by drones through their high-definition cameras during the inspection of transmission lines;

[0014] Sample making: pre-processing the collected image data, including cleaning and labeling steps; cleaning is to remove irrelevant information in the image; labeling is to mark the position of abnormal targets in the image;

[0015] Sample division: Divide the prepared samples into training samples, test samples, and validation samples in a ratio of 70%, 15%, and 15%; training samples are used to train the model, test samples are used to test the performance of the model, and validation samples are used to verify the model during training to prevent overfitting.

[0016] In one embodiment, the model is established using the YOLO-PowerLite model.

[0017] In one embodiment, the YOLO-PowerLite model is improved based on the YOLOv8n model;

[0018] (1) In the feature fusion part, the original Bottleneck structure in the C2f module is replaced with AK_Bottleneck, thus creating a new C2f_AK module;

[0019] (2) Optimize the decoupling head structure used in the original YOLOv8n model;

[0020] (3) The CA attention mechanism is introduced into the backbone network of the YOLO-PowerLite model; by emphasizing the channel and spatial information of the image, the model's attention to key features is improved and its sensitivity to background interference is reduced.

[0021] In one embodiment, the C2f_AK module structure includes:

[0022] Two convolutional layers and several AK_Botleneck modules, where the AK_Botleneck module is a residual block;

[0023] The C2f_AK module extracts and fuses the information in the input feature map through convolution operations and iterations of the AK_Botleneck module.

[0024] In one scheme, the decoupled head structure is optimized to reduce the total number of parameters and computational burden of the model by sharing parameters in the front part of the category prediction and bounding box regression branches; at the same time, by retaining the task-specific terminal layers, it ensures that the model learns effective feature representations for each task; the computational efficiency of the model is improved, and sufficient task specificity is maintained to achieve high performance.

[0025] In one scheme, the CA attention mechanism introduced in the backbone network includes two stages:

[0026] Coordinate information embedding and coordinate attention generation;

[0027] In the coordinate information embedding stage, for a given input X, the CA attention mechanism uses two scales of pooling kernels (H, 1) and (1, W) to encode each channel in the horizontal and vertical directions respectively to enhance the capture of spatial information;

[0028] In the coordinate attention generation stage, the feature maps obtained from the coordinate information embedding stage are first merged, and then the dimension is reduced through convolution transformation; after batch normalization and nonlinear activation processing, an intermediate feature map is generated.

[0029] In one embodiment, the target identification includes:

[0030] Model deployment: Deploy the trained and tested model to the edge computing platform carried by the drone;

[0031] Image acquisition: During the flight, the drone uses the high-definition camera on board to obtain real-time images of the power transmission line; these images will be transmitted to the edge computing platform in real time;

[0032] Real-time recognition: The edge computing platform inputs the received real-time images into the model, and the model identifies abnormal targets in the images through calculations;

[0033] In one embodiment, the feedback includes:

[0034] Abnormal result generation: When the edge computing platform on the drone completes the recognition and analysis of the image, it generates abnormal target recognition results;

[0035] Result transmission: The results of abnormal target identification are transmitted to the ground control center in real time via wireless network;

[0036] Result integration: For long-term or large-scale inspections, many abnormality identification results will be generated; these results are sorted out to analyze the total number of abnormalities and the distribution of various abnormalities.

[0037] Beneficial effects of the present invention:

[0038] The power transmission line abnormal target detection method provided by the present invention, which uses the edge computing platform equipped with a drone, realizes real-time data processing and analysis by using edge computing technology, and adopts advanced target detection models to improve the speed and accuracy of detection. At the same time, the method can also timely feedback alarms and generate abnormal reports, so that operators can be informed of abnormal conditions of the line in a timely manner, thereby greatly improving the efficiency of power line detection and ensuring the stable supply of electricity, and has high practical value and economic value. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 is a flow chart of the method of the present invention;

[0040] Figure 2 This is a network structure diagram of the model of the present invention;

[0041] Figure 3 It is a structural diagram of C2f_AK and AK_Bottleneck of the present invention;

[0042] Figure 4 It is a structural diagram of AKConv of the present invention;

