A power transmission line external damage detection and early warning edge computing method
By combining a lightweight convolutional neural network model with residual connections and feature fusion modules, the shortcomings of transmission line external damage detection in terms of real-time performance and accuracy are solved, and high-precision detection and timely early warning of external damage targets are achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- STATE GRID SHAANXI ELECTRIC POWER CO LTD ECONOMIC & TECHNICAL RESEARCH INSTITUTE
- Filing Date
- 2023-05-17
- Publication Date
- 2026-05-15
AI Technical Summary
Existing deep learning-based methods for detecting external damage targets on power transmission lines have shortcomings in terms of real-time performance and accuracy. They are particularly difficult to meet the requirements of real-time performance and high accuracy in complex field scenarios, and existing lightweight models have reduced detection accuracy.
A lightweight convolutional neural network model is adopted, combined with residual connections and feature fusion modules. Through a feature extraction backbone network, a feature fusion enhancement module, and a module for predicting the location and category of external damage targets, high-precision detection of external damage targets such as construction machinery, tower cranes, cranes, wildfires, and birds is achieved.
It enables real-time, high-precision detection of external damage targets on transmission lines on edge computing devices, providing timely early warnings and reducing the number of model parameters and computational costs.
Smart Images
Figure CN116740631B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the technical field of intelligent monitoring of power systems, and particularly relates to an edge computing method for detecting and warning of external damage to transmission lines based on image processing and computer vision. Background Technology
[0002] With the rapid development of my country's urban and rural economies and the continuous expansion of its power grid, transmission lines are covering an increasingly wide area, often spanning multiple provinces and located far from main roads. Compared to substation systems, transmission lines are characterized by their length, wide coverage, and frequent exposure to the elements, making them susceptible to damage from external forces such as construction machinery and natural disasters, which can lead to power grid safety accidents. Traditional transmission line inspections mainly include manual inspections, video photography, and helicopter aerial photography. However, these three methods require subjective judgment from personnel regarding the lines and equipment, which can easily result in missed or incorrect assessments, making it impossible to detect potential hazards in transmission lines in real time and provide timely warnings. Therefore, as a crucial component of the safe operation of the power grid, the detection and timely warning of external damage to transmission lines is of paramount importance.
[0003] In recent years, with the rapid development of artificial intelligence, image processing, and computer vision technologies, research on external damage detection technology for transmission lines has emerged. Existing technologies include "a method for detecting external force damage to transmission line towers based on convolutional neural networks," which discloses a method for detecting external force damage to transmission line towers and can achieve classification detection of such damage. Existing technologies also include "a method and device for alarming transmission lines based on deep learning," which uses a trained external damage target detection model to obtain the type and location of all first-order external damage targets. Furthermore, existing technologies propose "a method, device, equipment, and medium for monitoring and early warning of external damage to transmission lines using a satellite-ground coordinated approach," which analyzes optical satellite remote sensing orthophotos using a pre-trained convolutional neural network model to obtain the spatial distribution range of preset targets.
[0004] The existing deep learning-based methods for detecting external damage targets on transmission lines suffer from high computational complexity and large network model parameters, making them unsuitable for real-time edge computing requirements in engineering applications. To address the need for deep learning network models in mobile or edge embedded devices with limited storage and computing power, lightweight development of convolutional network models has emerged in recent years. The existing technology, "A method for detecting foreign objects on transmission lines based on improved YOLOv4," reduces the number of parameters to one-eighth of that of ordinary convolutions by introducing depthwise separable convolutions, thus achieving a lightweight network model. The existing technology, "A method for detecting abnormal targets on transmission lines based on improved YOLOv3," discloses a method for detecting abnormal targets on transmission lines that uses a large number of depthwise separable convolutions to replace standard convolutions to reduce the number of parameters and achieve a lightweight YOLOv3 model. The existing technology of "lightweight detection model, method, device and storage medium for small target defects" integrates depthwise separable convolution and pointwise convolution to reduce the number of model parameters. It also designs a lightweight module DC-SE by using a channel number multiplication strategy and channel attention mechanism SE to reduce the interference of complex background on insulator faults.
