Power transmission line typical fault identification method and system based on BiRG-YOLO algorithm

Abstract

The invention discloses a power transmission line typical fault identification method and system based on a BiRG-YOLO algorithm, and the method comprises the steps: obtaining the aerial image data of a power transmission line, carrying out the preprocessing of the aerial image data of the power transmission line, obtaining a processed aerial image, and dividing the processed aerial image into a training set and a test set; a YOLOv8n network is introduced, a BiFormer module and RepGhost-C2f are used for improving the YOLOv8n network, a BiRG-YOLO network is obtained, the training set is input into the BiRG-YOLO network for training, and an original fault recognition model is obtained; inputting the test set into the original recognition model, adjusting training parameters through reverse gradient updating, and optimizing the original fault recognition model to obtain a typical fault recognition model of the power transmission line; inputting the aerial image data of the power transmission line to be identified into the typical fault identification model of the power transmission line to obtain a fault identification result of the power transmission line; a BiFomrer module is introduced into the YOLOv8n, a RepGhost-C2f module is introduced into the YOLOv8n, the model obtained through final training can detect tiny faults in images in a targeted mode, and the accuracy of typical fault analysis of the power transmission line is improved.

Description

Technical Field

[0001] The present invention relates to the technical field of overhead transmission line fault detection, and particularly to a method and system for identifying typical faults of transmission lines based on the BiRG-YOLO algorithm. Background Technique

[0002] As the lifeblood of the country's economic development, the power industry has received great attention from the government in recent years. Along with the country's strategic deployment of energy security and green development, the construction and maintenance requirements of China's power grid have increased significantly. Especially with the large-scale access of new energy, the stability and security of transmission lines face higher challenges. Traditional inspections usually adopt manual inspections, where staff conduct on-site inspections and rely on technology and experience to judge faults. However, China's transmission lines are widely distributed, covering suburbs, rivers, valleys, etc. They are far away and have complex terrains, resulting in low inspection efficiency and high danger. Traditional manual inspections are no longer able to meet the current inspection needs. To effectively address this situation, the country has introduced a series of policies aimed at strengthening the intelligent and digital management of power facilities and improving the efficiency and accuracy of power inspections and monitoring. The efficient and safe operation of the power industry also relies on refined management and standardized inspections, but manual inspections have problems such as low efficiency and difficulty in comprehensive coverage. Therefore, the national energy department has actively promoted the development and application of automated and intelligent inspection equipment, aiming to implement a "prevention first" safety strategy in power grid operation and avoid accidents. With the rapid development of drone technology and computer technology, the fault inspection method for key component equipment in transmission lines based on aerial images has become the main development direction of line inspections.

[0003] In the field of traditional image processing, representative methods such as the maximum inter-class variance method, multi-salient aggregation algorithm, and binocular vision all require manual segmentation of the component image area. The algorithms are complex, the recognition effect is poor, and for insulator images with complex backgrounds, it is necessary to manually determine and extract feature quantities for the given images, which will cause errors, resulting in low effectiveness of feature extraction and low fault recognition rate. The insulator components in transmission lines are small, and the backgrounds of aerial images are complex. The segmentation and recognition methods corresponding to insulator images with different backgrounds vary, with poor generalization and practicability, unable to achieve intelligent detection, and single detection types, being greatly affected by the external environment and difficult to be put into actual application.

[0004] With the development of deep learning, the application of deep learning algorithms in the component recognition of transmission lines has gradually attracted the attention of relevant scholars in various research fields. Deep learning does not require manual feature extraction, only a large amount of labeled data, and trains a model with as excellent performance as possible. Compared with traditional image processing techniques, the deep neural network of deep learning has stronger learning ability, so that more comprehensive image features can be obtained. Based on the object detection algorithm of deep learning, its task is to find all objects in the image when the objects have different postures, appearances, shapes and the environment has different brightness and angles, and determine their categories and positions. It is divided into two methods: Two-stage and One-stage. Two-stage is a region proposal-based method, and the iconic algorithms are RCNN, Fast RCNN and Faster RCNN. The core idea is to obtain the proposed region first and then classify within the current region, but the detection time is longer. One-stage is a method without region proposal, and the iconic algorithms are SSD and YOLO. The core idea is to directly predict the position and attributes of the object based on the entire image with a single convolutional network. At present, YOLOv8 has attracted attention in the industrial field due to its excellent performance and ease of deployment, and its accuracy and speed have been greatly improved compared with the previous YOLO networks. However, the background of the aerial images of transmission lines is complex, and the typical fault components of transmission lines account for a small proportion in the whole aerial image, which belongs to small object detection. These factors have a great impact on the fault detection of transmission lines. At present, due to the above reasons, most of the current algorithms and systems have low fault detection accuracy and are difficult to be actually used in line inspection. Summary of the Invention

[0005] The present invention provides a method and system for identifying typical faults in transmission lines based on the BiRG-YOLO algorithm, so as to overcome the technical problems that the positioning and detection accuracy of the fault area are affected due to the complex background and small proportion of the components to be detected in the aerial images of transmission lines, and it is difficult to be actually used in line inspection.