[0043] Figure 5 It is a characteristic network design diagram of the present invention;

[0044] Figure 6 This is the design diagram of the head of the present invention;

[0045] Figure 7 This is a structural diagram of the CA attention module of the present invention. DETAILED DESCRIPTION

[0046] In order to facilitate the understanding of the present invention, the present invention will be described more fully below with reference to the relevant drawings. Typical embodiments of the present invention are given in the drawings. However, the present invention can be implemented in many different forms and is not limited to the embodiments described in the present invention. On the contrary, the purpose of providing these embodiments is to make the disclosure of the present invention more thorough and comprehensive.

[0047] Unless otherwise defined, all technical and scientific terms used in the present invention have the same meaning as those commonly understood by those skilled in the art of the present invention. The terms used in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention.

[0048] like Figure 1 As shown, the method comprises:

[0049] S1. Sample preparation: collect various abnormal target image data, prepare various abnormal target samples in high-voltage transmission lines and their forest channels, and divide them into training samples, test samples and verification samples;

[0050] (1) Collecting image data: The collected abnormal target image data mainly comes from the images captured by the drone’s high-definition camera during the inspection of the transmission line. These images contain various abnormal targets, such as broken insulators, bird nests, and overgrown weeds.

[0051] (2) Producing high-voltage transmission line samples: Preprocessing the collected image data, including cleaning, labeling and other steps. Cleaning mainly removes irrelevant information in the image, such as background, sky, etc. Labeling is to mark the location of abnormal targets in the image to provide "answers" for model training.

[0052] (3) Producing abnormal samples in forest corridors: In addition to the high-voltage transmission lines themselves, forest corridors are also a focus of attention. Because lush trees may have an impact on transmission lines, it is also necessary to collect abnormal target samples in this part, such as trees that are too tall, fallen trees, etc.

[0053] (4) Sample division: Divide the prepared samples into training samples, test samples, and validation samples according to a certain ratio. Generally speaking, the ratio can be 70%, 15%, and 15%. Training samples are used to train the model, test samples are used to test the performance of the model, and validation samples are used to verify the model during training to prevent overfitting.

[0054] S2. Model establishment: By improving the YOLOv8n model, a lightweight model architecture is formed and the model is trained, tested and verified through the established samples;

[0055] In order to ensure the accuracy of abnormal target detection on power transmission lines and to achieve lightweight model to reduce deployment cost, a series of targeted improvements were made based on the YOLOv8n model. The final improved network model is called YOLO-PowerLite. The improved model network structure is as follows: Figure 2 shown.

[0056] First, in the feature fusion part, the C2f module in YOLOv8n was significantly improved. By replacing the standard Bottleneck structure in the C2f module with AK_Bottleneck based on Alterable Kernel Convolution (AKConv), a new C2f_AK module was created. This improvement significantly reduced the number of parameters and computational complexity, which not only provided the model with the advantages of lightweight and efficiency, but also facilitated its deployment on drone edge computing devices for real-time detection of abnormal targets in power transmission lines. In addition, based on the bidirectional feature pyramid network (BiFPN), the feature fusion part was reconstructed to improve the efficiency and effectiveness of feature processing of abnormal targets in power transmission lines.

[0057] Secondly, in the design of the detection head, the decoupled head structure used in the original YOLOv8n model is optimized. Although the decoupled head structure improves model performance by providing dedicated network paths for category prediction and bounding box regression, the increased number of parameters and computational burden it brings may become a bottleneck in resource-constrained environments. Therefore, it is proposed to further reduce the total number of parameters and computational requirements of the model by sharing parameters in the front end of the category prediction and bounding box regression tasks, while maintaining efficient feature representation by retaining task-specific terminal layers.

[0058] Finally, considering that the transmission line images taken by drones often have complex backgrounds, such as trees, grass, and other power facilities, these complex backgrounds will have a great impact on the accurate detection of abnormal targets in the transmission lines, so the Coordinate Attention (CA) mechanism is introduced into the backbone network of the model. The CA attention mechanism effectively improves the model's attention to key features and reduces its sensitivity to background interference by emphasizing the channel and spatial information of the image.