[0005] In existing technologies for power transmission line target detection, replacing standard convolutions with depthwise separable convolutions reduces the number of parameters and makes the network model lighter, but it also lowers the detection accuracy. This is particularly problematic in outdoor environments where power transmission lines are susceptible to various environmental interferences, placing higher demands on the real-time performance and accuracy of detection algorithms for external damage targets. Furthermore, considering that external damage detection models need to predict not only the category but also the location of the damaged target, and that shallow features in deep learning networks provide rich location information while deep features provide rich semantic category information, fusing semantically rich deep features with shallow features containing precise location information may result in excessively long information propagation paths, negatively impacting the algorithm's performance.
[0006] In summary, considering the complexity of transmission line scene images, overcoming the shortcomings of existing technologies and inventing an edge computing method for detecting and warning of external damage targets on transmission lines that is suitable for transmission line scenes and takes into account both real-time performance and high accuracy is a key problem that urgently needs to be solved in this field. Summary of the Invention
[0007] In view of the above problems, this invention proposes an edge computing method applicable to power transmission line scenarios for detecting and warning of external damage targets. The method includes: establishing a power transmission line image dataset to train, validate, and test a power transmission line external damage target detection model; constructing a lightweight power transmission line external damage detection model to detect the category and location information of external damage targets such as construction machinery, tower cranes, cranes, wildfires, and birds; detecting safe areas of the power transmission line to detect the power transmission lines in the acquired images and identify safe areas; and designing a power transmission line external damage warning module, which designs a warning strategy and generates a warning signal based on the detected categories and locations of external damage targets and the safe areas of the power transmission line.
[0008] Specifically, the present invention employs the following technical solution to realize the above-mentioned method for detecting and warning of external damage to transmission lines, including:
[0009] S1. Establish a power transmission line image dataset to enable the training, verification, and testing of the power transmission line external damage target detection model.
[0010] S2. Establish a lightweight transmission line external damage detection model to detect the category and location information of external damage targets such as construction machinery, tower cranes, cranes, wildfires, and birds. The aforementioned lightweight transmission line external damage detection model includes a feature extraction backbone network, a feature fusion enhancement module, and a module for predicting the location and category of external damage targets. Considering that in transmission line scene images, external damage equipment such as cranes, tower cranes, and construction machinery are generally large targets, shallow features are beneficial for detecting large targets. While Feature Pyramid Networks (FPNs) can combine semantically rich features with precise location information, it may lead to excessively long information propagation paths. Residual learning, through residual blocks and identity mappings, can reduce model complexity on the one hand, and relax information propagation between different network layers on the other. Therefore, in the above model, the feature extraction backbone network is mainly constructed based on residual connections. In the aforementioned feature fusion enhancement module, a bottom-up propagation path is used to enhance the transmission of shallow location information to deeper layers, while a top-down path is used to enhance the propagation of rich semantic category information.
[0011] S3. The transmission line safety area detection module is designed to detect transmission lines and identify safety areas in the acquired transmission line scene images. This includes steps such as transmission line image preprocessing, edge detection, transmission line detection, and transmission line safety area identification.
[0012] S4. Design of the transmission line external damage prevention early warning module: Based on the detected external damage target categories and location information of the transmission line and the safety zone identification of the transmission line, an early warning strategy is designed to generate an early warning signal.
[0013] Further, step S1 includes:
[0014] S1-1. Acquire images of power transmission line scenes;
[0015] Specifically, cameras are fixed at appropriate locations on the transmission line towers to collect a certain number of images of external targets that may damage the transmission line, including or including construction machinery, tower cranes, cranes, wildfires, birds, etc.
[0016] S1-2, Preprocessing of transmission line images;
[0017] Specifically, the above-mentioned power transmission line scene images are preprocessed by filtering, denoising, and format conversion;
[0018] S1-3, Transmission line image dataset;
[0019] Specifically, the preprocessed transmission line images are labeled and divided into training set, validation set and test set;
[0020] S1-4, Enhancement of transmission line image dataset;
[0021] Specifically, the training set and validation set divided in steps S1-3 above are expanded and enhanced using image enhancement technology to train and validate the transmission line external damage target detection model.