[0006] In order to achieve the above object, the technical solution of the present invention is:

[0007] A method for identifying typical faults in transmission lines based on the BiRG-YOLO algorithm, comprising:

[0008] S1: Obtain the aerial image data of the transmission line, preprocess the aerial image data of the transmission line to obtain the processed aerial image, and divide it into a training set and a test set;

[0009] S2: Introduce the YOLOv8n network, and improve the YOLOv8n network using the BiFormer module and RepGhost-C2f to obtain the BiRG-YOLO network. Input the training set into the BiRG-YOLO network for training to obtain the original fault recognition model;

[0010] S3: Input the test set into the original recognition model, adjust the training parameters through reverse gradient update, and optimize the original fault recognition model to obtain the transmission line typical fault recognition model;

[0011] S4: Input the aerial image data of the transmission line to be recognized into the transmission line typical fault recognition model to obtain the fault recognition result of the transmission line.

[0012] Further, introducing the YOLOv8n network and using the BiFormer module and RepGhost-C2f to improve the YOLOv8n network to obtain the BiRG-YOLO network includes:

[0013] The BiRG-YOLO network includes an input end, a Backbone network, a Neck network, and a Head network;

[0014] The input end is used to process the image and adjust the size of the image;

[0015] The Backbone network is used to reduce the resolution of the image, extract the features of the targets in the image, and output feature maps of three sizes;

[0016] The Neck network is the same as the Neck network in the original YOLOv8n network, including the FPN network and the PAN network. It receives the feature maps of three sizes output from different layers in the Backbone network, performs bidirectional feature fusion through the FPN network and the PAN network, and fuses the features in each size of the feature map into the other two sizes of the feature maps respectively, and outputs the fused feature maps of three sizes and passes them to the Head network;

[0017] The Head network is the same as the Head network in the original YOLOv8n network, including three groups of decoupled heads that respectively process the fused feature maps of three sizes. Each group of decoupled heads includes a convolutional layer, a classification branch, and a regression branch, and is used to predict the fused feature maps of three sizes, including representing the detected target positions using bounding boxes, outputting the class probabilities of each bounding box, and the confidence levels of the detected targets;

[0018] The Bottleneck module in the C2f module of the original YOLOv8n network is replaced with the RepGhostBottleneck to form the RepGhost-C2f module;

[0019] In the Backbone network of the original YOLOv8n network, the first and second C2f modules are replaced with BiFormer modules, and the third and fourth C2f modules are replaced with RepGhost-C2f modules;

[0020] The improved Backbone network then includes a first convolutional module, a second convolutional module, a first BiFormer module, a third convolutional module, a second BiFormer module, a fourth convolutional module, a first RepGhost-C2f module, a fifth convolutional module, a second RepGhost-C2f module, and an SPPF module connected in sequence.

[0021] Furthermore, the training set is input into the BiRG-YOLO network for training to obtain the original fault recognition model, including:

[0022] At the input end, the Mosaic data augmentation method is used to perform random cropping and splicing operations on the input images;

[0023] The processed images are input into the Backbone network to extract semantic features at different levels of the images, obtaining three sizes of basic feature maps;

[0024] The three sizes of basic feature maps are input into the Neck network for bidirectional feature fusion to obtain three sizes of fused feature maps;

[0025] The three sizes of fused feature maps are input into the Head network, divided into feature maps of S×S grids. Through the adaptive anchor box strategy, according to the width and height of the true boxes in the training set, anchor boxes of different sizes of features are generated. The class probability and confidence of each anchor box are predicted through the classification branch; the coordinates of the bounding box corresponding to the anchor box are predicted through the regression branch; the target boxes are screened out through the non-maximum suppression method, the coordinates of the target boxes are obtained, and the bounding boxes are drawn in the original image to frame the targets.

[0026] Furthermore, the test set is input into the original recognition model, and the training parameters are adjusted through reverse gradient update to optimize the original fault recognition model to obtain the transmission line typical fault recognition model, including:

[0027] S31. Initialize the weights and hyperparameters of the original recognition model;

[0028] S32. Define the loss function, use CIOU_Loss as the anchor box loss function, and use the binary cross-entropy function as the classification loss function. The calculation formulas of CIOU_Loss are shown in (1) and (2).