[0059] In YOLOv8n, the C2f module is responsible for fusing high-level semantic features with low-level detail features to improve detection accuracy. In order to make lightweight improvements to YOLOv8n, the original Bottleneck structure in the C2f module is replaced with AK_Bottleneck, thus creating a new C2f_AK module. The structures of C2f_AK and AK_Bottleneck are as follows: Figure 3 shown.

[0060] The structure of the C2f_AK module includes two convolutional layers and several AK_Botleneck modules, where the AK_Botleneck module is a residual block used to increase the depth and complexity of the network. The principle of the C2f_AK module is to extract and fuse the information in the input feature map through convolution operations and iterations of the AK_Botleneck module to obtain a richer feature map, thereby improving the accuracy and efficiency of target detection. The AK_Bottleneck structure is based on an innovative lightweight network technology, the Alterable Kernel Convolution (AKConv) module, which aims to optimize the performance and efficiency of the network by dynamically adjusting the shape and size of the convolution kernel. The core principle of AKConv technology lies in its unique ability to allow the convolution kernel to dynamically change shape and parameter configuration according to the requirements of the input features. This flexibility allows AKConv to not only adapt to different data characteristics, but also optimize the use of computing resources according to task requirements, so that the number of convolution parameters decreases linearly. The AKConv structure is as follows Figure 4 As shown in the figure. AKConv defines the initial position for convolution kernels of any size through a new coordinate generation algorithm. And in order to adapt to the changes in the target, AKConv introduces an offset to adjust the shape of the sample at each position. The introduction of this method allows the C2f_AK module to maintain the powerful detection capability of the YOLOv8n model while greatly improving its lightweight performance.

[0061] In the feature fusion module of YOLOv8n, Path Aggregation Network (PANet) is used, which shows significant advantages compared with the traditional Feature Pyramid Network (FPN). Figure 5 As shown in the figure, PANet effectively improves the fusion process between features of different scales through its innovative bidirectional information flow mechanism. This mechanism not only promotes the transmission of low-level detail information to high-level, but also ensures that the semantic information of high-level can enhance the expression of low-level features, so that the model performs better when processing images with rich details and complex backgrounds. Although PANet has made significant progress in multi-scale feature fusion, there is still room for improvement. First, PANet fails to fully demonstrate its potential when processing large-scale feature maps, and may ignore some key information, thus affecting the overall performance of object detection. In addition, the feature map will lose part of its original information during the process of upsampling and downsampling. This loss of information reduces the efficiency of feature map reuse.

[0062] In order to effectively solve the above problems, this paper reconstructs the feature fusion part of YOLOv8n based on the Bidirectional Feature Pyramid Network (BiFPN). Compared with PANet, BiFPN has key optimizations and improvements. First, it removes nodes that have only a single input edge and do not participate in feature fusion. This approach is based on a simple idea: if a node cannot contribute to the fusion of different features, it will have a relatively small effect on the network structure for the purpose of feature fusion. This adjustment simplifies the network structure and forms a more efficient bidirectional network. Next, BiFPN adds additional connections between the original input and output nodes at the same level, which can achieve effective fusion of more features without significantly increasing the cost. Finally, unlike PANet, which has only one top-down and one bottom-up path, BiFPN treats each bidirectional path (top-down and bottom-up) as a separate feature network layer and promotes higher-level feature fusion by repeating these layers multiple times.

[0063] The original YOLOv8n detection head adopts a decoupled head structure, such as Figure 6 As shown in Figure 1. The decoupled head structure processes category prediction and bounding box regression as two independent tasks, each of which is calculated through its own independent convolutional block. The main advantage of this design is that it provides a dedicated network path for each task, allowing the model to learn more specialized feature representations for category classification and spatial positioning, which has the potential to improve the overall performance of the model.

[0064] However, this separation of the decoupled head structure also brings significant disadvantages. First, the independent network path for each task increases the total number of model parameters, thereby increasing the amount of computation and storage requirements. This is particularly problematic in environments with limited computing resources and may limit the deployment and application of the model. Second, while task specificity helps improve performance, it also leads to reduced flexibility in the model optimization process because the network layers for category prediction and bounding box regression cannot share information.