[0022] Further, step S2 includes:
[0023] S2-1. Construct the backbone network for feature extraction of the external damage detection model for transmission lines;
[0024] Specifically, the enhanced transmission line images from the S1-4 dataset are used as input, and convolution and pooling operations are performed to extract effective features layer by layer. Residual connections are introduced to establish residual blocks, reducing model complexity. These residual blocks consist of a residual part and an identity mapping part. The residual part includes two or more convolution operations, while the identity mapping part directly connects the input to the output part through short connections. These short connections relax the propagation of feature information between different layers, facilitating the fusion of location and category features of externally damaged targets on transmission lines.
[0025] S2-2, Feature fusion module for external damage detection model of transmission lines;
[0026] Specifically, the feature fusion module is used to aggregate features extracted from different layers of the S2-1 backbone network and combine bottom-up and top-down methods to achieve feature fusion of different layers.
[0027] Considering that the external damage detection model for transmission lines needs to predict not only the category information of the external damage target, but also the location information of the external damage target, accurate location information of the external damage target provides an important basis for whether to implement an early warning. The aforementioned shallow features with rich location information are beneficial for the identification of large external damage targets such as general cranes, tower cranes, construction machinery, and other construction equipment in transmission line scene images. Therefore, this feature fusion module draws on the bottom-up and top-down feature aggregation methods of PathAggregation Network (PANet), and also adopts the tensor connection method in YOLOv4 to achieve feature stacking. The feature fusion module includes bottom-up path enhancement, adaptive pooling, and fully connected fusion.
[0028] Furthermore, specifically, the aforementioned bottom-up path enhancement adds a top-up shallow feature information propagation path on the basis of FPN to achieve the propagation of location information of large external targets such as general cranes, tower cranes and other construction machinery.
[0029] Furthermore, specifically, the aforementioned adaptive pooling aggregates multi-layer feature information for each candidate box through simple connections, avoiding the drawback of arbitrarily assigning different feature information in FPN, which is beneficial for the transmission line external damage detection model to output more accurate location and category information.
[0030] Furthermore, specifically, the aforementioned fully connected fusion consists of a main branch and a sub-branch. The main branch includes four 3×3 convolutions and one deconvolution, used to predict the class mask information of the breached target. The sub-branch includes two 3×3 convolutions and one fully connected layer, used to predict the position mask information for segmenting the breached target (foreground) from the background. Further, the class mask output from the main branch and the position mask output from the sub-branch are combined using a unit summation operation and a sigmoid activation function to output the feature mask of the breached target.
[0031] S2-3, Lightweight Transmission Line External Damage Detection Model Prediction Module;
[0032] Specifically, this module considers the first-order detector YOLO-Head to perform regression and classification prediction on the multi-scale features output by the above S2-2 feature fusion module, so as to realize the category and location information of the external damage target of the transmission line.
[0033] S2-4 Training and Validation of Lightweight Transmission Line External Damage Detection Model;
[0034] Specifically, the lightweight convolutional neural network model is pre-trained using a natural image dataset to obtain a pre-trained model. Then, transfer learning techniques are used to fine-tune the pre-trained model using the training set in the enhanced transmission line image dataset from steps S1-4 to obtain a trained lightweight convolutional neural network model. Finally, the model is validated using the validation set in the enhanced transmission line image dataset from steps S1-4.
[0035] Further, step S3 includes:
[0036] S3-1, Preprocessing of images of centralized power transmission line scenes;
[0037] Specifically, in order to further improve the real-time performance of external damage detection and early warning in transmission line scene images, during the testing phase, each transmission line scene image in the test set divided in steps S1-3 above is first processed into grayscale; then, the grayscale processed images are preprocessed with filtering, noise reduction and other preprocessing operations.
[0038] S3-2, Power transmission line inspection;
[0039] Specifically, the image after preprocessing in step S3-1 is subjected to edge detection using an edge detection operator to obtain a binary image containing rich edge information. The binary image is then subjected to post-processing such as filtering and denoising to eliminate noise points, scattered points, and broken lines in the edge detection image.