[0029]

[0030] Among them, C IoU_Loss represents the anchor box loss function, IoU represents the intersection over union, α is a parameter used to balance the ratio, b and b gt are the center points of the anchor box and the target box respectively, ρ represents the Euclidean distance between the two center points, c1 represents the diagonal distance of the smallest rectangle that simultaneously covers the anchor box and the target box, w and h are the width and height of the anchor box respectively, w gt and h gt are the width and height of the target box respectively, and v is the aspect ratio;

[0031] S33. Define the optimizer and use SGD for optimization;

[0032] S34. Input the test set into the original recognition model, calculate the values of the anchor box loss function and the classification loss function based on the prediction results and the true results, perform weighted summation on the values of the two loss functions to obtain the total loss value, and calculate the gradient according to the total loss value using the chain rule; use the SGD optimizer to update the model parameters in the reverse direction according to the gradient to obtain the transmission line typical fault recognition model.

[0033] Based on the same inventive concept, a transmission line typical fault recognition system based on the BiRG - YOLO algorithm is also proposed, including: an image acquisition module, an image pre - processing module, a detection model training module, a user interaction module, and a detection module;

[0034] The image acquisition module is used to obtain the aerial image data of the transmission line;

[0035] The image pre - processing module is used to pre - process the aerial image data of the transmission line and divide the pre - processed aerial image data of the transmission line into a training set and a test set; the pre - processing includes: image normalization, data augmentation, and marking of insulators and faults;

[0036] The detection model training module is used to construct a BiRG - YOLO network based on the YOLOv8n network, input the training set into the BiRG - YOLO network for training to obtain the original recognition model, and input the test set into the original recognition model for optimization to obtain the transmission line typical fault recognition model;

[0037] The user interaction module is used to encapsulate the transmission line typical fault recognition model through PyQT5, call the transmission line typical fault recognition model, construct a client, and control the system to perform fault recognition according to the instructions input by the user;

[0038] The detection module is used to input the image data to be detected into the transmission line typical fault recognition model through PyQT5, detect the image data to be detected, and obtain the transmission line typical fault recognition result.

[0039] Beneficial effects: Compared with the existing transmission line fault identification methods, in view of the characteristics of aerial images of transmission lines and typical fault components, the present invention introduces the BiFomrer module into YOLOv8n. By enhancing the feature extraction ability, the small target detection ability is improved, which is particularly suitable for identifying tiny faults in complex environments such as insulator bursts and flashovers. The RepGhost-C2f module is introduced into YOLOv8n. By means of the reparameterization idea, the number of parameters is reduced, the speed is increased while maintaining the detection accuracy, the model volume is reduced, and the deployability of the present invention is improved. Finally, the trained model can specifically detect tiny faults in the image and improve the accuracy of fault analysis. Description of the Drawings

[0040] In order to more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can also be obtained based on these drawings.

[0041] Figure 1 It is a method flow chart of a method for identifying typical faults in transmission lines based on the BiRG-YOLO algorithm provided by the present invention;

[0042] Figure 2 It is a schematic structural diagram of the BiRG-YOLO network of the present invention;

[0043] Figure 3 It is an example diagram of the feature map extracted by the BiRG-YOLO network of the present invention;

[0044] Figure 4 It is a schematic overall structural diagram of the BiFormer module;

[0045] Figure 5 It is a schematic structural diagram of the BiFormer Block module;

[0046] Figure 6 It is a schematic structural diagram of the RepGhost-Bottleneck module;

[0047] Figure 7 It is an example diagram of the detection result of the present invention;

[0048] Figure 8 It is a system structure diagram of a system for identifying typical faults in transmission lines based on the BiRG-YOLO algorithm provided by the present invention;

[0049] Figure 9 It is an interactive interface display diagram of the client in the embodiment of the present invention. Detailed implementation manners

[0050] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are some, but not all, of the embodiments of the present invention. All other embodiments obtained by those of ordinary skill in the art based on the embodiments of the present invention without creative efforts shall fall within the protection scope of the present invention.

[0051] This embodiment provides a method for identifying typical faults in transmission lines based on the BiRG-YOLO algorithm, as Figure 1 shown, including:

[0052] S1: Obtain aerial image data of the transmission line, preprocess the aerial image data of the transmission line to obtain the processed aerial image, and divide it into a training set and a test set;

[0053] S2: Introduce the YOLOv8n network, and use the BiFormer module and RepGhost-C2f to improve the YOLOv8n network to obtain the BiRG-YOLO network. Input the training set into the BiRG-YOLO network for training to obtain the original fault identification model;

[0054] S3: Input the test set into the original identification model, adjust the training parameters by backpropagation gradient update, and optimize the original fault identification model to obtain a typical fault identification model for the transmission line;

[0055] S4: Input the aerial image data of the transmission line to be identified into the typical fault identification model of the transmission line to obtain the fault identification result of the transmission line.