[0065] In view of these considerations, the proposed improvement aims to overcome these challenges by introducing a shared parameter structure while retaining some key advantages of the decoupled head structure. Figure 6 As shown in (b). By sharing parameters in the front part of the category prediction and bounding box regression branches, the design reduces the total number of parameters and computational burden of the model, while retaining the task-specific terminal layers to ensure that the model can learn effective feature representations for each task. This approach not only improves the computational efficiency of the model, but also maintains sufficient task specificity to achieve high performance.

[0066] When processing aerial images of power transmission lines, the complexity of the background becomes a significant challenge. The presence of trees, grass, and other power facilities often interferes with model detection, resulting in poor performance in the task of suppressing irrelevant background information. In order to improve the model's focus on key features and reduce its reliance on background information, the Coordinate Attention (CA) mechanism

[30] is introduced into the backbone network of the model. The CA attention mechanism module is a lightweight attention mechanism that focuses on the channel and spatial information of the image, which makes the model perform better when dealing with complex and blurred backgrounds. The structure of the CA attention mechanism module is shown in the figure. Figure 7 shown.

[0067] The implementation of the CA attention mechanism includes two main stages: coordinate information embedding and coordinate attention generation. In the coordinate information embedding stage, although global pooling is often used in channel attention to globally encode spatial information, it compresses this information into a single channel descriptor, making it difficult to maintain the integrity of the position information. However, preserving position information is crucial for capturing spatial structures in visual tasks. Therefore, for a given input X, the CA attention mechanism uses two scales of pooling kernels (H, 1) and (1, W) to encode each channel in the horizontal and vertical directions, respectively, to enhance the capture of spatial information. Entering the coordinate attention generation stage, the feature maps obtained from the coordinate information embedding stage are first merged, and then the dimensionality is reduced by convolution transformation to simplify the complexity of the model. Subsequently, after batch normalization and nonlinear activation, an intermediate feature map containing rich information is generated. The feature map is then split into two independent tensors along the spatial dimension. Finally, the two tensors are processed by two 1×1 convolutions and Sigmoid activation functions respectively, so that they are converted into tensors with the same number of channels as the input X, thus completing the construction of the entire CA attention mechanism.

[0068] In this paper, the CA attention mechanism is integrated into the backbone network. One major advantage of this is that as a lightweight attention mechanism, the CA attention mechanism can significantly improve the model's ability to focus on key information of abnormal targets while taking into account the limited computing and storage resources of the drone hardware platform, thereby improving detection performance. In order to maximize the use of this attention mechanism, the CA attention mechanism is chosen to be used at the end of the model backbone network in order to achieve better performance.

[0069] For model training, we used Intel(R) Xeon(R) Platinum 8375C processor and NVIDIA GeForceRTX4090 graphics card. The deep learning model framework used Pytorch 2.1.1 and Python 3.10, CUDA version 11.8, and operating system Ubuntu 22.04. To ensure the fairness and comparability of the model effect, we did not use pre-trained weights for the training process of various models in all ablation experiments and comparative experiments. Some important hyperparameter settings of the model during the training phase are shown in Table 1.

[0070] Table 1

[0071]

[0072]

[0073] In order to adapt to different application scenarios and hardware device requirements, the YOLOv8n model has been specifically designed. By adjusting the two key parameters of width and depth, five different scale model versions have been derived, named YOLOv8n, YOLOv8s, YOLOv8m, YOLOv8l and YOLOv8x. The number of parameters and resource consumption of each version increases accordingly to meet the requirements of different detection performance. The width, depth and maximum number of channels corresponding to these five models are shown in Table 2.

[0074] Table 2

[0075]

[0076] Considering that edge computing devices are usually limited by computing resources and energy consumption, we chose the lightest YOLOv8n as the baseline model when improving the lightweight model and preparing for deployment.