[0040] Specifically, line detection is performed on the edge binary image obtained from the above steps to detect power transmission lines.
[0041] S3-3, Safety zone inspection of power transmission lines;
[0042] Specifically, the straight-line image of the transmission line obtained in step S3-3 is processed to obtain several rectangular regions that may or may not contain the transmission line, and duplicate and redundant boxes are suppressed. A geometric constraint strategy that conforms to the geometric characteristics of the transmission line is designed to further filter and obtain the location information of the transmission line region. The perimeter of the detected transmission line region is then expanded outward to obtain a local safety warning boundary for the transmission line with sufficient margin.
[0043] Further, step S4 includes:
[0044] S4-1, Detection of external damage targets on power transmission lines;
[0045] Specifically, each transmission line scene image in the test set divided in steps S1-3 is input into the lightweight convolutional neural network model trained and validated in steps S2-4 for testing, so as to obtain the category and location information of external damage targets to transmission lines, such as construction machinery, tower cranes, cranes, wildfires, and birds.
[0046] S4-2, Early warning judgment for external damage to transmission lines;
[0047] Specifically, the detection frame area of each external damage target detected in step S4-1 above is subjected to a mathematical intersection operation with the local warning area for transmission line safety obtained in step S3-4 above. If the result of the intersection operation is not empty, a warning signal is generated; otherwise, no warning signal is generated.
[0048] Compared with the prior art, the present invention has the following advantages:
[0049] This invention is based on a lightweight convolutional neural network that can achieve high-precision detection of the category and location information of multiple external damage targets on power transmission lines, such as construction machinery, tower cranes, cranes, wildfires, and birds. The model has fewer parameters and relatively low computational cost. The trained model is expected to be integrated into edge computing devices to achieve real-time and high-precision detection of external damage targets on power transmission lines.
[0050] This invention combines a lightweight convolutional neural network model with high detection accuracy with a transmission line safety area detection method based on traditional image processing techniques such as edge detection, Hough transform, and minimum bounding rectangle, enabling timely early warning of external damage targets on transmission lines. Attached Figure Description
[0051] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments will be briefly described below.
[0052] Figures 1-4 This is a schematic diagram of each step in the edge calculation method for detecting and warning of external damage to transmission lines provided in an embodiment of the present invention.
[0053] Figure 5 This is a schematic diagram of the edge calculation method for detecting and warning of external damage to transmission lines according to an embodiment of the present invention. Detailed Implementation
[0054] To facilitate understanding by those skilled in the art, the present invention will be further described in detail below with reference to specific embodiments. The following specific embodiments are for illustrative purposes only and are not intended to represent all embodiments, nor should they be considered as limiting the scope of the invention.
[0055] This embodiment provides a method for detecting and warning of external damage to transmission lines, which includes the following steps: S1, establishing a transmission line image dataset; S2, establishing a lightweight convolutional neural network model; S3, detecting safe areas of transmission lines; S4, detecting external damage targets of transmission lines and making warning decisions.
[0056] Further, step S1 includes:
[0057] S1-1. Acquire images of power transmission line scenes;
[0058] Specifically, cameras are fixedly installed at appropriate locations on the transmission line towers to capture images of the transmission line scene.
[0059] S1-2, Preprocessing of transmission line images;
[0060] Specifically, the images in the above-mentioned power transmission line scene are screened, and images that do not include external targets such as construction machinery, tower cranes, cranes, wildfires, and birds are removed, as well as blurry or severely obscured images. Considering that the power transmission line scene is complex and susceptible to interference, noise, or severe weather, image enhancement techniques such as filtering and denoising are used to enhance each image in the power transmission line scene, and preprocessing operations such as data format conversion are performed.
[0061] S1-3. Establish a power transmission line image dataset;
[0062] Specifically, the labelimg annotation software was used to manually annotate the transmission line images after the preprocessing in steps S1-2 above. Rectangular bounding boxes were used to enclose the locations of external targets in the transmission line images, such as construction machinery, tower cranes, cranes, wildfires, birds, and foreign objects on the conductors, and category labels were assigned. The annotated transmission line image dataset was then divided into training, validation, and test sets in a 7:2:1 ratio.