[0056] Specifically, first obtain the aerial image data of the transmission line, preprocess the aerial image data of the transmission line to obtain the processed aerial image, and divide it into a training set and a test set. Preprocessing the image can unify the image storage size, reduce the amount of calculation data, improve the recognition effect of the image, and further improve the generalization ability of the model;

[0057] Secondly, based on the introduction of the YOLOv8n network, the BiFormer module and RepGhost-C2f are used to improve the YOLOv8n network, resulting in the BiRG-YOLO network. The training set is input into the BiRG-YOLO network for training to obtain the original fault recognition model; the training set is input into the BiRG-YOLO network for training to obtain the original fault recognition model; the test set is input into the original recognition model, and the training parameters are adjusted by backpropagation gradient update to optimize the original fault recognition model, obtaining a transmission line typical fault recognition model; by constructing the BiRG-YOLO network and training the BiRG-YOLO network, the obtained transmission line typical fault recognition model can enhance the feature extraction ability, improve the small target detection ability, and is particularly suitable for identifying tiny faults in complex environments such as insulator bursts and flashovers, improving the detection accuracy;

[0058] Finally, the aerial image data of the transmission line to be identified is input into the transmission line typical fault recognition model to obtain the fault recognition result of the transmission line.

[0059] In a specific embodiment, the scheme for obtaining the aerial image data of the transmission line, preprocessing the aerial image data of the transmission line to obtain the processed aerial image, and dividing it into a training set and a test set is as follows:

[0060] S11. Acquisition of the dataset: Since there is currently no publicly available dataset for transmission line inspection images, in this scheme, an aerial transmission line fault database is independently established, and a large amount of transmission line image data is obtained by means of drone aerial photography. To improve the quality of the dataset and the robustness of the model, open-source datasets collected from the Internet are added to the dataset;

[0061] S12. Image normalization: According to the length of the shorter side of the image, the resolution of all images is uniformly normalized in proportion and adjusted to a size of 500×500;

[0062] S13. Data augmentation: Data augmentation processing is performed on the fault samples, morphological operations of brightness adjustment, rotation, and flipping are respectively performed to simulate different shooting angles, different lighting conditions, etc., and duplicate shooting images are removed. The total number of enhanced fault negative samples is 9755, further improving the generalization ability of the model;

[0063] S14. Making image labels: The insulator is marked in the image using the labelimg tool. Define fall as the fault of the grading ring falling off, burst as the fault of the glass insulator bead bursting, break as the fault of the insulator being defective, nest as the fault of birds building nests, and flashover as the flashover; the dataset is shown in Table 1,

[0064] Table 1

[0065] Category Number of enhanced samples Image size fall 1584 4608×3456 burst 1834 4608×3456 break 2416 5472×3648 nest 1556 5472×3648 flashover 2365 5472×3648

[0066] S15. Divide the images into a training set and a test set at a ratio of 7:3.

[0067] Due to the limitation of computing resources and the inconsistent sizes of the image samples obtained by aerial photography equipment, the difficulty of insulator positioning is greatly increased. Therefore, in this solution, the images are normalized to unify the image storage size and reduce the amount of calculation data. Since it is difficult to collect data on faulty components of transmission lines, resulting in a small amount of faulty data, multiple line samples including sunny, cloudy, and rainy days in all four seasons are actually photographed, and duplicate photographed images are included. Therefore, in this solution, the images are enhanced to improve the recognition effect of the images and further improve the generalization ability of the model.

[0068] In a specific embodiment, the YOLOv8n network is introduced, and the BiFormer module and RepGhost-C2f are used to improve the YOLOv8n network to obtain the BiRG-YOLO network. The solution for training the original fault recognition model by inputting the training set into the BiRG-YOLO network is as follows:

[0069] As Figure 2 shown, the BiRG-YOLO network includes an input end, a Backbone network, a Neck network, and a Head network;

[0070] The input end is used to process the images and adjust the sizes of the images;

[0071] The Backbone network is used to reduce the resolution of the images, extract the features of the targets in the images, and output feature maps of three sizes;

[0072] The Neck network is the same as the Neck network in the original YOLOv8n network, including an FPN network and a PAN network. It receives the feature maps of three sizes output by different layers in the Backbone network, performs bidirectional feature fusion through the FPN network and the PAN network, fuses the features in each size of the feature maps into the other two sizes of the feature maps respectively, and outputs the fused feature maps of three sizes and transmits them to the Head network;

[0073] The Head network is the same as the Head network in the original YOLOv8n network, including three groups of decoupled heads that respectively process the fused feature maps of three sizes. Each group of decoupled heads includes a convolutional layer, a classification branch, and a regression branch, and is used to predict the fused feature maps of three sizes, including representing the detected target positions using bounding boxes, outputting the class probabilities of each bounding box, and the confidence levels of the detected targets; the output feature maps are as Figure 3 shown;

[0074] In the Backbone network of the original YOLOv8n network, the first and second C2f modules are replaced with BiFormer modules, and the third and fourth C2f modules are replaced with RepGhost-C2f modules;