[0077] Evaluation Metrics

[0078] In order to comprehensively evaluate the performance of the abnormal target detection model for transmission lines, this study not only focuses on the accuracy of the model, but also takes into account the lightweight requirements of the model for deployment in embedded devices. Therefore, precision, recall, average precision (AP), mean average precision (mAP), FLOPs, Params and model size are selected as key evaluation indicators.

[0079] Precision refers to the proportion of all predicted targets correctly identified by the model, and the calculation formula is as follows. Among them, true positive (TP) is the number of abnormal targets of the transmission line correctly identified by the model, while false positive (FP) and false negative (FN) represent the number of abnormal targets that actually exist but are incorrectly identified and missed by the model, respectively.

[0080]

[0081] Recall refers to the proportion of all actual targets correctly identified by the model. The calculation formula is as follows.

[0082]

[0083] AP is equal to the area under the precision-recall curve. The closer its value is to 1, the better the model performance. The calculation formula is as follows.

[0084]

[0085] mAP is the average value of multi-category AP and is a commonly used evaluation indicator in target detection. It intuitively reflects the performance of the current model. The calculation formula is as follows.

[0086]

[0087] FLOPs can reflect the computational complexity of the model, and its calculation formula is as follows.

[0088] FLOPs = 2 × H × W (C in K 2 +1)C out

[0089] Model lightweighting is evaluated by the number of parameters (Params), and the calculation formula is as follows.

[0090] Params=C in ×K 2 ×C out

[0091] In addition, model size can be used to measure the ease of integration of the model and its suitability for use in lightweight devices.

[0092] S3, target identification: deploy the model to the edge computing platform carried by the drone to identify broken insulators of transmission lines and abnormal bird nests in real time;

[0093] (1) Model deployment: Deploy the trained and tested model to the edge computing platform carried by the drone. This process may involve some technical challenges, such as model compression and software and hardware compatibility. Model compression is to enable the model to run on the resource-constrained edge computing platform, while software and hardware compatibility ensures smooth model deployment in terms of computing power, memory size, etc.

[0094] (2) Image acquisition: During the flight, the drone uses the high-definition camera on board to obtain real-time images of the power transmission line. These images will be transmitted to the edge computing platform in real time.

[0095] (3) Real-time recognition: The edge computing platform inputs the received real-time images into the model, and the model identifies abnormal targets in the images through calculations, such as broken insulators, bird nests, and other abnormal targets.

[0096] S4. Feedback results: Feedback abnormal target identification results to security managers to provide support for security prevention and control.

[0097] Abnormal result generation: When the edge computing platform on the drone completes the recognition and analysis of the image, it will generate an abnormal target recognition result. This result contains a variety of information, such as the type of abnormality (such as broken insulators of high-voltage transmission lines, bird nests, etc.), the location of the abnormality, and the time of recognition.

[0098] Result transmission: The results of abnormal target identification are transmitted to the ground control center in real time via wireless network. Here, the coverage and stability of the network need to be considered to ensure the successful transmission of information. In addition, sufficient security measures are also required to prevent the information from being stolen or tampered with during the transmission process.

[0099] Result integration: For long-term or large-scale inspections, many abnormal identification results may be generated. What needs to be done at this time is to organize these results and analyze, for example, the total number of abnormalities, the distribution of various abnormalities, etc.

[0100] Result display: In the ground control center, there is a special display screen or the software is used to display the received abnormal identification results on the computer for the safety management personnel to review.

[0101] Response measures: Based on the feedback results, safety managers can determine whether further measures are needed, such as sending personnel for on-site inspections, adjusting the operating status of transmission lines, or initiating maintenance procedures. The timeliness and accuracy of feedback results will directly affect the effectiveness of these decisions.

[0102] Those skilled in the art can understand that all or part of the processes in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a computer program, and the program can be stored in a computer-readable storage medium, and when the program is executed, it can include the processes of the embodiments of the above-mentioned methods. The storage medium can be a disk, an optical disk, a read-only memory (ROM) or a random access memory (RAM).