[0063] S1-4, Enhancement of transmission line image dataset;
[0064] Specifically, the training set and validation set divided in steps S1-3 above are expanded and enhanced using image enhancement techniques such as brightness enhancement, cropping, and rotation, increasing the number of transmission line image datasets from 1350 to approximately 4000, which are then used for training and validation of the transmission line external damage target detection model.
[0065] Further, step S2 includes:
[0066] S2-1, Lightweight transmission line external damage detection backbone network;
[0067] Specifically, considering the number of model parameters, inference speed, and robustness to extracting feature information from transmission line scene images, this embodiment takes ResNet50 as the feature extraction backbone network of the lightweight transmission line external damage detection model as an example.
[0068] S2-2, Feature Fusion Module for External Damage Detection Model of Transmission Lines;
[0069] Specifically, this embodiment selects the Neck part of YOLOv4 as the feature enhancement module. The main structure consists of the SPP module and the PANet module. It uses three feature layers of different scales obtained by the backbone network feature extraction to perform preliminary feature fusion enhancement. Then, the enhanced feature maps of the three different scales are upsampled for feature fusion. Finally, the feature maps of the three scales are reconstructed through downsampling. The fusion of multi-layer features is achieved by combining bottom-up and top-down methods.
[0070] S2-3, Lightweight Transmission Line External Damage Detection Model Prediction Module;
[0071] Specifically, in this embodiment, the YOLO Head part is used as the target detection and prediction module for external damage to transmission lines. Its main function is to obtain the output, compare it with the real data annotation, calculate the loss function (which mainly consists of three parts: ① localization loss ② confidence loss ③ classification loss), and then reshape the data format as needed. At the same time, the original grid coordinates are activated accordingly. The above-mentioned more effective feature layer is used to obtain the bounding box and category prediction of the external damage image of the transmission line scene.
[0072] S2-4 Training and Validation of Lightweight Transmission Line External Damage Detection Model;
[0073] Specifically, in this embodiment, the lightweight transmission line external damage detection model is pre-trained using the natural scene COCO dataset to obtain a pre-trained model. Then, the pre-trained model is fine-tuned using the training set in the enhanced transmission line image dataset in steps S1-4 using transfer learning techniques to obtain a trained lightweight convolutional neural network model. Finally, the model is validated using the validation set in the enhanced transmission line image dataset in steps S1-4.
[0074] Further, step S3 includes:
[0075] S3-1, Preprocessing of images of centralized power transmission line scenes;
[0076] Specifically, in order to further improve the real-time performance of external damage detection and early warning in transmission line scene images, during the testing phase, each transmission line scene image in the test set divided in steps S1-3 above is first processed into grayscale; then, the grayscale processed images are preprocessed with filtering, noise reduction and other preprocessing operations.
[0077] S3-2, Power transmission line inspection;
[0078] Specifically, the image preprocessed in step S3-1 is subjected to edge detection using an edge detection operator to obtain a binary image containing rich edge information. Considering that the Canny operator employs non-maximum suppression and a double thresholding method, resulting in good edge detection performance, this embodiment uses the Canny operator for edge detection to eliminate interference from non-edge points.
[0079] Specifically, the binary image containing edge information is subjected to post-processing such as filtering and denoising to eliminate noise points, scattered points, and broken lines in the edge detection image. Furthermore, the post-processed binary edge image is subjected to Hough transform for line detection to achieve the detection of transmission lines.
[0080] S3-3, Safety zone inspection of power transmission lines;
[0081] Specifically, the transmission line straight-line image obtained in step S3-2 above is processed using the Minimum Bounding Rectangle (MBR) algorithm to obtain several rectangular regions that may or may not contain the transmission line. Then, the greedy non-maximum suppression (NMS) algorithm is used to suppress duplicate and redundant boxes. Since the transmission line region has obvious length, width, and regional characteristics, a geometric constraint strategy that conforms to the geometric characteristics of the transmission line, such as aspect ratio, is designed to further filter and obtain the transmission line region information. The perimeter of the detected transmission line region is then expanded outward to obtain a local safety warning boundary for the transmission line with sufficient margin.