[0075] The improved Backbone network then includes a first convolutional module, a second convolutional module, a first BiFormer module, a third convolutional module, a second BiFormer module, a fourth convolutional module, a first RepGhost-C2f module, a fifth convolutional module, a second RepGhost-C2f module, and an SPPF module connected in sequence;

[0076] The BiFormer module BiFormer adopts a 4-layer pyramid structure, and its structure is as Figure 4 shown. The network is divided into four stages. Stage 1 uses overlapping image patch embedding, and stages 2-4 are image merging modules that reduce the resolution and increase the number of channels. Each stage uses a BiFormer Block to transform the features of the input data;

[0077] The structure of the BiFormer Block is as Figure 5 shown, including a depth convolutional module, two layer normalization modules, a BRA module, and a multi-layer perceptron module.

[0078] The depth convolutional module is a 3*3 convolutional layer that uses depthwise separable convolution to implicitly encode relative position information, uses the BRA module to capture bidirectional relative attention, and finally uses a two-layer perceptron (MLP) module for relationship modeling and position embedding;

[0079] The Bottleneck module in the C2f module of the original YOLOv8n network is replaced with RepGhostBottleneck to form the RepGhost-C2f module, which extracts the features of the output image of the previous layer and performs feature fusion.

[0080] In this solution, BiFormer is a vision transformer based on the sparse attention mechanism. Its core component is the Bi-Level Routing Attention (BRA). The core idea of BRA is to filter out most of the least relevant key-value pairs at the rough region level, only retaining a small number of routing regions, removing redundant information, and then applying fine-grained attention in the union of the selected routing regions. BRA restricts the scope of attention calculation, reducing the computational complexity while improving performance. BiFormer relies on the structural characteristics of the transformer and uses BRA for sparse sampling, replacing the downsampling of convolution in Convolutional Neural Networks (CNNs). In this solution, the BiFormer module is used to replace the first two C2f modules, avoiding the large-scale stacking of convolutional layers in the entire backbone network, improving the problem of information loss caused by downsampling, being friendly to small object detection, and fully retaining the information in the feature map for deep use.

[0081] In this solution, the Bottleneck module in the original C2f is replaced by RepGhost-Bottleneck to achieve implicit feature reuse. RepGhost is a lightweight neural network that abandons the inefficient Concat operation, and the information fusion process is implicitly executed by the Add operation, improving the inference speed. Moreover, RepGhost uses two different structures in the training and inference stages, and the model structure in the inference stage is simpler, reducing the use of the Add operation and activation functions, further enhancing the real-time performance of the model.

[0082] The structure diagrams of the two branches of RepGhost are as Figure 6 shown. RepGhost Bottleneck is extremely simple in the inference stage, with only one shortcut block branch and a single-chain operation composed of a 1×1 convolution, depthwise convolution, and ReLU. This has a huge impact on reducing the model size and improving the inference speed. RepGhost Bottleneck ensures the feature fusion of different layers in the training stage through reparameterization, maintaining the generation of different feature maps, thus ensuring the feature extraction ability of the model.

[0083] In this solution, the Neck network samples the original Neck network structure of Yolov8n. Its structure follows the CSP structure of the backbone network, which can better fuse the features extracted by the backbone network. It uses a structure that combines the Feature Pyramid Network (FPN) and the Path Aggregation Network (PAN). FPN constructs high-level semantic feature maps at all scales from top to bottom. Its receptive field is large, relatively abstract, and the extracted features are beneficial for completing classification, but some information will be lost, which is not conducive to precise positioning. Therefore, the features are then passed from bottom to top through PAN to make up for and strengthen the positioning information, and finally the features are passed to the output end.

[0084] By replacing the first two C2f modules of the Backbone network of the original YOLOv8n with BiFomrer modules, the ability to extract features is enhanced, and the ability to detect small targets is improved. It is particularly suitable for identifying tiny faults in complex environments such as insulator burst and flashover. The last two C2f modules of the Backbone network of the original YOLOv8n are replaced with RepGhost-C2f modules, and the number of parameters is reduced through the idea of reparameterization, the speed is increased while maintaining the detection accuracy, the model volume is reduced, and the deployability of this model is improved.