[0103] It should be understood that the detailed description of the technical solutions of the present invention by means of the preferred embodiments is illustrative rather than restrictive. A person skilled in the art may modify the technical solutions described in the embodiments, or replace some of the technical features by equivalents, based on reading the specification of the present invention; and these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for detecting abnormal targets on power transmission lines using an edge computing platform carried by a drone, characterized in that: The method includes: Sample preparation: collect various abnormal target image data, prepare various abnormal target samples in high-voltage transmission lines and their forest channels, and divide them into training samples, test samples and verification samples; Model building: By improving the YOLOv8n model, a lightweight model architecture is formed and the model is trained, tested and verified through the established samples; The model is established using the YOLO-PowerLite model; The YOLO-PowerLite model is improved based on the YOLOv8n model; (1) In the feature fusion part, the original Bottleneck structure in the C2f module is replaced with AK_Bottleneck, thus creating a new C2f_AK module; (2) Optimize the decoupling head structure used in the original YOLOv8n model; (3) The CA attention mechanism is introduced into the backbone network of the YOLO-PowerLite model. The C2f_AK module structure includes: Two convolutional layers and several AK_Botleneck modules, where the AK_Botleneck module is a residual block; The C2f_AK module extracts and fuses the information in the input feature map through convolution operations and iterations of the AK_Botleneck module; Target recognition: Deploy the model to the edge computing platform carried by the drone to identify broken insulators on power transmission lines and abnormal bird nests in real time; Feedback results: Feedback abnormal target identification results to security managers to provide support for security prevention and control.

2. The method for detecting abnormal targets on power transmission lines using an edge computing platform mounted on a drone according to claim 1 is characterized in that: The sample preparation includes: Collect image data: Image data comes from images captured by drones through their high-definition cameras during the inspection of transmission lines; Sample creation: pre-processing the collected image data, including cleaning and labeling steps; cleaning is to remove irrelevant information in the image; labeling is to mark the position of abnormal targets in the image; Sample division: Divide the prepared samples into training samples, test samples, and validation samples in a ratio of 70%, 15%, and 15%; training samples are used to train the model, test samples are used to test the performance of the model, and validation samples are used to verify the model during training to prevent overfitting.

3. The method for detecting abnormal targets on power transmission lines using an edge computing platform mounted on a drone according to claim 1 is characterized in that: The decoupled head structure is optimized to reduce the total number of parameters and computational burden of the model by sharing parameters in the front part of the category prediction and bounding box regression branches; At the same time, by retaining the task-specific terminal layers, we ensure that the model learns effective feature representations for each task; improving the computational efficiency of the model and maintaining sufficient task specificity to achieve high performance.

4. The method for detecting abnormal targets on power transmission lines using an edge computing platform mounted on a drone according to claim 1 is characterized in that: The CA attention mechanism introduced in the backbone network includes two stages: Coordinate information embedding and coordinate attention generation; In the coordinate information embedding stage, for a given input X, the CA attention mechanism uses two scales of pooling kernels (H, 1) and (1, W) to encode each channel in the horizontal and vertical directions respectively to enhance the capture of spatial information; In the coordinate attention generation stage, the feature maps obtained from the coordinate information embedding stage are first merged, and then the dimensionality is reduced by convolution transformation; After batch normalization and nonlinear activation processing, an intermediate feature map is generated.

5. The method for detecting abnormal targets on power transmission lines using an edge computing platform mounted on a drone according to claim 1 is characterized in that: The target identification includes: Model deployment: Deploy the trained and tested model to the edge computing platform carried by the drone; Image acquisition: During the flight, the drone uses the high-definition camera on board to obtain real-time images of the power transmission line and transmits them to the edge computing platform in real time; Real-time recognition: The edge computing platform inputs the received real-time images into the model, and the model identifies abnormal targets in the images through calculation.

6. The method for detecting abnormal targets on power transmission lines using an edge computing platform mounted on a drone according to claim 1 is characterized in that: The feedback results include: Abnormal result generation: When the edge computing platform on the drone completes the recognition and analysis of the image, it generates abnormal target recognition results; Result transmission: The results of abnormal target identification are transmitted to the ground control center in real time via wireless network; Result integration: Long-term or large-scale inspections will produce many abnormal identification results; the results are sorted and the total number of abnormalities and the distribution of various abnormalities are analyzed.

Images