[0082] Further, step S4 includes:
[0083] S4-1, Detection of external damage targets on power transmission lines;
[0084] Specifically, each transmission line scene image in the test set divided in steps S1-3 is input into the lightweight convolutional neural network model trained and validated in steps S2-4 for testing, so as to obtain the category and location information of external damage targets to transmission lines, such as construction machinery, tower cranes, cranes, wildfires, and birds.
[0085] S4-2, Early warning judgment for external damage to transmission lines;
[0086] Specifically, the detection frame area of each external damage target detected in step S4-1 above is subjected to a mathematical intersection operation with the local warning area for transmission line safety obtained in step S3-3 above. If the result of the intersection operation is not empty, a warning signal is generated; otherwise, no warning signal is generated.
[0087] Although the present invention has been described in detail through the above preferred embodiments, it should be understood that the above description should not be considered as a limitation of the present invention.
Claims
1. A method for edge calculation of external damage detection and early warning for transmission lines, characterized in that, The method includes the following steps: S1. Establish a transmission line image dataset to enable the training, verification, and testing of the transmission line external damage target detection model; S2. Establish a lightweight transmission line external damage detection model to detect the category and location information of external damage targets such as construction machinery, tower cranes, cranes, wildfires, and birds. The lightweight transmission line external damage detection model includes a feature extraction backbone network, a feature fusion enhancement module, and an external damage target location and category prediction module. In the above model, the feature extraction backbone network is mainly constructed with residual connections. In the feature fusion enhancement module, a bottom-up propagation path is used to enhance the transmission of shallow location information to deeper layers, while a top-down path is used to enhance the propagation of rich semantic category information. S3. Establish a transmission line safety area detection module to realize the detection of transmission lines and the identification of safety areas in the collected transmission line scene images, including the steps of transmission line image preprocessing, edge detection, transmission line detection and transmission line safety area identification; S4. Establish a transmission line external damage prevention early warning module. Based on the detected external damage target category and location information of the transmission line and the safety zone identification of the transmission line, design an early warning strategy to generate an early warning signal. Step S1 includes: S1-1. Acquire images of power transmission line scenes; Cameras were fixed on transmission line towers to capture multiple images of external damage targets to the transmission line, including or separately including construction machinery, tower cranes, cranes, wildfires, and birds. S1-2, Image preprocessing of power transmission lines; Preprocessing of power transmission line scene images includes filtering, denoising, and format conversion; S1-3, Transmission line image dataset; The preprocessed transmission line images are labeled and divided into training, validation and test sets; S1-4, Enhancement of transmission line image dataset; The training set and validation set divided in steps S1-3 are expanded and enhanced using image enhancement technology to train and validate the transmission line external damage target detection model. Step S2 includes: S2-1. Construct the backbone network for feature extraction of the external damage detection model for transmission lines; The enhanced transmission line images from the S1-4 dataset are used as input, and convolution and pooling operations are performed on them to extract effective features layer by layer. By introducing residual connections, residual blocks are established to reduce the complexity of the model. The residual block includes a residual part and an identity mapping part. The residual part includes two or more convolution operations, and the identity mapping part directly connects the input to the output part through short connections. The aforementioned short connections play a role in relaxing the propagation of feature information between different layers, which is beneficial to the fusion of location feature information and category feature information of external damage targets of transmission lines. S2-2, Feature Fusion Module for External Damage Detection Model of Transmission Lines; The feature fusion module is used to aggregate features extracted from different layers of the S2-1 backbone network, and combine bottom-up and top-down methods to achieve feature fusion of different layers. Considering that the external damage detection model for transmission lines not only needs to predict the category information of the external damage target, but also the location information of the external damage target, accurate location information of the external damage target provides an important basis for whether to implement early warning. The shallow features of the location information are conducive to the identification of large external damage targets such as cranes, tower cranes, construction machinery and construction equipment in the transmission line scene image. Therefore, the feature fusion module draws on the bottom-up and top-down feature aggregation methods of the path aggregation network PANet, and also draws on the tensor connection method in YOLOv4 to achieve feature stacking. The feature fusion module includes bottom-up path enhancement, adaptive pooling and fully connected fusion. S2-3, Lightweight Transmission Line External Damage Detection Model Prediction Module; The prediction module considers the first-order detector YOLO-Head to perform regression and classification prediction on the multi-scale features output by the above S2-2 feature fusion module, so as to realize the category and location information of the external damage target of the transmission line; S2-4 Training and Validation of Lightweight Transmission Line External Damage Detection Model; The lightweight convolutional neural network model is pre-trained using a natural image dataset to obtain a pre-trained model. Then, the pre-trained model is fine-tuned using the training set in the enhanced transmission line image dataset from steps S1-4 using transfer learning techniques to obtain a trained lightweight convolutional neural network model. Finally, the model is validated using the validation set in the enhanced transmission line image dataset from steps S1-4.