[0085] In a specific embodiment, the scheme of inputting the training set into the BiRG-YOLO network for training to obtain the original fault recognition model is as follows:

[0086] At the input end (Input), the Mosaic data augmentation method is used to randomly crop and splice the input images to improve the model training effect, and the size of the input images is uniformly adjusted to 640x640;

[0087] The processed images are input into the Backbone network to extract semantic features at different levels of the images, and three sizes of basic feature maps are obtained; in this solution, after the combination of the first consecutive BiFormer and convolutional modules, the size of the feature map is 80×80, and after the combination of the second consecutive BiFormer and convolutional modules, the size of the feature map is 40×40. These two sizes of feature maps have undergone sufficient feature extraction through convolutional modules and BiFormer modules, and already contain a large amount of feature information, and the sizes are different, which are suitable for detecting targets of different sizes; when the 40×40 feature map enters the RepGhostC2f module, the feature map is fused. The above-mentioned fused 40×40 feature map will enter a convolutional module and become a 20×20 feature map, which is then input into RepGhost C2f again to obtain a 20×20 feature map with richer features;

[0088] The basic feature maps of three sizes are input into the Neck network for bidirectional feature fusion to obtain fused feature maps of three sizes. After cross-scale feature fusion by the Neck network, three final feature maps with sizes of 80×80, 40×40, and 20×20 can be obtained. These three final feature maps will be fed into the Head network for result prediction of targets of large, medium, and small sizes.

[0089] The fused feature maps of three sizes are input into the Head network and divided into feature maps of S×S grids (the value of S is different for feature maps of different sizes). Through the adaptive anchor box strategy, anchor boxes of different sizes are generated according to the width and height of the ground truth boxes in the training set. The class probability and confidence of each anchor box are predicted through the classification branch. The coordinates of the bounding box corresponding to the anchor box are predicted through the regression branch. The target boxes are filtered out by the non-maximum suppression method, the coordinates of the target boxes are obtained, and the bounding boxes are drawn on the original image to frame the targets.

[0090] In a specific embodiment, the scheme for inputting the test set into the original recognition model, adjusting the training parameters through negative feedback operations, and optimizing the original fault recognition model to obtain a transmission line typical fault recognition model is as follows:

[0091] S31. Initialize the weights and hyperparameters of the original recognition model;

[0092] S32. Define the loss functions. Use CIOU_Loss as the anchor box loss function and the binary cross-entropy function as the classification loss function. The calculation formula of CIOU_Loss is shown in (3) and (4).

[0093]

[0094] Among them, C IoU_Loss represents the anchor box loss function, IoU represents the intersection over union, α is a parameter for balancing the ratio, b and b gt are the center points of the anchor box and the target box respectively, ρ represents the Euclidean distance between the two center points, c1 represents the diagonal distance of the smallest rectangle that simultaneously covers the anchor box and the target box, w and h are the width and height of the anchor box respectively, w gt and h gt are the width and height of the target box respectively, and v is the aspect ratio.

[0095] S33. Define the optimizer and use SGD for optimization;

[0096] S34. Input the test set into the original recognition model, calculate the values of the anchor box loss function and the classification loss function based on the prediction results and the true results, perform a weighted sum of the values of the two loss functions to obtain the total loss value, and calculate the gradient according to the total loss value using the chain rule; use the SGD optimizer to update the model parameters backward according to the gradient to obtain the typical fault recognition model for transmission lines.

[0097] In this solution, optimizing the model can effectively avoid overfitting and improve the performance of the model on the test set, and is applicable to complex tasks such as object detection.

[0098] In a specific embodiment, the solution for inputting the aerial image data of the transmission line into the typical fault recognition model of the transmission line to obtain the fault recognition result of the transmission line is as follows:

[0099] Input the aerial image dataset of the transmission line into the trained typical fault recognition model of the transmission line for detection. For the input aerial image, generate feature maps of 3 different scales and divide them into feature maps of S×S grids (for feature maps of different scales, the value of S is different). Through the adaptive anchor box strategy, generate anchor boxes of different sizes according to the width and height of the true boxes in the training set. Predict the class probability and confidence of each anchor box through the classification branch; predict the coordinates of the bounding box corresponding to the anchor box through the regression branch; use the non-maximum suppression method to select the most accurately predicted box. When the class of each anchor box in all grid regions of the original image is predicted, perform information integration, output the complete target information of the entire image, and finally frame the target in the original image, which is the final fault recognition result of the transmission line. The recognition result is as Figure 7 shown.

[0100] This embodiment also provides a typical fault recognition system for transmission lines based on the BiRG-YOLO algorithm, as Figure 8 shown, including: an image acquisition module, an image preprocessing module, a detection model training module, a user interaction module, and a detection module;

[0101] The image acquisition module is used to obtain the aerial image data of the transmission line;

[0102] The image preprocessing module is used to preprocess the aerial image data of the transmission line and divide the preprocessed aerial image data of the insulator into a training set and a test set; the preprocessing includes: image normalization, data augmentation, and marking of insulators and faults;

[0103] The detection model training module is used to construct a BiRG-YOLO network based on the YOLOv8n network, input the training set into the BiRG-YOLO network for training to obtain the original recognition model, and input the test set into the original recognition model for optimization to obtain the typical fault recognition model for transmission lines;

[0104] The user interaction module is used to encapsulate the typical transmission line fault recognition model through PyQT5, call the typical transmission line fault recognition model, construct a client, and control the system to perform fault recognition according to the instructions input by the user;

[0105] Specifically, in the user interaction module, PyQT5 is used to design an interactive interface as a client for users. The functions of this client include: input of real-time video stream, selection of local video input, graphics card selection, weight model selection, local picture input, folder input, etc. The algorithm model can be added or reduced by replacing the configuration file at the system backend, enhancing the system scalability and facilitating user use; the data end is as Figure 9 shown;

[0106] The detection module is used to input the image data to be detected into the typical transmission line fault recognition model through PyQT5, detect the image data to be detected, and obtain the typical transmission line fault recognition result;

[0107] Specifically, for the aerial image of the transmission line input by the detection module, it is passed to the typical transmission line fault recognition model through the client based on PyQT5, generating feature maps of multiple different scales and dividing them into feature maps of S×S grids. The grid where the target center is located is responsible for detecting the target with the corresponding anchor box. Each grid will predict multiple target boxes and the confidence of the target boxes, and perform class prediction; the output end is used to screen out the most accurate target box. When all grid areas in the original transmission line image have predicted their respective classes, information synthesis is performed to output the complete target information of the entire picture, and finally the faulty components are framed in the image.

[0108] This system provides a user interaction QT interface, encapsulates the typical transmission line fault recognition model, facilitates users to intuitively and real-time view the fault detection results, and improves the efficiency of fault analysis and decision-making.

[0109] Compare the typical transmission line fault recognition model obtained by using the present invention with several improvement methods of YOLOv8N:

[0110] In the ablation experiment, as shown in Table 2, the present invention uses mAP0.50, mAP0.5:0.95, mAP0.75, the number of parameters, and the average detection time of a picture as evaluation indicators. Serial number 1 is the original YOLOv8n, serial number 2 is adding the BiFormer module only to the shallow layer of the backbone network of the original YOLOv8n, serial number 3 is adding the RepGhost-C2f module only to the deep layer of the backbone network of the original YOLOv8n, and serial number 4 is the typical transmission line fault recognition model of the present invention.

[0111] Table 2

[0112]

[0113] As can be seen from Table 2, the BiFormer module significantly improves the detection accuracy of the model, but affects the number of parameters and speed. The RepGhost-C2f module can reduce the number of parameters and improve the speed, but has a slight impact on the accuracy. The typical transmission line fault recognition model combines the advantages of detection accuracy, the number of parameters, and detection speed, and has the best performance.

[0114] Table 3 shows the data analysis of each type of fault based on the ablation experiment in Table 2.

[0115] Table 3

[0116]

[0117]

[0118] As can be seen from Table 3, the typical transmission line fault recognition model performs excellently in each type of fault.

[0119] The current mainstream algorithms are used for comparison on the same dataset, including YOLOv3, YOLOv5s, and YOLOv7 algorithms.

[0120] Table 4

[0121] Algorithm Map0.5(%) Map0.5:0.95(%) Number of parameters (Param) Time(ms) YOLOv3 85.1 52.3 65.2M 10.2 YOLOv5s 90.2 56.6 7.4M 7.4 YOLOv7 91.9 59.8 36.72M 9.3 YOLOv8n 91.5 58.5 3.01M 4.5 Ours 93.2 60.2 2.65M 4.4

[0122] As can be seen from Table 4, compared with other mainstream algorithms, the method of the present invention has improved speed while significantly improving the accuracy, and at the same time maintains the smallest number of model parameters, which is more convenient for deployment.

[0123] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions described in the foregoing embodiments, or perform equivalent replacements for some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A typical fault identification method for power transmission lines based on BiRG-YOLO algorithm, characterized in that: include: S1: Acquire aerial image data of power transmission lines, pre-process the aerial image data of power transmission lines, obtain processed aerial images, and divide them into a training set and a test set; S2: Introduce the YOLOv8n network, use the BiFormer module and RepGhost-C2f to improve the YOLOv8n network, obtain the BiRG-YOLO network, input the training set into the BiRG-YOLO network for training, and obtain the original fault recognition model; S3: inputting the test set into the original recognition model, adjusting the training parameters by reverse gradient updating, optimizing the original fault recognition model, and obtaining a typical fault recognition model for the transmission line; S4: Inputting the aerial image data of the power transmission line to be identified into the typical fault identification model of the power transmission line to obtain a fault identification result of the power transmission line.