2. The edge calculation method for external damage detection and early warning of transmission lines according to claim 1, characterized in that: Furthermore, the bottom-up path enhancement adds a top-up shallow feature information propagation path on the basis of FPN to achieve the propagation of the position information of construction machinery such as cranes and tower cranes that break large targets.
3. The edge calculation method for external damage detection and early warning of transmission lines according to claim 1, characterized in that: Furthermore, the adaptive pooling aggregates multi-layer feature information for each candidate box through simple connections, avoiding the disadvantage of arbitrarily assigning different feature information in FPN, which is conducive to the transmission line external damage detection model outputting more accurate location and category information.
4. The edge calculation method for external damage detection and early warning of transmission lines according to claim 1, characterized in that: Furthermore, the fully connected fusion consists of a main branch and a sub-branch. The main branch includes four 3×3 convolutions and one deconvolution to predict the class mask information of the breached target. The sub-branch includes two 3×3 convolutions and one fully connected layer to predict the position mask information of the breached target and the background. Furthermore, the class mask output by the main branch and the position mask output by the sub-branch are combined with a unit summation operation and the sigmoid activation function to output the feature mask of the breached target.
5. The edge calculation method for external damage detection and early warning of transmission lines according to claim 1, characterized in that: Further, step S3 includes: S3-1, Preprocessing of images of centralized power transmission line scenes; To further improve the real-time performance of external damage detection and early warning in transmission line scene images, during the testing phase, each transmission line scene image in the test set divided in steps S1-3 is first converted to grayscale; then, the grayscale images are preprocessed by filtering and denoising. S3-2, Power transmission line inspection; The image after preprocessing in step S3-1 is subjected to edge detection using an edge detection operator to obtain a binary image containing rich edge information. The binary image is then subjected to post-processing such as filtering and denoising to eliminate noise points, scattered points and broken lines in the edge detection image. Specifically, line detection is performed on the edge binary image obtained from the above steps to achieve the detection of power transmission lines; S3-3, Safety zone inspection of power transmission lines; The transmission line straight image obtained in step S3-2 is processed to obtain several rectangular regions that contain or do not contain the transmission line. Duplicate and redundant boxes are suppressed. A geometric constraint strategy that conforms to the geometric characteristics of the transmission line is designed to further filter and obtain the location information of the transmission line region. The perimeter of the detected transmission line region is expanded outward to obtain a local safety warning boundary for the transmission line with a margin.
6. The edge calculation method for external damage detection and early warning of transmission lines according to claim 5, characterized in that: Further, step S4 includes: S4-1, Detection of external damage targets on power transmission lines; Each transmission line scene image in the test set divided in steps S1-3 is input into the lightweight convolutional neural network model trained and validated in steps S2-4 for testing, so as to obtain the category and location information of external damage targets to transmission lines, such as construction machinery, tower cranes, cranes, wildfires, and birds. S4-2, Early warning judgment for external damage to transmission lines; The detection frame area of each external damage target detected in step S4-1 is mathematically intersected with the local warning area for transmission line safety obtained in step S3-4. If the result of the above intersection operation is not empty, an early warning signal is generated; otherwise, no early warning signal is generated.