2. According to claim 1, a typical fault identification method for a power transmission line based on the BiRG-YOLO algorithm is characterized in that: The YOLOv8n network is introduced, and the BiFormer module and RepGhost-C2f are used to improve the YOLOv8n network to obtain the BiRG-YOLO network, including: The BiRG-YOLO network includes the input end, the Backbone network, the Neck network, and the Head network; The input end is used to process the image and adjust the image size; The Backbone network is used to reduce the resolution of the image, extract the features of the target in the image, and output feature maps of three sizes; The Neck network is the same as the Neck network in the original YOLOv8n network, including the FPN network and the PAN network. It receives the feature maps of three sizes output by different layers in the Backbone network, performs bidirectional feature fusion through the FPN network and the PAN network, fuses the features of the feature map of each size into the feature maps of the remaining two sizes, outputs the fused feature map of three sizes and passes it to the Head network; The head network is the same as the head network in the original YOLOv8n network, including three sets of decoupling heads that process the fused feature maps of three sizes respectively. Each set of decoupling heads includes a convolutional layer, a classification branch, and a regression branch, which are used to predict the fused feature maps of three sizes, including using bounding boxes to represent the detected target locations, outputting the category probability of each bounding box, and the confidence of the detected target existence; The Bottleneck module in the C2f module in the original YOLOv8n network is replaced with the RepGhost Bottleneck to form the RepGhost-C2f module; In the Backbone network in the original YOLOv8n network, the first and second C2f modules are replaced by BiFormer modules, and the third and fourth C2f modules are replaced by RepGhost-C2f modules; The improved Backbone network includes a first convolution module, a second convolution module, a first BiFormer module, a third convolution module, a second BiFormer module, a fourth convolution module, a first RepGhost-C2f module, a fifth convolution module, a second RepGhost-C2f module and an SPPF module, which are connected in sequence.

3. According to claim 2, a typical fault identification method for a power transmission line based on the BiRG-YOLO algorithm is characterized in that: The training set is input into the BiRG-YOLO network for training to obtain the original fault recognition model, including: Mosaic data augmentation method is used at the input end to randomly crop and splice the input image; The processed image is input into the Backbone network to extract the semantic features of different levels of the image and obtain basic feature maps of three sizes; The basic feature maps of three sizes are input into the Neck network for bidirectional feature fusion to obtain fused feature maps of three sizes; The fused feature maps of three sizes are input into the Head network and divided into feature maps of S×S grids. Through the adaptive anchor box strategy, anchor boxes with different size features are generated according to the width and height of the real box in the training set. The classification branch predicts the category probability and confidence of each anchor box; the regression branch predicts the coordinates of the bounding box corresponding to the anchor box; The target box is filtered out through the non-maximum suppression method, the coordinates of the target box are obtained, and the bounding box is drawn in the original image to frame the target.

4. According to claim 3, a typical fault identification method for a power transmission line based on the BiRG-YOLO algorithm is characterized in that: The test set is input into the original recognition model, and the training parameters are adjusted by reverse gradient updating to optimize the original fault recognition model to obtain a typical fault recognition model for power transmission lines, including: S31, initializing the weights and hyperparameters of the original recognition model; S32. Define the loss function, use CIoU_Loss as the anchor box loss function, and use the binary cross entropy function as the classification loss function. The calculation formula of CIOU_Loss is shown in (1) and (2). Among them, C IoU_Loss represents the anchor box loss function, IoU represents the intersection over union ratio, α is a parameter used to balance the ratio, b and b gt are the center points of the anchor box and the target box, respectively. ρ represents the Euclidean distance between the two center points. c1 represents the diagonal distance of the smallest rectangle that covers both the anchor box and the target box. w and h are the width and height of the anchor box, respectively. w gt and h gt are the width and height of the target box, respectively, and v is the aspect ratio; S33. Define the optimizer and use SGD for optimization; S34. Input the test set into the original recognition model, calculate the value of the anchor box loss function and the value of the classification loss function based on the predicted results and the actual results, perform weighted summation on the values ​​of the two loss functions to obtain the total loss value, and use the chain method to calculate the gradient according to the total loss value; use the SGD optimizer to reversely update the model parameters according to the gradient to obtain a typical fault recognition model for transmission lines.

5. A typical fault identification system for power transmission lines based on the BiRG-YOLO algorithm, using the typical fault identification method for power transmission lines based on the BiRG-YOLO algorithm as claimed in claim 1, characterized in that: include: Image acquisition module, image preprocessing module, detection model training module, user interaction module and detection module; The image acquisition module is used to obtain aerial image data of power transmission lines; The image preprocessing module is used to preprocess the transmission line aerial image data, and divide the preprocessed transmission line aerial image data into a training set and a test set; The preprocessing includes: image normalization, data enhancement, and marking of insulators and faults; The detection model training module is used to build a BiRG-YOLO network based on the YOLOv8n network, input the training set into the BiRG-YOLO network for training to obtain an original recognition model, input the test set into the original recognition model for optimization, and obtain a typical fault recognition model for transmission lines; The user interaction module is used to encapsulate the typical fault identification model of the transmission line through PyQT5, call the typical fault identification model of the transmission line, build a client, and control the system to perform fault identification according to the command input by the user; The detection module is used to input the image data to be detected into the typical fault recognition model of the power transmission line through PyQT5, detect the image data to be detected, and obtain the typical fault recognition result of the power transmission line.

Images